CN108256881A - A kind of traffic filtering method and apparatus - Google Patents

Abstract

The invention discloses a flow filtering method and device, and relates to the technical field of computers. The method includes: receiving each traffic sent by the advertising exchange server; extracting traffic characteristics from each traffic; using a preset revenue calculation model to calculate the first value data corresponding to the traffic characteristics; the revenue calculation model according to each category The target value data of the traffic and the characteristics of each traffic in the historical traffic records and the corresponding actual value data and consumption income data are obtained through training; each traffic in the historical traffic records is the traffic that has received the notification of successful bidding from the advertising exchange server; the judgment of the first Whether the value data is less than the first threshold; when the first value data is less than the first threshold, the traffic is filtered. It solves the problems of waste of DSP system resources caused by the existing flow distribution mechanism and reduces the cost performance of DSP, and achieves the beneficial effects of filtering the traffic before sending the traffic to each DSP, reducing the waste of DSP system resources, and improving the cost performance.

Description

A flow filtering method and device

technical field

The invention relates to the field of computer technology, in particular to a flow filtering method and device.

Background technique

RTB (Real Time Bidding, real-time bidding) is a bidding technology that uses third-party technology to evaluate and bid for each user's display behavior on millions of websites. Different from the frequency of mass purchases, real-time bidding avoids invalid audience arrival and makes purchases for meaningful users. Its core is DSP (Demand Site Platform, demand-side platform). For the media, RTB can bring more advertising sales, realize sales process automation and reduce various expenses. Moreover, with the rapid popularization of smart phones and the maturity of the mobile network environment, the mobile DSP market will be further explored, and programmatic buying will become a major trend in the digital marketing era.

In the Internet advertising market, a DSP can monitor a large amount of traffic in real time, but not all traffic can generate enough revenue. Generally speaking, only 10%-20% of the traffic monitored by a DSP can generate sufficient revenue. However, for the flow of which the final income is not good, the existing DSP will still execute all the calculation processes, which will lead to a waste of DSP system resources and reduce the cost performance of the DSP.

Contents of the invention

In view of the above problems, the present invention is proposed to provide a traffic filtering method and a corresponding traffic filtering device that overcome the above problems or at least partially solve the above problems.

According to one aspect of the present invention, a traffic filtering method is provided, including:

Receive all traffic sent by the advertising transaction server;

Extract flow features from each flow;

Using the preset revenue calculation model to calculate the first value data corresponding to the traffic characteristics; the revenue calculation model is based on the target value data of each category of traffic and the characteristics of each traffic in the historical traffic records and the corresponding actual value data and Obtained through consumption revenue data training; each traffic in the historical traffic records is the traffic that received the successful bidding notification from the advertising exchange server;

judging whether the first value data is smaller than a first threshold;

When the first value data is less than a first threshold, the traffic is filtered.

Optionally, the step of filtering the traffic includes:

determine the category to which the traffic belongs;

Filter the traffic of the category according to the specified ratio.

Optionally, after the step of filtering the traffic of the category according to a specified ratio, the method further includes:

Use the remaining traffic after filtering the traffic of the category according to the specified ratio as the sampling traffic;

Obtain the sampling records of each sampling traffic that receives the successful bidding notification from the advertising exchange server at a specified time period;

extracting flow characteristics of each sampled flow from the sampling records;

Using a preset revenue calculation model to calculate the first value data corresponding to the traffic characteristics;

judging whether the mean value of multiple first value data is less than a first threshold;

If the mean value of the plurality of first value data is not less than the first threshold, stop filtering the traffic of the category.

Optionally, before the step of receiving each flow sent by the advertising exchange server, it also includes:

According to the target value data of each category of traffic and the characteristics of each traffic in the historical traffic records, as well as the corresponding actual value data and consumption revenue data, the revenue calculation model is trained.

Optionally, the step of training the revenue calculation model according to the target value data of each category of traffic and the characteristics of each traffic in the historical traffic records and the corresponding actual value data and consumption revenue data includes:

For each flow, the flow characteristic of the flow is used as an independent variable parameter, and the difference between the consumption income data of the flow and the actual value data, and the ratio of the target value data of the category to which the flow belongs are used as the dependent variable parameter;

Construct the training feature vector of the traffic with the respective variable parameters and dependent variable parameters;

A machine training model is used to train the training feature vector to obtain a revenue calculation model.

