CN114036396B

CN114036396B - Content processing method and device, storage medium and electronic equipment

Info

Publication number: CN114036396B
Application number: CN202111415574.1A
Authority: CN
Inventors: 戴朝约; 黄华
Original assignee: Netease Media Technology Beijing Co Ltd
Current assignee: Netease Media Technology Beijing Co Ltd
Priority date: 2021-11-25
Filing date: 2021-11-25
Publication date: 2025-01-17
Anticipated expiration: 2041-11-25
Also published as: CN114036396A

Abstract

The disclosed embodiments provide a content processing method, device, storage medium and electronic device, which filter the content related to the user's explicit negative feedback behavior in the filtering link in the recommendation calculation process of the recommendation system; and suppress the ranking weight of the content related to at least the user's implicit negative feedback behavior in the detailed sorting link. Thus, in the calculation process of the recommendation system, the content related to the user's explicit and implicit negative feedback behavior is filtered and suppressed in a targeted manner, effectively reducing the parts of the recommended content that the user does not like or is not interested in, effectively ensuring the accuracy of the recommendation system, and effectively improving the user experience.

Description

Content processing method and device, storage medium and electronic equipment

Technical Field

Embodiments of the present disclosure relate to the field of data processing technologies, and more particularly, to a content processing method, apparatus, storage medium, and electronic device.

Background

This section is intended to provide a background or context for the embodiments of the disclosure recited in the claims, which description herein is not admitted to be prior art by inclusion in this section.

Recommendation algorithms have been commonly applied to various user services in the internet. However, some current recommendation algorithms have inaccurate problems in the way content that users may dislike is filtered, resulting in content recommended to users still containing a large amount of content that is not of interest to users.

Disclosure of Invention

In the recommendation algorithm, the following ways are generally adopted at present to eliminate content which may not be liked by the user in the recommended content. One way is to mask all content related to the tags from being recommended to the user according to the behavior that the user obviously shows dislike, such as the tags of the mask content clicked by the user, but the tags of the content may not be comprehensive, so that the content of the same type cannot be limited to be continuously recommended to the user, and the other way is to show that the preference of the user to the content is possibly marked by scoring, but the scoring is possibly influenced by other factors possibly related to the preference of the user, and when other factors are inconsistent with the preference of the user, the user is likely to dislike the content, but the score is not low and is recommended, and the method is inaccurate. It can be seen that the above methods may cause that the final recommended content still contains content that is disliked by the user, resulting in poor user experience.

For this reason, an improved solution is highly needed to accurately filter the dislike content of the user in the recommendation algorithm, so as to improve the user experience.

In this context, embodiments of the present disclosure provide a content processing method, apparatus, storage medium, and electronic device.

According to a first aspect of the present disclosure, a content processing method is provided, which is applied to a recommendation system, wherein a recommendation calculation flow of the recommendation system includes a coarse ranking link, a filtering link and a fine ranking link of content, and the method includes:

In the filtering link, filtering content related to the explicit negative feedback behavior of the user;

And in the fine ranking link, suppressing the ranking weight of the content at least related to the implicit negative feedback behavior of the user.

According to a second aspect of the present disclosure, a content processing device is provided, and the content processing device is applied to a recommendation system, wherein a calculation flow of the recommendation system includes a coarse ranking link, a filtering link and a fine ranking link of content, and the device includes:

the explicit negative feedback processing module is used for filtering the content related to the explicit negative feedback behavior of the user in the filtering link;

And the implicit negative feedback processing module is used for inhibiting the ordering weight of the content at least related to the implicit negative feedback behavior of the user in the fine ranking link.

According to a third aspect of the present disclosure, there is provided a storage medium having stored thereon a computer program which, when executed by a processor, implements the content processing method according to any of the embodiments of the first aspect.

According to a fourth aspect of the present disclosure, there is provided an electronic device comprising:

Processor, and

A memory for storing executable instructions of the processor;

wherein the processor is configured to execute via execution of the executable instructions:

a content processing method as in any one of the embodiments of the first aspect.

According to the embodiment of the disclosure, the content related to the explicit negative feedback behavior of the user is filtered through a filtering link in a recommendation calculation flow of a recommendation system, and the ranking weight of the content related to at least the implicit negative feedback behavior of the user is restrained in the fine ranking link. Therefore, the method and the device realize the targeted filtering and suppression of the content related to the explicit and implicit type negative feedback behaviors of the user in the calculation process of the recommendation system, effectively reduce the unfavorable and uninteresting parts of the user in the recommendation content, effectively ensure the accuracy of the recommendation system and effectively improve the user experience.

Drawings

The above, as well as additional purposes, features, and advantages of exemplary embodiments of the present disclosure will become readily apparent from the following detailed description when read in conjunction with the accompanying drawings. Several embodiments of the present disclosure are illustrated by way of example, and not by way of limitation, in the figures of the accompanying drawings, in which:

FIG. 1 illustrates a schematic diagram of a recommendation calculation flow for a recommendation system in an example of the present disclosure.

Fig. 2 shows a flow diagram of a content processing method in an embodiment of the present disclosure.

Fig. 3 is a flowchart illustrating a procedure performed in step S201 of the content processing method in an embodiment of the present disclosure.

Fig. 4 shows a schematic diagram of a user negative feedback model in an embodiment of the present disclosure.

Fig. 5 shows a flowchart of the step S202 of the content processing method in an embodiment of the present disclosure.

FIG. 6 illustrates a flow diagram of a method of training a machine learning model in an embodiment of the present disclosure.

Fig. 7 shows a block diagram of a content processing apparatus in an embodiment of the present disclosure.

FIG. 8 shows a schematic diagram of a storage medium in an embodiment of the present disclosure, and

Fig. 9 shows a block diagram of an electronic device in an embodiment of the disclosure.

In the drawings, the same or corresponding reference numerals indicate the same or corresponding parts.

Detailed Description

The principles and spirit of the present disclosure will be described below with reference to several exemplary embodiments. It should be understood that these embodiments are presented merely to enable one skilled in the art to better understand and practice the present disclosure and are not intended to limit the scope of the present disclosure in any way. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

Those skilled in the art will appreciate that embodiments of the present disclosure may be implemented as a system, apparatus, device, method, or computer program product. Accordingly, the present disclosure may be embodied in the form of entirely hardware, entirely software (including firmware, resident software, micro-code, etc.) or in a combination of hardware and software.

According to an embodiment of the disclosure, a content processing method, a content processing device, a storage medium and an electronic device are provided.

Any number of elements in the figures are for illustration and not limitation, and any naming is used for distinction only, and not for any limiting sense.

The principles and spirit of the present disclosure are described in detail below with reference to several representative embodiments thereof.

Summary of The Invention

With the wide application of the internet and mobile internet, users can obtain various contents through the internet. The content refers to a general term for all information distributed on the internet that can be read, and has various types of presentation forms, such as articles, videos, music, and the like.

Because the interest preference of the users to the content is different, different content needs to be distributed to different users through a computer algorithm so as to meet different personalized requirements of each user, which is called personalized recommendation. The personalized recommendation is typically performed by a recommendation system.

As illustrated in fig. 1, the recommendation calculation flow of the recommendation system for personalized recommendation includes a recall link 101, a coarse ranking link 102, a filtering link 103, and a fine ranking link for content. Illustratively, the fine row links are shown as including fine row links 104 and rearrangement links 105. In other embodiments, there may be no rearrangement links 105, which is not limited to this example.

Content recall, among other things, may also be generally referred to as tag recall, where content that may be of interest to a user is determined, for example, based on how well the user's interest tag matches the content tag. Furthermore, the data volume of the recalled content is huge and huge, the recalled content is ranked and scored and filtered according to the interests of the user, the content quality and the like, and the content with the ranking being back is removed, so that recommended content is finally obtained and recommended to the user.

The coarse ranking link 102 may perform coarse ranking scoring on recall content, and rank according to the score, and take out the portion with higher coarse ranking score for subsequent links. The characteristics of the content according to which the coarse row scoring is performed can be obtained through statistics or can be abstracted. Taking a text article as an example, the feature dimensions of the scoring references comprise text relevance, browsed quantity and the like, the browsed quantity can be obtained statistically, and the text relevance is obtained through text matching calculation. The features and feature numbers of coarse row scoring are closely related to the content and may be quite different from one type of content to another.