Optionally, the step of using a preset revenue calculation model to calculate the first value data corresponding to the traffic characteristics includes:

Substituting the flow characteristics of the flow into the income calculation model to calculate the first value data representing the ratio between the difference between the consumption income and the expected actual value and the target value of the flow.

According to another aspect of the present invention, a flow filtering device is provided, comprising:

The flow receiving module is used to receive each flow sent by the advertising exchange server;

A first traffic feature acquisition module, configured to extract traffic features from each traffic;

The first value data acquisition module is used to calculate the first value data corresponding to the traffic characteristics by using a preset revenue calculation model; the revenue calculation model is based on the target value data of each category of traffic and each traffic in the historical traffic record The characteristics and the corresponding actual value data and consumption revenue data are obtained through training; each flow in the historical flow record is the flow that received the successful bidding notification from the advertising exchange server;

A first judging module, configured to judge whether the first value data is smaller than a first threshold;

A filtering module, configured to filter the traffic when the first value data is less than a first threshold.

Optionally, the filter module includes:

A traffic category judging submodule, configured to judge the category to which the traffic belongs;

The filtering sub-module is configured to filter the traffic of the category according to a specified ratio.

Optionally, after the filtering module, it also includes:

A sampling traffic acquisition module, configured to use the remaining traffic after filtering the traffic of the category according to a specified ratio as the sampling traffic;

The sampling record acquisition module is used to obtain the sampling records of each sampling flow received from the advertising exchange server for bidding success notification every specified time period;

A second traffic feature acquisition module, configured to extract the traffic feature of each sampled flow from the sampling record;

Sampling the first value data acquisition module, configured to use a preset revenue calculation model to calculate the first value data corresponding to the traffic characteristics;

The second judging module is used to judge whether the mean value of multiple first value data is smaller than the first threshold;

A filtering stop module, configured to stop filtering the traffic of the category if the average value of the plurality of first value data is not less than the first threshold.

Optionally, before the flow receiving module, it also includes:

The revenue calculation model training module is used to train the revenue calculation model according to the target value data of each category of traffic, the characteristics of each traffic in the historical traffic records, and the corresponding actual value data and consumption revenue data.

Optionally, the income calculation model training module includes:

The parameter confirmation sub-module is used for each flow, taking the flow characteristic of the flow as an independent variable parameter, taking the difference between the consumption revenue data and the actual value data of the flow, and the target value data of the category to which the flow belongs The ratio of is used as the dependent variable parameter;

The training feature vector construction sub-module is used to construct the training feature vector of the traffic with the respective variable parameters and dependent variable parameters;

The income calculation model training sub-module is used to use the machine training model to train the training feature vector to obtain the income calculation model.

Optionally, the first value data acquisition module includes:

The first value data calculation sub-module is used to substitute the flow characteristics of the flow into the income calculation model, and calculate the first value representing the ratio between the difference between the consumption income and the expected actual value and the target value of the flow. value data.

According to a flow filtering method and device of the present invention, each flow sent by the advertising exchange server can be received; flow characteristics are extracted from each flow; and the first value data corresponding to the flow characteristics is calculated by using a preset revenue calculation model ; The revenue calculation model is obtained according to the target value data of each category of traffic and the characteristics of each traffic in the historical traffic records and the corresponding actual value data and consumption revenue data training; each traffic in the historical traffic records is received by the advertising transaction server Traffic notified of successful bidding; judging whether the first value data is less than a first threshold; when the first value data is less than the first threshold, filtering the traffic. This solves the problem that the existing advertising transaction server sends all the traffic to each DSP, resulting in waste of DSP system resources and reducing the cost performance of the DSP. The beneficial effects of enabling the advertising transaction server to filter the traffic before sending it to each DSP are achieved, thereby reducing waste of DSP system resources and improving cost performance.

The above description is only an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention, it can be implemented according to the contents of the description, and in order to make the above and other purposes, features and advantages of the present invention more obvious and understandable , the specific embodiments of the present invention are enumerated below.