The filtering unit 103 may be configured to filter low quality portions of the coarse-ranking result obtained after the coarse-ranking unit 102, for example, to filter out content related to labels and bias (low colloquial, violent, nausea, etc.), low browsing count, old, etc.

The fine ranking step 104 is configured to score the fine ranking of the filtered result, rank the fine ranking according to the score of the fine ranking, and filter the content with a lower score to form a fine ranking result.

The rearrangement unit 105 may be configured to reorder at least a part of the content (such as the part with the front score) in the fine-ranking result, so as to obtain the recommended content for recommendation to the user.

It will be appreciated that the links of recall, coarse ranking, filtering, fine ranking, rearrangement, etc. in the recommended calculation flow may be changed, for example, the filtering flow may be added after recall, coarse ranking, fine ranking, etc. for example, filtering may be omitted when the content quality is better, rearrangement may be omitted when the recommendation accuracy is not high, and fig. 1 is only a representative example.

In the recommended computation flow, some means are employed below in order to remove content, low-quality content, or the like that may not be of interest to the user.

One approach is to directly mask content with a content tag that is "uninteresting" based on the user's selection of the content tag. For example, if the user clicks on content that does not like the "pet" type, the recommendation system does not recommend content related to "pet" to the user. However, this approach, while simple, has limitations such as selecting tags that dislike a certain content source (e.g., author, media), then the author's content is not subsequently recommended to the user, but in fact the user is of a type that dislikes content, then the other authors' same type of content may still be pushed to the user. In addition, the filtering means is not capable for the user to skip the content which is not selected to browse and possibly the content which is not interested by the user.

Another approach is to express the user's preferences by scoring content tags, such as coarse scoring, fine scoring, etc. However, when the recommendation system makes a recommendation, a large amount of content which is possibly uninteresting to the user but possibly popular still exists in the result of scoring ranking due to the influence of popular behaviors, and the purpose of suppressing the content which is uninteresting to the user is not achieved. Moreover, relying solely on scoring does not allow for the artificial and controlled suppression of the recommendation of a certain type of content to adjust the feedback sensitivity of the recommendation system. The higher the feedback sensitivity is, the more timely the feedback behavior to the user is indicated to adjust the recommended content. For example, when a user browses a certain type of content, the next recommendation will give a more informative recommendation of such content.

Based on the above several approaches, it is possible to include a large amount of content that is not interesting or liked by the user in the final recommended content, which affects the user experience.

In view of this, the present disclosure further enhances analysis based on explicit and implicit negative feedback behavior of the content by the user in the recommendation system, and performs corresponding filtering and suppression of the ranking score on the content unsuitable for recommendation, thereby effectively improving user experience.

Here, the definition of the "negative feedback behavior" will be explained. "negative feedback behavior" is in contrast to "positive feedback behavior" which expresses behavior that the user is interested in the content, prefers to like, such as browsing, clicking, etc. Correspondingly, "negative feedback behavior" means behavior that is not interesting, preferred or negative to the content representation. The "negative feedback behavior" can be further divided into "explicit negative feedback behavior" and "implicit negative feedback behavior". "explicit negative feedback behavior" refers to user-active negative feedback behavior such as clicking a dislike button on recommended content, clicking a decrease some type of tab type content recommendation button, and so forth. The "implicit negative feedback behavior" refers to that after receiving the content recommendation display, the user does not have positive active reading behavior, such as swiftly scrolling through the displayed content without clicking to enter reading.

Having described the basic principles of the present disclosure, various non-limiting embodiments of the present disclosure are specifically described below.

It should be specifically noted that, in the content processing method, the content processing apparatus and the like in the embodiments of the present disclosure, in the case of obtaining the approval of the user (such as an electronic protocol, confirmation of authorization and the like), the feedback behavior of the user on the content is collected, so as to improve the user experience on the premise of ensuring the security of the user data.

Exemplary method embodiment

Referring to fig. 2, a flow chart of a content processing method according to an embodiment of the disclosure is shown. The content processing method can be applied to a recommendation system and can be applied to a recommendation calculation flow of the recommendation system. The recommended computation flow of fig. 1 is illustratively incorporated in fig. 2 as a reference.

The content processing method comprises the following steps:

in the filtering step, filtering the content related to the explicit negative feedback behavior of the user.

Step S201 is performed in the filtering section 103 to strengthen the filtering action of the filtering section 103. In addition, the content data quantity flowing through the recommendation calculation flow is reduced link by link, and obvious dislike content of the user is filtered in advance at the position which is as close to the upstream as possible, so that the data quantity which enters the subsequent link and needs to be processed can be effectively reduced.

In some embodiments, step S201 may be applied to the filtering element 103 to filter the content. In a possible implementation example, the operation logic of step S201 may implement an explicit negative feedback processing module that can be used in the filtering link, and may be implemented based on software, hardware, or a combination of software and hardware.

And S202, suppressing the ordering weight of the content at least related to the implicit negative feedback behavior of the user in the fine-ranking link.

In some embodiments, the fine-ranking link may be a rearrangement link or a fine-ranking link, that is, the step S202 may be applied to the rearrangement link or the fine-ranking link, so as to suppress the ranking weight of the content targeted by at least the implicit negative feedback action when the data amount related to the content in the rearrangement link or the fine-ranking link is relatively small, and the content whose score is suppressed will be correspondingly later and more likely to be filtered out in the ranking result, thereby efficiently and accurately reducing the uninteresting part of the recommended content for the user.

It should be noted that, in fig. 2, step S202 is exemplarily applied to the rearrangement link to improve the processing efficiency. However, it is understood that the step S202 may be applied in the fine-ranking step, and may be selected according to the processing efficiency, the accuracy, etc., and is not limited to the above example. In a possible implementation example, the operation logic of the step S202 may be implemented as an implicit negative feedback processing module that can be used in the fine-ranking link or the rearrangement link, where the implicit negative feedback processing module may be implemented based on software, hardware, or a combination of software and hardware.

The implementation of each step in the content processing method will be described below by means of a plurality of examples.

As shown in fig. 3, a schematic flow chart of step S201 in an exemplary embodiment of the present disclosure is shown.

Referring to fig. 3, the process includes steps S301 to S303.

And step S301, determining and storing negative content attributes in the content attributes aimed at by the explicit negative feedback behavior.

Step S302, matching the negative content attribute with the content attribute of the matched content to obtain a matching result;

step S303, filtering matched content to which the content attribute of the matching result meets a preset matching condition belongs.

On the one hand, the description will be made with respect to step S301.

In some embodiments, the content attributes may include content tags. The content attributes may be derived from direct documentation or abstraction of the content. Taking a teletext article or video as an example, the content properties may illustratively comprise keywords, points of interest, content sources, categories, etc. For example, a keyword "football" is described in an article, and then "football" may be one of the content attributes of the article, a video about lobster cooking, a user's interest point "seafood" in the video may be one of the content attributes of the video, an author of the article is a, a may be one of the content attributes of the article, a publisher of the video is self-media B, B may be one of the content attributes of the video, and a short video about puppies interacting with the host, a "pet" may be one of the content attributes of the video.

In some embodiments, the negative content attribute refers to a content attribute for which the user's explicit negative feedback behavior is directed. For example, when a user browses an article, by selecting "dislike" the article to express a dislike meaning, one or more of a plurality of content attributes that the article corresponds to may be determined as a negative content attribute.

In some embodiments, the dislike degree of the user on the content attribute can be analyzed through explicit negative feedback behavior of the user, and if the dislike degree is reached, the content attribute can be considered as a negative content attribute. In a possible implementation example, the negative content attribute may be filtered based on the frequency of explicit negative feedback actions received during the backtracking period reaching a preset threshold. The backtracking period may be a period from a current time to a historical time, and the historical time may be preset, or a time when a content attribute is judged to be a negative content attribute according to a condition, where the content attribute is subjected to explicit negative feedback at the earliest. In a possible example, with a target user as a content recommendation target, by collecting the explicit negative feedback actions made by the target user on the content in a backtracking period, statistics are performed corresponding to content-related content attributes. For example, the user X browses the content a, the content B, the content C and the content D in the backtracking period, and the user X makes explicit negative feedback actions on the content a, such as dislike evaluation C1, makes negative feedback actions on the content C, such as dislike evaluation C2, and the content attributes of the content a and the content C both include the content attribute Y, so that the number of times of the explicit negative feedback actions on the content attribute Y is counted to be 2. If the length of the backtracking period is 300 seconds, the frequency of explicit negative feedback behavior of the user X with respect to the content attribute Y is 1/150 times/second. And whether the calculated frequency reaches a preset threshold value or not is judged to be reached, namely that the dislike degree of the user on a certain content attribute is enough, and the content attribute can be confirmed to be a negative content attribute.

Whether the content attribute is the negative content attribute is judged by taking whether the statistical result (such as the frequency) of the content attribute subjected to the explicit negative feedback behavior reaches a preset threshold value as a condition, compared with the judgment according to the explicit negative feedback behavior for a certain time, the reliability is much higher, and the method is helpful for determining the real reason that the user does not like a certain content.

Specifically, for example, the user makes an explicit negative feedback action of "dislike" for a certain content, but does not explicitly dislike the reason for the content, namely, for which content attribute of the content, each content attribute in the content is considered to accept an explicit negative feedback action. By analogy, the frequency is obtained by the explicit negative feedback times for each content attribute accumulated in the backtracking period, when the frequency reaches a preset threshold value, namely, the meaning that the user may repeatedly dislike a certain content attribute is indicated, the fact that the user does not like the content attribute can be confirmed, namely, the reason of the explicit negative feedback action of the user is clarified.

Of course, the manner of calculating the frequency may not be limited to the above example. In a possible embodiment, for one content attribute, it may be determined whether the content attribute is a negative content attribute on the condition of the following formula (1).

Wherein n represents the number of times that the content related to the content attribute is subjected to explicit negative feedback action of a user in the backtracking period, Δt represents the backtracking period, β is a number of times smoothing parameter, μ is a weight coefficient, and b is a bias term.

To promote more accurate determination of the negative content attributes, in some embodiments the conditions for determining the negative content attributes may be more stringent. Alternatively, the negative content attributes may be filtered based on the condition that they are subject to explicit negative feedback behavior and not subject to positive feedback behavior during the backtracking period. That is, only the number of times of receiving the explicit negative feedback action and not receiving the positive feedback action is counted among the n times in the formula (1), because once the content related to a certain content attribute receives the explicit negative feedback action and then receives the positive feedback action, for example, after the user clicks and dislikes a certain content a having a content attribute M and browses a certain content B having the content attribute M, it may be stated that the content attribute is not actually disliked by the user, but may be linked by other content attributes of the same content that are disliked by the user, and erroneous judgment is generated.

In some possible scenarios, explicit negative feedback actions by a user may not all be accurately trusted. For example, explicit negative feedback behavior (even if there is an explicit cause, it may be caused by the misoperation) generated by the user misoperation, etc. may not conform to the actual interest preference of the user. By excluding these unreliable explicit negative feedback actions, the obtained negative feedback attributes are more accurate, which is beneficial to avoiding the bad effect that the related content of the content attributes which the user may be interested in is shielded due to misoperation of the user.

Accordingly, in some embodiments, the content processing method may further include at least one of:

1) A trusted explicit negative feedback behavior is determined.

In some embodiments, trusted explicit negative feedback behavior may be determined from reference information formed by user selection. For example, the reference information is used to document that the user definitely dislikes a certain content, such as a cause. If the user explicitly expresses reasons for dislike, such as "dislike content source", "dislike author", "dislike pet", etc., the preliminary may be considered as trusted explicit negative feedback. The data of these reasons can be formed by the selection of the displayed options by the user, or can be formed by semantic recognition according to the text and voice input by the user.

2) Explicit negative feedback behavior that is not trusted is corrected or excluded.

In some embodiments, to verify whether the user's current explicit negative feedback behavior is correct, it may be compared to historical explicit negative feedback data for the content for which the explicit negative feedback behavior is intended. Optionally, the first reason may be obtained according to the historical negative feedback data, and whether the first reason is consistent with the second reason of the current explicit negative feedback behavior is checked, if so, the current explicit negative feedback behavior is trusted, or else, the current explicit negative feedback behavior is not trusted.

In some embodiments, the types of the first and second reasons illustratively include at least one of content bias, dislike of content attributes associated with content details, dislike of content sources. Wherein content bias means that the content is not forward, such as hypo, bloodline violence, nausea, etc., is not happy with content attributes related to content details, such as keywords, points of interest, categories, etc. of dislike content by the user, is not happy with content source, such as author, media, etc. of dislike content by the user. It will be appreciated that the type setting of the above reasons may be changed according to the requirements and the types of the content, for example, if the types of the content attribute related to the content details are not favored, sub-types such as dislike keywords, dislike interest points, dislike categories, etc. may be set, or the types such as dislike keywords, dislike interest points, dislike categories, etc. may be directly substituted for dislike the types of the content attribute related to the content details.

In some embodiments, the first cause may be calculated by calculating a ratio of the number of times of the historical explicit negative feedback actions corresponding to at least one cause in the historical explicit negative feedback action data to the total amount of the historical explicit negative feedback actions, and obtaining a comparison result of the ratio and a preset ratio threshold value for estimating the first cause. For example, the number of times of the historical explicit negative feedback actions to which the user group representation is biased is calculated, the ratio of the number of times of the historical explicit negative feedback actions to which the content is biased is 1/5, and whether the ratio reaches a preset ratio threshold, for example 1/6, is judged. Since 1/5>1/6, it means that the first cause is highly likely to be biased. The preset proportion threshold may be preset, or may be one of the proportions corresponding to various reasons related to the content, for example, a minimum value, a median value, a maximum value, etc.

In connection with the above, the types of explicit negative feedback behavior that are not trusted are illustrated.

For example, if an article is disliked by the history of the user population (colloquial, bloody violence, nausea, etc.), the total number of disliked times the article is disliked by the history is relatively low (the ratio is below some preset ratio threshold), and the article is currently fed back as a biased content by the user. Wherein a low historical bias feedback duty cycle indicates that the article may not be biased to content, but the user feedback is biased to content, and the two are not consistent, indicating that the user feedback is not trusted.

For another example, if the number of feedback times that an article is subjected to a historical dislike content source is relatively low in the total number of historical dislike times of the article and is currently fed back as a dislike content source by a user, the two do not coincide, indicating that the user's current feedback is not trusted.

For another example, if an article is subjected to a history of feedback that dislikes a particular content item (including a particular category, keyword, point of interest, etc.), the total number of history dislikes for the article is relatively high and is currently being fed back as a source disliked by a user. Here, a relatively high total number of dislike history of a specific content item indicates that the content is dislike for a more likely reason that the specific content is dislike in detail, and that the dislike of the content by the current user is a dislike content source, and the dislike of the specific content item and the dislike content source are not consistent, which means that the current dislike content source is not credible in behavior.

For the current explicit negative feedback behavior of the user determined to be unreliable, the explicit negative feedback behavior can be eliminated and not used as a basis for analyzing the user preference. Or modified and continued to be employed, e.g., a content bias based on a 90% dislike duty cycle of a historical explicit negative feedback data analysis of an article, and the current explicit negative feedback behavior of the content indicates that the user dislikes the author of the content, then the modification of this current explicit negative feedback behavior to a "content bias" may be considered.

The matching principle in step S302 is explained. If the number of the negative content attributes matched and hit in each content attribute is larger, that is, the number of the negative content attributes contained in the content is larger, the dislike degree of the content by the user is higher. In a possible example, a preset matching condition may be set according to the number of matching hits, e.g. the preset matching condition includes that the number of content attributes of matching hits in the matched content reaches a preset threshold value, etc. Thus, the matched content satisfying the preset matching condition can be filtered.

In a possible example, the content attributes may include, for example, keywords, points of interest, content sources, categories, bias, etc., which may exist in text form. Correspondingly, the matching between the negative content attribute and the content attribute of the matched content can be realized by a text similarity matching mode. In a possible example, since the content attribute may include types such as keywords, points of interest, content sources, categories, bias, etc., the preset threshold may be set comprehensively according to various types of content attributes in the preset matching condition. For example, filtering content for 4 match hits for interest and 2 match hits for category, or filtering content for 5 match hits for keywords, 4 match hits for interest and 2 match hits for category.