Description of drawings

Various other advantages and benefits will become apparent to those of ordinary skill in the art upon reading the following detailed description of the preferred embodiment. The drawings are only for the purpose of illustrating a preferred embodiment and are not to be considered as limiting the invention. Also throughout the drawings, the same reference numerals are used to designate the same parts. In the attached picture:

FIG. 1 shows a flowchart of steps of a traffic filtering method according to an embodiment of the present invention;

FIG. 1A shows a schematic diagram of an RTB process according to an embodiment of the present invention;

FIG. 1B shows a flowchart of traffic filtering according to an embodiment of the present invention;

Fig. 1C shows a schematic diagram of a similar advertising space on the same page according to an embodiment of the present invention;

FIG. 1D shows a schematic diagram of the cumulative distribution of media monitoring volumes with different click monitoring rates in a website according to an embodiment of the present invention;

FIG. 1E shows a schematic diagram of a function graph for calculating the traffic filtering rate according to the click-to-listen rate according to an embodiment of the present invention;

Fig. 1F shows a schematic diagram of the gross profit brought by every 10,000 listens under different click-to-listen rates according to an embodiment of the present invention;

FIG. 1G shows a schematic diagram of the cumulative distribution of media net profit in the overall net profit under different click-to-listen rates according to an embodiment of the present invention;

FIG. 2 shows a flowchart of steps of a traffic filtering method according to an embodiment of the present invention;

Fig. 3 shows a schematic structural diagram of a flow filtering device according to an embodiment of the present invention; and

Fig. 4 shows a schematic structural diagram of a flow filtering device according to an embodiment of the present invention.

Detailed ways

Exemplary embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. Although exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be embodied in various forms and should not be limited by the embodiments set forth herein. Rather, these embodiments are provided for more thorough understanding of the present disclosure and to fully convey the scope of the present disclosure to those skilled in the art.

Embodiment one

A traffic filtering method provided by an embodiment of the present invention is introduced in detail.

Referring to FIG. 1 , it shows a flowchart of steps of a traffic filtering method in an embodiment of the present invention.

Step 101, receiving each flow sent by the advertising exchange server.

RTB is a bidding technology that uses third-party technology to evaluate and bid on each user's display behavior on millions of websites. FIG. 1A is a schematic diagram of an RTB process. The RTB process involves an advertising exchange (AdExchange, ADX) platform, a DSP, and a supply-side platform (Sell-Side Platform, SSP). Its core is the DSP platform. The DSP platform is like the headquarters of the display advertisement bidding: the first step is to send a request with user display information through its data tracking capability; the second step is to calculate the user display based on the estimation algorithm. value, and return the bid to the bid response engine; in the third step, the order is issued to the advertising trading system, and the advertising code is loaded to the major media, and finally delivered to the most accurate users. The advertising exchange server is an Internet advertising exchange platform. Like a stock exchange platform, the advertisement exchange server contacts the buyer and seller of the advertisement exchange, that is, the advertiser and the owner of the advertising space. A demand-side platform allows advertisers, such as advertisers and agencies, to more easily access and more efficiently purchase ad inventory because the platform aggregates inventory from various ad exchanges. With this platform, there is no need for another cumbersome purchase step - a purchase request. Supply-side platforms enable ad slot owners such as publishers to also intervene in the ad transaction, making their inventory available. Through this platform, publishers hope to get the highest eCPM for their inventory without having to sell it at a low price.

Among them, the DSP platform allows advertisers and advertising agencies to access more easily and purchase advertising inventory more efficiently, because the platform brings together the inventory of various advertising exchanges. With this platform, there is no need for another cumbersome purchase step - a purchase request. SSPs enable publishers to also intervene in the ad exchange, making their inventory available. Through this platform, publishers hope to get the highest eCPM for their inventory without having to sell it at a low price. An open online advertising marketplace (similar to a stock exchange) that connects publishers and advertisers. The inventory of advertising space in the trading platform is not necessarily all premium inventory, as long as the publisher wants to provide it, it can be found there.

The following example illustrates the specific process of RTB:

Step 1, the user uses a browser to access a media website, such as website A. Website A presents the content of this website, such as news, to users. At the same time, media websites will also display advertisements to users on some advertising slots. Although the main job of media websites is to present high-quality content to users, they also need to make money through advertisements so that the business of providing content can continue. They can use their own sales force to sell advertising inventory. However, for media sites, RTB is an increasingly important channel for selling advertising inventory. Here we assume that website A decides to sell the advertising space in the upper right corner through RTB.