For each user as a different individual, a set of negative content attributes is stored correspondingly, and the set of negative content attributes can exist in the form of a negative content attribute library. As the content browsed by the user increases, the negative content attribute in the corresponding negative content attribute library increases, and the occupied data volume increases accordingly. In addition, a situation may occur in which the preference of the user changes. For example, the authors who dislike the original have improved, the points of interest that dislike the original are now interested in again, etc., and, over time, the more old negative content attribute users are more likely to improve. Thus, in some embodiments, rules may be set to update stored negative content attributes, such as to eliminate negative content attributes that the user has improved.

In some embodiments, the elimination of negative content attributes may be triggered based on the attribute occupancy capacity and/or the attribute survival time of the negative content attributes. The attribute occupation capacity refers to storage space occupied by the negative content attribute, and the attribute survival time refers to the existing time of the stored negative content attribute. In a possible example, the process of eliminating the stored negative content attribute may be triggered by the attribute occupancy capacity and/or the attribute survival time reaching a preset threshold as a trigger condition.

In some embodiments, in the eliminating process, the importance of each negative content attribute may be calculated to eliminate the negative content attribute according to the importance, for example, eliminating the negative content attribute with the importance lower than the threshold value, or eliminating one or more negative content attributes with the lowest importance. Alternatively, the importance level may be calculated based on the survival time of the negative content attribute and the matching score. Wherein the matching score is positively correlated with the number of times the negative content attribute is hit by matching, i.e., the more times the negative content attribute is hit by matching, the higher the matching score of the negative content attribute. The importance is inversely related to the survival time and is positively related to the matching score, i.e. the importance of a negative content attribute is higher if the survival time of the negative content attribute is shorter and the number of hits to be matched is greater.

In an exemplary embodiment, the importance may be calculated by the following formula (2):

Where Δt _i is the survival time of the ith element (i.e., the stored negative content attribute), which may be the period from the update time to the current time. Eta is a time parameter, alpha is a weight coefficient, and s _i is a matching score of the ith element.

In some embodiments, the negative content attributes may be categorized into a plurality of types, each of which may correspond to a cause of an explicit negative feedback behavior. For example, the types of reasons may include, for example, dislike content sources, dislike categories, dislike keywords, etc., while the corresponding negative content attributes may be associated under the type of the corresponding reason according to relevance. For example, the types of dislike content sources include negative content attributes such as dislike author A, dislike media B, the categories of dislike include negative content attributes such as dislike make-up, pets, and the types of dislike keywords include negative content attributes such as cactus, sword, etc.

Each cause may be correspondingly configured with an attribute capacity threshold and/or attribute survival time for updating of the negative content attributes of the respective category. That is, the updating (e.g., adding, dropping) of the negative content attribute under the classification of each cause is independent of each other. For example, the capacity, length or number of negative content attributes stored corresponding to the "dislike content source" reaches a preset threshold, and may trigger a process of eliminating negative content attributes under the "dislike content source" without triggering a process of eliminating negative content attributes under the classification of other reasons, such as "dislike category".

In some embodiments, the negative content properties may be updated accordingly by collecting feedback behavior of the user to discover changes in the user's interests. For example, if the user becomes uninteresting about the content of interest, the content attribute may become a negative content attribute and add to the storage, or if the user becomes liked about the content of disliked content, the negative content attribute related to the content will change and need to be removed from the stored set of negative content attributes.

In a possible example, by collecting feedback behavior of the user on the content, if the user is found to be doing positive feedback behavior on the content and is directed to a negative content attribute, the negative content attribute may be removed. The positive feedback behavior is directed to a negative content attribute, and may be directed directly or inferred. For example, the positive feedback behavior of a user directly indicates "like a pet" which is preceded by a negative content attribute that the user does not like, and the "pet" may be removed from the stored negative content attribute set, or the user behavior analysis may be performed with reference to the principle of formula (1), where the positive feedback behavior of a user is "like" for a certain content, and the positive feedback behavior of the user is performed according to the positive feedback behavior of the user for a period of time, so that the number or frequency of positive feedback behaviors for a certain negative content attribute may be counted, and when a preset threshold is reached, the positive feedback behavior of the user is changed, and the negative content attribute may be removed from the stored negative content attribute set.

By updating the stored negative content attributes, such as eliminating the stored excessive (corresponding attribute capacity threshold) and/or old (corresponding attribute survival time) negative content attributes, such as adding the negative content attributes that the user does not like and removing the changed negative content attributes, the content can be filtered more accurately to adapt to the user's preference more accurately.

As also shown in FIG. 4, in some embodiments, the matching calculation and updating of the negative content attribute set 401 in the above-described embodiments may be accomplished by constructing a user negative feedback model 400. The user negative feedback model 400 may access a stored set of negative content attributes. The user negative feedback model 400 matches each content attribute 403 of each matched content 402 with each stored negative content attribute, and when a preset matching condition is satisfied, the corresponding content is filtered.

Optionally, the user negative feedback model 400 may also perform an update of the stored set of negative content attributes 401. In one aspect, the user negative feedback model 400 may also trigger a negative content attribute elimination procedure, schematically illustrated as content attribute 401a being eliminated, based on the attribute occupancy capacity and/or the attribute survival time related triggering conditions. Further alternatively, the elimination procedure of the attribution negative content attribute may be triggered according to the triggering condition corresponding to each cause. On the other hand, the negative feedback model 400 of the user analyzes according to the real-time feedback behavior of the user to add the negative content attribute which is not preferred by the user, and is schematically shown as a new content attribute 401b in the figure, and removes the changed negative content attribute, and is schematically shown as a removed content attribute 401c in the figure.

Referring to fig. 5, a flow chart of step S202 in an embodiment of the disclosure is shown. The flow in fig. 5 includes:

step S501, collecting interactive behavior data of a user on content.

In some embodiments, the interactive behavior data may be recorded data when the user browses the content, where feedback behavior of the user to the content may be included, and the feedback behavior includes at least implicit negative feedback behavior of the user, that is, for example, behavior that skips without browsing the content. The feedback behavior can also comprise positive feedback behavior, explicit negative feedback behavior and the like of the user, and the preference is analyzed by collecting various types of interaction behavior of the user, so that the sorting weight of the content can be adjusted more accurately according to the preference. Labels corresponding to the content, such as categories of content, keywords, content sources, whether biased, etc., may also be determined. In addition, optionally, the interactive behavior data may further include content attributes of the content interacted with by the user, such as content tags, including categories of the content, keywords, content sources, whether the content is biased, and the like.

Taking contents such as image text articles and videos as examples, the ID of the content recommended to the user for display in the last period, the content attribute of the corresponding content and the interactive behavior data of the user on the display content can be collected. Feedback behavior of the user on the presentation content, for example, real-time feedback behavior of the recommended information content, may be obtained from the interactive behavior data, and the feedback behavior may include positive feedback behavior, such as click reading, praise, sharing, and the like, for example. Illustratively, the feedback behavior may include explicit negative feedback behavior, such as clicking on dislike articles, clicking on dislike authors, clicking on keyword labels on dislike articles. Optionally, some feature information of the user interaction behavior, such as the duration of the display exposure of the recommended information content, the reading duration of the user, etc., may be obtained through the interaction behavior data, and may be used to assist in, for example, judging the positive degree, the negative degree, etc., of the content expressed by the user feedback behavior.

Step S502, calculating preference weight coefficients of the targeted content attributes according to the collected interaction behavior data.

The preference weight coefficient represents a probability of preference of the user for each content attribute. Wherein the content attributes vary from recommended content to recommended content. For example, the content attributes of the information class may include content source, category, keyword, bias (e.g., whether popular, violent, horror, etc.), article, short video, etc.

The preference weight coefficient reflects the preference degree of the user on the content attribute, wherein the preference degree comprises the influence of implicit negative feedback behavior of the user. The preference weight coefficient can act on the sorting score of the content attribute in the rearrangement link or the fine-ranking link, so that the sorting weight of the content attribute at least subjected to the implicit negative feedback action of the user in the rearrangement link or the fine-ranking link is reduced, and the position of the corresponding content in the sorting can be back or even filtered out, so that the content cannot appear in the final recommended content.

In some embodiments, a preference value of the user for the content attribute targeted by the feedback behavior may be calculated according to the feedback behavior, and then the preference value is mapped to a preset value range to obtain the preference weight coefficient.

In some embodiments, the preference value may be calculated by a preset formula. The preset formula calculates the sum of two items, wherein one item is a preference item for representing user preference, and the other item is a heuristic item for representing heuristic user preference.

The preference may relate to the positive feedback behavior, the negative feedback behavior of the user on the relevant content of the content property. The more frequent the positive feedback behavior, the stronger the user preference, and the more frequent the negative feedback behavior, the weaker the user preference. In addition, the user's preferences for content may also change over time, and the older feedback behavior may be less reliable. In a possible implementation example, the value of the preference is positively correlated with a positive feedback index, the positive feedback index including the result of the summation of the behavior preference weights of the user under the effect of the time decay of the occurrence moment of the user, the value of the behavior preference weights being positively correlated with the positive degree and negatively correlated with the negative degree of the interaction behavior, the time decay effect being positively attenuated along the time axis.

The heuristics may relate to exposure situations where content is presented to the user. The more exposure of the content to the user is not browsed, the less preferred the user is. In a possible implementation example, the exposure condition may be represented by an exposure index, where the value of the heuristic is inversely related to the exposure index, and the exposure index includes a summation result of each exposure of the related content of the content attribute under a time attenuation effect of the exposure time, and the time attenuation effect is attenuated positively along a time axis.

Illustratively, the following formula (3) provides an example of a preset formula for calculating the preference value. Formula (3) is only one implementation, and is not limiting.

Wherein the left term of the plus sign is a preference and the right term of the plus sign is a heuristic. p represents a forward feedback index, and is obtained by weighting feedback behaviors of the content on a time axis according to a user. And multiplying each weighted term by a time attenuation coefficient obtained according to the time difference between the behavior occurrence time and the current time, and finally obtaining a forward feedback index by weighted summation. The positive feedback behavior can take positive values and give values of different magnitudes according to different positive degrees, and the negative feedback behavior can take negative values and give values of different magnitudes according to different negative degrees. For example, when a user has a reading behavior weight of 1 for content B related to content attribute A at time 1 which is 1 hour away from the current time, the time attenuation coefficient of time 1 is 0.1, when the user has a dislike behavior weight of-10 for content C related to content attribute A at time 2 which is 30 minutes away from the current time, the time attenuation coefficient of time 2 is 0.2, when the user has a sharing behavior weight of 2 for content D related to content attribute A at time 3 which is 15 minutes away from the current time, the time attenuation coefficient of time 3 is 0.4, and the smaller the value of the time attenuation coefficient is, the larger the attenuation effect is. In this example, p=1×0.1+ (-10) 0.2+2×0.4= -1.1 can be calculated.

E represents the exposure index, and corresponds to the summation of the time attenuation coefficients obtained by the exposure time of the content attribute. For example, the contents are exposed at time 1, time 2, and time 3, which are 1 hour, 30 minutes, and 15 minutes from the current time, and the time attenuation coefficients are 0.1, 0.2, and 0.4, respectively, and the sum is 0.7. The time attenuation coefficients used in the calculation of p and e may be the same or different.

Alpha and beta are bias in preference items, the values of the alpha and beta can be different corresponding to different content attributes, and gamma and lambda are harmonic coefficients between preference items and heuristics, such as gamma=0.4, lambda=500 and the like.

In some embodiments, in addition to calculating the preference value by way of a preset formula, the preference value for the content attribute may also be predicted by a trained machine learning model. In a possible example, the machine learning model may employ a supervised classification approach to classify corresponding content attributes to predict the probability of user preference over different content attributes as a preference value. Illustratively, machine learning models that may be employed include, for example, gradient-lifting decision trees (Gradient Boosting Decision Tree, GBDT), deep neural networks, and the like.

As shown in fig. 6, a flow diagram of a training method of a machine learning model in one embodiment of the present disclosure is shown. The machine learning model may be used to predict preference values for content attributes. The training method comprises the following steps:

step S601, a training sample is obtained, wherein the training sample comprises user characteristics and label information.

In some embodiments, the user characteristics include user self characteristics extracted from network activities of the user and user attribute preference characteristics extracted from user interaction data of content to which the content attributes pertain, the user attribute preference characteristics being related to the content attributes. Illustratively, the user's own characteristics may include, but are not limited to, at least one of whether the user is active, the last session duration, the number of sessions on the day, the user's recent click-through rate, the user's recent negative feedback rate, and so forth. Illustratively, the user attribute preference feature may include, but is not limited to, at least one of a number of exposures to related content of a content attribute, a number of clicks, a number of negative feedback actions, a browsing duration, and the like.

The tag information is used for indicating that the content attribute subjected to the positive feedback action of the user is a positive sample and indicating that the content attribute subjected to the negative feedback action of the user is a negative sample. That is, the tag information includes two kinds, first information representing a positive sample and second information representing a negative sample, for example, 0 and 1.

Step S602, training the machine learning model by using the training sample, so that the machine learning model is used for predicting the probability value of the content attribute belonging to the positive sample according to the input user characteristic as the preference value.

The machine learning model can be configured with a loss function, the output error of the machine learning model is measured through the loss function, and parameters of the machine learning model are adjusted until preset training conditions are met, such as the output error is reduced below a preset threshold value or the training times reach preset times, and the like, so that the machine learning model indicates that training is completed and is applied to a prediction scene, and the preference value of the user in each classification content attribute is predicted according to the input of the current user characteristics.

In some embodiments, the content attributes that do not require suppression of the ranking score may be screened out using a preset screening rule before the preference weight coefficient is calculated from the preference value. The preset screening rule comprises at least one of screening out content attributes with preference values larger than a preset threshold value, screening out content attributes with exposure times lower than the preset threshold value, and screening out content attributes with interval duration of the latest exposure time from the current moment exceeding the preset threshold value. Specifically, content attributes with preference values greater than a preset threshold value are preferable to users, and the sorting score of the content attributes is not required to be suppressed, so that the content attributes are screened out. The exposure times lower than the preset threshold value indicate that the exposure times are less, the content attribute with the less exposure times cannot be recommended without inhibiting the sorting score, and the content attribute can be screened out. The interval duration of the latest exposure time from the current moment exceeds a preset threshold value, which indicates the old content attribute, can not be recommended without inhibiting the sorting score, and can be screened out. It can be understood that different content recommendation occasions can have different screening rules and parameter settings, and examples of some screening rules are given here corresponding to information browsing scenes such as text and video, but not limited to these.

In some embodiments, the preference values may be mapped into a desired range of preset values to obtain preference weight coefficients. The mapping method may be a linear mapping method, for example. The preference weight coefficient can be obtained by mapping from the preference value by the following example of the formula (4).

Wherein θ is a preference weight coefficient, α is a maximum value of the preference weight coefficient, b is a minimum value of the preference weight coefficient, ε is a parameter value for setting an initial threshold value of θ, P is a preference value, and α, b, ε may change corresponding to different content attributes, for example, one content attribute, α=1, b=0.4, ε=0.05, and so on.

It should be noted that, in a possible example, the preference weight coefficient may also be obtained directly by analyzing the interaction data of the user on the content, without generating the preference value. For example, the probability value predicted by the machine learning model is directly used as a preference weight coefficient or the like.

Returning to fig. 5, step S503 is to adjust the ranking weight of the content attribute in the fine-ranking link according to the preference weight coefficient.

In some embodiments, the greater the preference weight coefficient of a content attribute, the greater the degree of preference, the lower the suppression of the ranking weight should be, i.e., the preference weight coefficient of a content attribute should be inversely related to its suppression of the ranking weight. Thus, for the calculation of the adjustment, a linear weighting algorithm of the following formula (5) may be used, for example:

Wherein s is the fine row score before adjustment, And (3) scoring the adjusted fine row, wherein θ is the preference weight coefficient, sigma is a retention score bias corresponding to the content attribute, and different content attributes can be set with different sigma values.