Step 2, website A transmits the information to the advertising trading platform. The transmitted information includes URL (UniformResource Locator, Uniform Resource Locator), advertisement location, user Cookie ID (Cookie Identity, Cookie identity code), etc.

In step 3, the advertising exchange platform organizes a bidding and sends traffic bidding responses to multiple DSPs. Assume advertiser 1 is one of the DSPs.

Step 4: When the DSP server corresponding to advertiser 1 receives the traffic bidding response from the advertising exchange platform, it transmits the data to the bidding engine.

Step 5: The bidding engine sends the user ID to the user and advertiser information database to see if the needs of the user and the advertiser match.

Step 6, the bidding engine receives the information of the user and the advertiser, and decides whether to participate in the bidding and the bidding price.

Step 7, the bidding engine generates a bidding response and transmits it to the DSP server. The bid response includes the bid price and other information such as which ad server to get the ad creative from.

Step 8: The DSP server corresponding to the advertiser 1 sends a bidding response to the advertising exchange platform.

In step 9, the advertising exchange platform conducts bidding after receiving responses from all DSP servers or when the deadline is reached. For example, the deadline of the advertising exchange platform is 100ms, which means that the time difference between the advertising exchange platform sending the traffic bidding response and receiving the DSP bidding response must be within the deadline. Assume that advertiser 1 is the winner of this advertising auction transaction.

In step 10, the advertising exchange platform notifies the user browser of the winner of the auction. The browser sends an advertisement exposure request to the advertisement server where the advertisement creative is located.

Step 11, the advertisement server sends the advertisement creative to the user's browser.

Step 12, the user sees an advertisement on the website. Then if the user is interested in the advertisement, he will click on the advertisement to enter the advertiser's landing page. Users browse the advertiser's website and may take further actions. For example, if the advertiser is an e-commerce company, the user may make a purchase on the e-commerce website.

During the real-time bidding process, the advertising exchange server can send traffic to each advertiser in the demand-side platform, and the advertiser can conduct real-time bidding on the received traffic. For different types of traffic, the advertiser can obtain different revenues, and there may be some traffic, and the advertiser may not be able to obtain revenue, so for this type of traffic, the advertiser does not need to participate in the bidding. Therefore, in the embodiment of the present invention, it is possible to filter each traffic sent by the advertising exchange server, so as to avoid invalid traffic competition or unnecessary traffic competition.

Then firstly, it is necessary to receive the traffic sent by the advertising transaction server before bidding. In the embodiment of the present invention, if the advertisement transaction server sends traffic, the traffic sent by the advertisement transaction server is received; and if the advertisement transaction server sends a traffic request for each traffic, the traffic request sent by the advertisement transaction server is received. This embodiment of the present invention is not limited. The traffic request may include the aforementioned traffic bidding response, which is not limited in this embodiment of the present invention.

Figure 1B is a flowchart of traffic starting from ADX and finally returning to ADX after traffic filtering is added. The biggest difference from the existing situation is that before the traffic is sent to the back-end module, it will be decided in advance whether it is necessary to bid on this traffic. If there is no bidding, an empty Response will be returned directly, otherwise it will be the same as the previous bidding process Same.

Step 102, extract traffic features from each traffic.

We preliminarily determined that the quality of a traffic depends on whether it can bring clicks, so we can define an indicator, Click Listen Rate (CLR), and the calculation of CLR is as follows:

Click monitoring rate = number of clicks/number of monitoring = number of clicks/number of impressions*number of impressions/number of bids*number of bids/number of monitoring=click rate*success rate*bidding rate

It can be seen that although in the whole process, there are many factors that affect the click monitoring rate, some are related to bidding, and some are related to post-effects. But in the final analysis, the more clicks a traffic can bring us, the better its quality is. This is the most intuitive concept, and there is no need to pay special attention to the factors affecting the intermediate links.

Originally, only the click rate is enough, but considering the following two scenarios, we also need to consider the revenue.

If a lot of traffic on an ad slot is based on CPM, its click effect has little impact on the final revenue. Therefore, if the bidding effect is good, even if the click-through rate is low, it is still profitable. In this case should not be filtered out.