In summary, the embodiment of the disclosure filters the content related to the explicit negative feedback behavior of the user through the filtering link in the recommendation calculation flow of the recommendation system, and suppresses the ranking weight of the content related to at least the implicit negative feedback behavior of the user in the fine ranking link. Therefore, the method and the device realize the targeted filtering and suppression of the content related to the explicit and implicit type negative feedback behaviors of the user in the calculation process of the recommendation system, effectively reduce the unfavorable and uninteresting parts of the user in the recommendation content, effectively ensure the accuracy of the recommendation system and effectively improve the user experience.

According to the embodiment of the disclosure, the content filtering based on the explicit negative feedback behavior is introduced in the recommendation system filtering link, so that the experience problem directly related to the explicit negative feedback behavior of the user is solved, the content attribute corresponding to the dislike content of the user can be shielded in time, and the user can have direct interactive control on the recommendation system. Through experiments, the explicit negative feedback processing experiment is used in a certain sexualized news recommended product, and the comparison effect test shows that the dislike rate of the user is reduced by about 28%. In addition, sequencing score inhibition based on at least implicit negative feedback behavior is introduced in a rearrangement link of the recommendation system, so that the recommendation quantity of the content which is not interested by the user at present is controllably reduced, feedback sensitivity to different contents is controllably set, and the experience of using recommendation by the user is improved. Through experiments, implicit negative feedback processing experiments are used in a certain sexualized news recommended product, and comparison effect tests show that the dislike rate of users is reduced by about 10%, and the average residence time is also stably improved. Through the combined use of the two approaches, the dislike rate of the user on the recommended content can be further reduced.

Exemplary apparatus

Having introduced a content processing method of an exemplary embodiment of the present disclosure, next, a content processing apparatus of an exemplary embodiment of the present disclosure is described with reference to fig. 7.

Referring to fig. 7, a content processing apparatus 700 according to an exemplary embodiment of the present disclosure may include an explicit negative feedback processing module 701 configured to filter content related to an explicit negative feedback behavior of a user in the filtering link, and an implicit negative feedback processing module 702 configured to suppress a ranking weight of at least content related to the implicit negative feedback behavior of the user in the fine-ranking link. Therefore, the method and the device realize the targeted filtering and suppression of the content related to the explicit and implicit type negative feedback behaviors of the user in the calculation process of the recommendation system, effectively reduce the unfavorable and uninteresting parts of the user in the recommendation content, effectively ensure the accuracy of the recommendation system and effectively improve the user experience.

In some embodiments, the explicit negative feedback processing module 701 includes a user negative feedback behavior processing module, configured to determine a negative content attribute from content attributes targeted by the explicit negative feedback behavior and store the negative content attribute, a matching filtering module, configured to match the negative content attribute with a content attribute of a matched content to obtain a matching result, and filter the matched content to which the content attribute of the matching result meets a preset matching condition.

In some embodiments, the user negative feedback behavior processing module includes a user behavior analysis module configured to filter the negative content attribute based on a condition that a frequency of explicit negative feedback behavior received in a backtracking period reaches a preset threshold.

In some embodiments, the user negative feedback behavior processing module is further configured to filter the negative content attribute based on a condition that the explicit negative feedback behavior is received and the positive feedback behavior is not received in the backtracking period.

In some embodiments, the content processing device 700 further includes at least one of the following modules:

the behavior credibility determining module is used for determining credible explicit negative feedback behavior;

and the behavior untrusted processing module is used for correcting or eliminating the untrusted explicit negative feedback behavior.

In some embodiments, the behavior credibility determination module is configured to determine a credible explicit negative feedback behavior according to reference information formed by user selection.

In some embodiments, the behavioral untrusted processing module includes:

The system comprises a behavior unreliable determining module, a behavior non-trusted determining module and a non-trusted determining module, wherein the behavior unreliable determining module is used for determining an unreliable explicit negative feedback behavior according to historical explicit negative feedback behavior data of content aimed by the explicit negative feedback behavior, and the unreliable explicit negative feedback behavior means that a first reason presumed according to the historical explicit negative feedback behavior data is not consistent with a second reason of the explicit negative feedback behavior;

and the behavior untrusted processing execution module is used for correcting the explicit negative feedback behavior until the second reason accords with the first reason or discarding the explicit negative feedback behavior.

In some embodiments, the behavior judgment module is configured to calculate a ratio of the number of times of the historical explicit negative feedback behavior corresponding to at least one cause in the historical explicit negative feedback behavior data to the total amount of the historical explicit negative feedback behavior, and obtain a comparison result of the ratio and a preset ratio threshold value to be used for estimating the first cause.

In some embodiments, the types of the first and second reasons include at least one of content bias, dislike for content attributes related to content details, dislike for content sources.

In some embodiments, the content processing device 700 further includes:

The attribute elimination module is used for triggering and executing an elimination process of the stored negative content attribute in response to the attribute occupation capacity and/or the attribute survival time reaching a preset threshold, and comprises the steps of calculating importance according to the survival time and the matching score of the negative content attribute, wherein the importance is in negative correlation with the survival time and in positive correlation with the matching score, the matching score is in positive correlation with the number of times that the negative content attribute is matched and hit, and the negative content attribute is eliminated according to the importance.

In some embodiments, each of the negative content attributes is classified into a plurality of types, each type corresponding to one cause of explicit negative feedback behavior, each classification corresponding to an attribute capacity threshold and/or attribute survival time configured for updating of the negative content attributes of the respective classification.

In some embodiments, the content processing apparatus 700 includes:

and the attribute adjustment module is used for removing negative content attributes subjected to positive feedback behaviors from the stored space.

In some embodiments, the preset matching condition includes the number of content attributes of the matching hit of the matched content reaching a preset threshold.

In some embodiments, the implicit negative feedback processing module 702 includes a user behavior collection module configured to collect interaction behavior data of a user on content, where the interaction behavior includes feedback behavior, the feedback behavior includes at least implicit negative feedback behavior, an attribute preference calculation module configured to calculate a preference weight coefficient of a targeted content attribute according to the collected interaction behavior data, and a score adjustment module configured to adjust a ranking weight of the content attribute in the thin-row link according to the preference weight coefficient, where a value of the preference weight coefficient of the content attribute is affected by the corresponding implicit negative feedback behavior to reduce the ranking weight of the content attribute in the thin-row link so as to suppress occurrence of related content in the generated recommended content.

In some embodiments, the attribute preference calculation module comprises an attribute preference weight coefficient calculation module for calculating a preference value of the user for the content attribute aimed at by the interaction behavior, and a weight mapping module for mapping the preference value to a preset value range to obtain the preference weight coefficient.

In some embodiments, the attribute preference weight coefficient calculation module comprises a calculation module, wherein the calculation module is used for calculating the preference value through a preset calculation formula, the preset calculation formula is expressed as the sum of a preference item for representing user preference and a heuristic item for representing heuristic user preference, the value of the preference item is positively related to a positive feedback index, the positive feedback index comprises the summation result of the behavior preference weight of each interaction behavior of a user under the action of time attenuation of the occurrence moment of the user, the value of the behavior preference weight is positively related to the positive degree and negatively related to the negative degree of the interaction behavior, the value of the heuristic item is negatively related to an exposure index, the exposure index comprises the summation result of each exposure of related content of the content attribute under the action of time attenuation of the exposure moment of the content attribute, and the time attenuation action is positively attenuated along a time axis.

In some embodiments, the attribute preference calculation module includes a model calculation module for predicting a preference value or preference weight coefficient of the user for the content attribute based on the user characteristics through a trained machine learning model.

In some embodiments, the content processing apparatus 700 includes a model training module for training the machine learning model, comprising:

The system comprises a sample acquisition module, a sample acquisition module and a processing module, wherein the sample acquisition module is used for acquiring training samples, the training samples comprise user characteristics and label information, the user characteristics comprise user self characteristics extracted according to network activities of users and user attribute preference characteristics related to content attributes extracted according to user interaction data of content to which the content attributes belong, the label information is used for indicating that the content attributes subjected to positive feedback behaviors of the users are positive samples and the content attributes subjected to negative feedback behaviors of the users are negative samples;

And the training execution module is used for training the machine learning model by using the training sample so that the machine learning model is used for predicting the probability value of the positive sample of the content attribute according to the input user characteristic as a preference value or a preference weight coefficient.

In some embodiments, the content processing apparatus 700 further includes an attribute filtering module configured to perform filtering out the content attribute using a preset filtering rule before the weight mapping module calculates the preference weight coefficient according to the preference weight coefficient of the content attribute, where the preset filtering rule includes at least one of filtering out the content attribute having the preference weight coefficient or the preference weight coefficient greater than a preset threshold, filtering out the content attribute having the exposure frequency lower than the preset threshold, and filtering out the content attribute having the interval duration of the latest exposure time from the current time exceeding the preset threshold.

Since the respective functional modules of the content processing apparatus 700 of the embodiment of the present disclosure are the same as the corresponding steps of the content processing method in the above embodiment, the description thereof will not be repeated here.

Exemplary storage Medium

Having described the content processing method and the content processing apparatus of the exemplary embodiment of the present disclosure, next, a storage medium of the exemplary embodiment of the present disclosure will be described with reference to fig. 8.

Referring to fig. 8, a storage medium 800 according to an embodiment of the present disclosure is described, which may contain program code and may be run on a device, such as a personal computer, to effect execution of the various steps and sub-steps in the processing methods of the present disclosure. In this document, a readable storage medium may be any tangible medium that can contain, or store a program for use by or in connection with an instruction execution system, apparatus, or device.

The program code can employ any combination of one or more readable media. The readable medium may be a readable signal medium or a readable storage medium. The readable storage medium can be, for example, but is not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or a combination of any of the foregoing. More specific examples (a non-exhaustive list) of a readable storage medium include an electrical connection having one or more wires, a portable disk, a hard disk, a Random Access Memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disc read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing.

The computer readable signal medium may include a data signal propagated in baseband or as part of a carrier wave with readable program code embodied therein. Such a propagated data signal may take any of a variety of forms, including, but not limited to, electro-magnetic, optical, or any suitable combination of the foregoing. A readable signal medium may also be any readable medium that is not a readable storage medium and that can communicate, propagate, or transport a program for use by or in connection with an instruction execution system, apparatus, or device.

Program code embodied on a readable medium may be transmitted using any appropriate medium, including but not limited to wireless, wireline, optical fiber cable, RF, etc., or any suitable combination of the foregoing.

Program code for carrying out operations of the present disclosure may be written in any combination of one or more programming languages, including an object oriented programming language such as Java, C++ or the like and conventional procedural programming languages, such as the "C" programming language or similar programming languages. The program code may execute entirely on the user's computing device, partly on a remote computing device, or entirely on the remote computing device or server. In the case of remote computing devices, the remote computing device may be connected to the user computing device through any kind of network, including a Local Area Network (LAN) or a Wide Area Network (WAN), or may be connected to an external computing device (e.g., connected via the Internet using an Internet service provider).

Exemplary electronic device

Having described the storage medium of the exemplary embodiments of the present disclosure, next, an electronic device of the exemplary embodiments of the present disclosure will be described with reference to fig. 9.

The electronic device 900 shown in fig. 9 is merely an example and should not be construed to limit the functionality and scope of use of embodiments of the present disclosure in any way.

As shown in fig. 9, the electronic device 900 is embodied in the form of a general purpose computing device. The components of the electronic device 900 may include, but are not limited to, the at least one processing unit 910 described above, the at least one storage unit 920 described above, and a bus 930 that connects the different system components (including the storage unit 920 and the processing unit 910).

Wherein the storage unit stores program code executable by the processing unit 810 such that the processing unit 910 performs the steps and sub-steps of the content processing method described in the above embodiments of the present disclosure. For example, the processing unit 910 may perform the steps as shown in fig. 2,3, 5, etc.

In some embodiments, the storage unit 920 may include volatile storage units, such as a random access memory unit (RAM) 9201 and/or a cache memory unit 9202, and may further include a read only memory unit (ROM) 9203.

In some embodiments, storage unit 920 may also include a program/utility 9204 having a set (at least one) of program modules 9205, such program modules 9205 including, but not limited to, an operating system, one or more application programs, other program modules, and program data, each or some combination of which may include an implementation of a network environment.

In some embodiments, bus 930 may include a data bus, an address bus, and a control bus.

In some embodiments, electronic device 900 may also communicate with one or more external devices 1000 (e.g., keyboard, pointing device, bluetooth device, etc.) via an input/output (I/O) interface 950. Optionally, the electronic device 900 further comprises an explicit unit 940 connected to an input/output (I/O) interface 950 for performing explicit. Also, electronic device 900 may communicate with one or more networks such as a Local Area Network (LAN), a Wide Area Network (WAN), and/or a public network, such as the Internet, through network adapter 960. As shown, the network adapter 960 communicates with other modules of the electronic device 900 over the bus 930. It should be appreciated that although not shown, other hardware and/or software modules may be used in connection with electronic device 900, including, but not limited to, microcode, device drivers, redundant processing units, external disk drive arrays, RAID systems, tape drives, data backup storage systems, and the like.

It should be noted that although in the above detailed description several modules or sub-modules of the recognition model generating means and the recognition means are mentioned, such a division is only exemplary and not mandatory. Indeed, the features and functionality of two or more units/modules described above may be embodied in one unit/module in accordance with embodiments of the present disclosure. Conversely, the features and functions of one unit/module described above may be further divided into ones that are embodied by a plurality of units/modules.

Furthermore, although the operations of the methods of the present disclosure are depicted in the drawings in a particular order, this is not required or suggested that these operations must be performed in this particular order or that all of the illustrated operations must be performed in order to achieve desirable results. Additionally or alternatively, certain steps may be omitted, multiple steps combined into one step to perform, and/or one step decomposed into multiple steps to perform.

While the spirit and principles of the present disclosure have been described with reference to several particular embodiments, it is to be understood that this disclosure is not limited to the particular embodiments disclosed nor does it imply that features in these aspects are not to be combined to benefit from this division, which is done for convenience of description only. The disclosure is intended to cover various modifications and equivalent arrangements included within the spirit and scope of the appended claims.

Claims

1. A content processing method is applied to a recommendation system, wherein a recommendation calculation flow of the recommendation system comprises a coarse ranking link, a filtering link and a fine ranking link of contents, and the method comprises the following steps:

in the fine ranking link, suppressing ranking weights of content at least related to the implicit negative feedback behavior of the user;

wherein the suppressing at least the ranking weight of the content related to the implicit negative feedback behavior of the user comprises:

the method comprises the steps of collecting interactive behavior data of a user on content, wherein the interactive behavior comprises feedback behavior which at least comprises implicit negative feedback behavior;

Calculating preference weight coefficients of the targeted content attributes according to the collected interaction behavior data;

the method comprises the steps of adjusting the ordering weight of the content attribute in the fine-ranking link according to the preference weight coefficient, wherein the value of the preference weight coefficient of the content attribute is influenced by the corresponding implicit negative feedback behavior to reduce so as to inhibit the ordering weight of the content attribute in the fine-ranking link and inhibit the occurrence of related content in the generated recommended content;

Wherein the calculating the preference weight coefficient of the targeted content attribute according to the collected interaction behavior data comprises:

Calculating a preference value of the user for the content attribute targeted by the interaction behavior through a preset formula, mapping the preference value to a preset value range to obtain a preference weight coefficient, wherein the preset formula is expressed as a sum of a preference item for representing the user preference and a heuristic item for representing the heuristic user preference, the value of the preference item is positively correlated with a positive feedback index, the positive feedback index comprises a summation result of the behavior preference weight of each interaction behavior of the user under the time attenuation action of the occurrence moment of the interaction behavior, the value of the behavior preference weight is positively correlated with the positive degree and negatively correlated with the negative degree of the interaction behavior, the value of the heuristic item is negatively correlated with an exposure index, the exposure index comprises a summation result of each exposure of the related content of the content attribute under the time attenuation action of the exposure moment of the content attribute, the time attenuation action is positively attenuated along a time axis, or

The method comprises the steps of predicting a preference value or a preference weight coefficient of a user on a content attribute according to user characteristics through a trained machine learning model, training the machine learning model, wherein the training of the machine learning model comprises the steps of obtaining training samples, the training samples comprise user characteristics and label information, the user characteristics comprise user attribute preference characteristics extracted according to network activities of the user and user interaction data of content to which the content attribute belongs, the user attribute preference characteristics are extracted according to user interaction data of the content to which the content attribute belongs, the label information is used for indicating that the content attribute subjected to positive feedback action of the user is a positive sample and the content attribute subjected to negative feedback action of the user is a negative sample, and training the machine learning model through the training samples, so that the machine learning model is used for predicting the probability value of the positive sample of the content attribute according to the input user characteristics and is used as the preference value or the preference weight coefficient.