If the bidding effect on the traffic of an ad slot is very good, and it just meets the needs of an advertiser with a high CPC bid, even if its click-through rate is a little low, it should not be filtered out.

The calculation of traffic revenue is divided into the following three parts:

(1) Internal cost

Every time a DSP monitors a flow, no matter whether it bids or not, there is a cost. This cost includes network, machine resources, manpower, etc. It has little to do with external factors. It is generally fixed and stable and can be regarded as a constant (use C Express), at present, the value given by the operation and maintenance department is 2 cents per thousand monitoring.

(2) External costs

DSP needs to pay for each successful bidding traffic. This fee is floating and related to the low price of the media and the bids of other DSPs, so it is called external cost. Only when the bidding is successful, there is an external transaction price (P), therefore, the external cost of a flow can be expressed as: BR*WR*P

(3) Internal income

Assuming that all advertisers in the DSP bid according to CPC (Cost Per Click), the DSP will deduct fees from the advertiser only after the successful bidding idea generates a click. Therefore, the internal revenue of monitoring traffic bid by CPC and CPM (CostPerMille, cost per thousand people) is expressed as BR (Bidding rate, bid rate) * WR (success rate) * CTR (Click-through Rate, click-through rate) *CPC and BR*WR*CPM/1000. Assuming that FPS (Fee Per Show, charged per impression) is used to represent the average deducted fee per impression, the values corresponding to CPC and CPM bids are CTR*CPC and CPM/1000.

To sum up, the monitoring revenue of a flow can be expressed as: RPL=BR*WR*FRS-C-BR*WR*P.

The cases described above are all based on the second-level domain name. In fact, under the same page, different advertising positions may have different effects. Generally speaking, the click-through rate of the first screen will be higher than that of the non-first screen. High, the click-through rate of large-sized ad slots will be higher than that of small-sized ones. This means that all ad slots under the same domain cannot be treated uniformly. The best situation is that ADX provides a unique number for different advertising positions, and DSP can directly use this unique number to count its corresponding bidding effect (bidding rate, bidding success rate) and post-effect (click rate, revenue situation) ). But in fact, not all ADXs provide unique identifiers for advertising slots. Therefore, DSPs are required to combine and map to an advertising slot based on different dimensional information in traffic requests.

The selection of specific dimensions needs to consider multiple indicators: there needs to be sufficient discrimination, the conditions between each dimension are independent (at least approximate), sufficiently general, applicable to all ADX, and so on.

Combined with the data verification of DSP Cube, finally, from the perspective of the media, the following dimensions were selected: second-level domain name, number of screens, and size of advertising space.

Coincidentally, these three dimensions are originally the input parameters for DSP to calculate the third-party advertising slot ID. In addition, due to the same set of dimension values, the effect may be different in different ADXs. The following table shows the data of Tanx and Baidu under the www.zhibo8.cc domain name. It can be seen that there is still a relatively large difference in the success rate and click-through rate. Although the domain name level is too wide, there will be some errors, but the comprehensive data is sufficient. D.

Table 1. Comparison of the effects of Tanx and Baidu in zhibo8

exchange bidding rate Success rate click rate CM rate Tanx 0.91 0.257 0.00093 0.54 Baidu 0.97 0.16 0.00326 0.67

Obviously, these dimensions cannot be mapped one-to-one, that is, multiple advertising slots may be mapped to the same combined dimension value, which requires that the advertising slots mapped to the same dimension value have similar bidding effects and after effects.

The picture below shows three advertisements on the same page. Through the mapping of these three dimensions, they will be treated as the same advertisement slot. Although the user has a higher probability of clicking on which ad slot, different people have different habits, and even involves behavioral analysis related to physiology and psychology, but at least the data in Table 2 verify that the post-effects of these three ad slots The difference is not large, so they can be mapped to the same combination of dimension values. As shown in Figure 1C, there are three advertisements on the same page. Through the mapping of these three dimensions, they will be treated as the same advertisement space. Although the user has a higher probability of clicking on which ad slot, different people have different habits, and even involves behavioral analysis related to physiology and psychology, but at least the data in Table 2 verify that the post-effects of these three ad slots The difference is not large, so they can be mapped to the same combination of dimension values.