2. The content processing method according to claim 1, wherein the filtering the content related to the explicit negative feedback behavior of the user comprises:

determining negative content attributes from all the content attributes aimed by the explicit negative feedback behavior, and storing the negative content attributes;

matching the negative content attribute with the content attribute of the matched content to obtain a matching result;

and filtering matched content to which the content attribute of the matching result meets the preset matching condition belongs.

3. The content processing method according to claim 2, wherein the determining a negative content attribute among the content attributes for which the explicit negative feedback behavior is directed includes:

And screening the negative content attribute based on the condition that the frequency of the explicit negative feedback behavior in the backtracking period reaches a preset threshold.

4. The content processing method according to claim 3, wherein the determining a negative content attribute among the content attributes for which the explicit negative feedback behavior is directed further comprises:

The negative content attributes are filtered based on the condition that explicit negative feedback behavior is experienced and positive feedback behavior is not experienced in the backtracking period.

5. A content processing method according to claim 1 or 3, further comprising at least one of:

1) Determining a trusted explicit negative feedback behavior;

6. The content processing method according to claim 5, wherein said determining trusted explicit negative feedback behavior comprises:

a trusted explicit negative feedback behavior is determined from the reference information formed by the user selection.

7. The content processing method according to claim 5, wherein the correcting or excluding the untrusted explicit negative feedback behavior includes:

Determining an unreliable explicit negative feedback behavior according to historical explicit negative feedback behavior data of content targeted by the explicit negative feedback behavior, wherein the unreliable explicit negative feedback behavior refers to that a first reason presumed according to the historical explicit negative feedback behavior data does not coincide with a second reason of the explicit negative feedback behavior;

correcting the explicit negative feedback behavior to coincide with the historical explicit negative feedback behavior data or excluding the explicit negative feedback behavior.

8. The content processing method according to claim 7, wherein the estimating the first cause from historical explicit negative feedback behavior data comprises:

and calculating the proportion of the times of the historical explicit negative feedback behaviors corresponding to at least one reason in the historical explicit negative feedback behavior data to the total amount of the historical explicit negative feedback behaviors, and acquiring a comparison result of the proportion and a preset proportion threshold value for presuming the first reason.

9. The content processing method according to claim 7, wherein the types of the first cause and the second cause include at least one of content bias, dislike of a content attribute related to content details, dislike of a content source.

10. The content processing method according to claim 2, characterized by further comprising:

and in response to the attribute occupancy capacity and/or the attribute survival time reaching a preset threshold, triggering execution of a process of eliminating the stored negative content attribute, including:

Calculating importance according to the survival time of the negative content attribute and the matching score, wherein the importance is in negative correlation with the survival time and in positive correlation with the matching score;

and eliminating the negative content attribute according to the importance degree.

11. The content processing method according to claim 10, wherein each of the negative content attributes is classified into a plurality of types, each type corresponding to one cause of explicit negative feedback behavior, each classification corresponding to an update of the negative content attribute configured with an attribute capacity threshold and/or an attribute survival time for the respective classification.

12. The content processing method as recited in claim 2, further comprising removing negative content attributes subject to positive feedback behavior from the stored space.

13. The content processing method according to claim 2, wherein the preset matching condition includes that the number of content attributes hit by the matching of the matched content reaches a preset threshold.

14. The content processing method according to claim 1, further comprising, before said mapping said preference value to a preset value range to obtain said preference weight coefficient:

Screening the content attribute by using a preset screening rule;

the preset screening rule comprises at least one of screening out content attributes with preference weight coefficients or preference weight coefficients larger than a preset threshold value, screening out content attributes with exposure times lower than the preset threshold value, and screening out content attributes with interval duration of the latest exposure time from the current moment exceeding the preset threshold value.

15. The content processing device is characterized by being applied to a recommendation system, wherein the calculation flow of the recommendation system comprises a coarse ranking link, a filtering link and a fine ranking link of contents, and the device comprises:

the implicit negative feedback processing module is used for inhibiting the ordering weight of the content at least related to the implicit negative feedback behavior of the user in the fine ranking link;

the implicit negative feedback processing module comprises:

the system comprises a user behavior collection module, a content processing module and a content processing module, wherein the user behavior collection module is used for collecting interactive behavior data of a user on content, and the interactive behavior comprises feedback behaviors which at least comprise implicit negative feedback behaviors;

The attribute preference calculation module is used for calculating a preference weight coefficient of the aimed content attribute according to the collected interaction behavior data;

The score adjustment module is used for adjusting the ordering weight of the content attribute in the fine-ranking link according to the preference weight coefficient, wherein the value of the preference weight coefficient of the content attribute is reduced by the influence of the corresponding implicit negative feedback behavior so as to inhibit the ordering weight of the content attribute in the fine-ranking link and inhibit the occurrence of related content in the generated recommended content;

Wherein the attribute preference calculation module is configured to:

16. The content processing apparatus of claim 15, wherein the explicit negative feedback processing module comprises:

The user negative feedback behavior processing module is used for determining and storing negative content attributes in the content attributes aimed at by the explicit negative feedback behavior;

and the matching filtering module is used for matching the negative content attribute with the content attribute of the matched content to obtain a matching result, and filtering the matched content of which the content attribute of the matching result meets the preset matching condition.

17. The content processing apparatus of claim 16, wherein the user negative feedback behavior processing module comprises:

And the user behavior analysis module is used for screening the negative content attribute based on the condition that the frequency of the explicit negative feedback behavior received in the backtracking period reaches a preset threshold.

18. The content processing apparatus of claim 17, wherein the user negative feedback behavior processing module is further configured to filter the negative content attribute based on a condition that explicit negative feedback behavior is received and positive feedback behavior is not received during the backtracking period.

19. The content processing apparatus according to claim 15 or 17, further comprising at least one of the following modules:

20. The content processing apparatus of claim 19, wherein the behavior trusted determination module is configured to determine trusted explicit negative feedback behavior based on reference information formed by user selection.

21. The content processing apparatus according to claim 16, wherein the behavior untrusted processing module comprises:

22. The apparatus according to claim 21, wherein the behavior judgment module is configured to calculate a ratio of a number of times of the historical explicit negative feedback behavior corresponding to at least one cause in the historical explicit negative feedback behavior data to a total amount of the historical explicit negative feedback behavior, and obtain a comparison result of the ratio with a preset ratio threshold for estimating the first cause.

23. The content processing apparatus of claim 21, wherein the types of the first and second reasons include at least one of content bias, dislike of content attributes related to content details, dislike of content sources.

24. The content processing apparatus according to claim 23, further comprising:

the attribute elimination module is configured to trigger execution of an elimination procedure for the stored negative content attribute in response to the attribute occupation capacity and/or the attribute survival time reaching a preset threshold, and includes:

25. The content processing apparatus of claim 24, wherein each of the negative content attributes is classified into a plurality of types, each type corresponding to one cause of explicit negative feedback behavior, each classification corresponding to an update of the negative content attribute configured with an attribute capacity threshold and/or attribute survival time for the respective classification.

26. The content processing apparatus according to claim 16, comprising:

27. The content processing apparatus according to claim 16, wherein the preset matching condition includes that the number of content attributes hit by the matching of the matched content reaches a preset threshold.

28. The content processing apparatus according to claim 15, further comprising an attribute filtering module configured to perform, before the weight mapping module calculates a preference weight coefficient from a preference weight coefficient of the content attribute, screening the content attribute using a preset screening rule, wherein the preset screening rule includes at least one of screening the content attribute having a preference weight coefficient or a preference weight coefficient greater than a preset threshold, screening the content attribute having an exposure number lower than the preset threshold, and screening the content attribute having an interval duration of a latest exposure time from a current time exceeding the preset threshold.

29. A storage medium having stored thereon a computer program, the computer program realizing when executed by a processor:

The content processing method according to any one of claims 1 to 14.

30. An electronic device, comprising:

Processor, and

A memory for storing executable instructions of the processor;

The content processing method according to any one of claims 1 to 14.