Table 2. Three ad slots mapped to the same dimension value on the same page

Bid Request (bidding request), that is, the traffic request in this application, in addition to media advertising space information, also includes user information, which may include third-party ID, gender, interest tags, device information, etc. This part of user data will have a great impact on the choice of creative delivery, which may also affect the subsequent bidding effect and post-effect. Currently, third-party ID data is too large to be suitable as a feature. The gender is again imprecise and inappropriate. The device information is relatively in line with our feature selection criteria. However, due to the fact that the analysis of devices is still relatively small, from a simple point of view, it will not be included in the feature combination in the first phase, but later, with the launch of Mobile, it is expected that there will be The dimensions of the device are added to the model.

In addition to the above information that can be directly extracted from the Bid Request, the most important user information is actually the effect of Cookie Mapping. It is certain that the traffic effect of successful Cookie Mapping is definitely better than that of Cookie Mapping failure. Although the inclusion of this part of information means that the traffic filtering module needs to be placed after the cookie mapping query, and the cookie mapping itself is an asynchronous operation, which will affect the efficiency, but this work is worth doing. DSP can tolerate every traffic Go through the process of Cookie Mapping, because even this is at least not worse than it is now.

Therefore, we only use whether the Cookie Mapping is successful as a feature of the user. After that, we will also consider the time of the cookie and the number of tags as features.

In the real-time bidding process of each advertiser, the value of the bidding amount has a great relationship with the traffic. For example, if the current traffic corresponds to users who often visit shopping websites, then if the advertiser is also a shopping advertisement, the bid amount may be relatively high, and the advertiser may obtain relatively high revenue; conversely, if the current The traffic corresponding to the user hardly visits the shopping website, so the advertiser will bid a relatively low amount, and if the bid is successful, the advertiser may obtain very low revenue.

Therefore, in the embodiment of the present invention, it is necessary to acquire the traffic characteristics of each traffic. That is, flow features are extracted from each flow. Among them, the traffic characteristics may include the time corresponding to the traffic, the region corresponding to the traffic, the media context corresponding to the traffic, the size of the advertisement space corresponding to the traffic, the user information corresponding to the traffic, etc., and the user information corresponding to the traffic may include user attributes, user interest dimensions, user browsing history, and so on. Wherein, the time corresponding to the traffic refers to the corresponding generation time of the traffic, that is, the time when the user browses the webpage where the advertisement slot corresponding to the traffic is located. The region corresponding to the traffic refers to the region where the user is located when the user browses the webpage where the advertisement space corresponding to the traffic is located, and may include the IP (Internet Protocol, Internet Protocol) address of the user browsing the webpage. The media context corresponding to the traffic may include the context content of the webpage where the advertisement slot corresponding to the traffic is located. User attributes may include user's gender, age and other personal information. The user's interest dimension refers to the user's points of interest, which may include hobbies, social circles, and so on. In the embodiment of the present invention, the specific content of the traffic characteristics to be extracted can be set according to requirements before this step, or in any step before this step, which is not limited in this embodiment of the present invention.

Step 103, using the preset revenue calculation model to calculate the first value data corresponding to the traffic characteristics; the revenue calculation model is based on the target value data of each category of traffic and the characteristics of each traffic in the historical traffic records and the corresponding actual Value data and consumption revenue data are obtained through training; each flow in the historical flow record is the flow that received the successful bidding notification from the advertising exchange server.

After the flow characteristics of each flow are acquired, the preset revenue calculation module may be used to calculate the first value data corresponding to the flow characteristics of each flow. The first value data is the revenue value recognized by the advertiser, rather than the conventional actual revenue.

For example, for a certain amount of traffic, an advertiser A’s target value data for this type of traffic is 10 yuan, but the advertiser A finally participated in the bidding with 1 yuan and the bid was successful, that is, the actual value data corresponding to this traffic is 1 yuan , at this time, if the user corresponding to the traffic brings the advertiser an income of 2 yuan, then the actual income is 1 yuan. But for this traffic, the revenue value recognized by the advertiser is 10 yuan. It can be seen that, in the embodiment of the present invention, the revenue value approved by the advertiser may include the difference between the target value data and the actual value data, and the actual revenue.

Moreover, in the embodiment of the present invention, for the convenience of comparison, the difference between the revenue approved by the advertiser and the target value data can be used as the revenue parameter for comparison. Then the income parameter at this time is the difference between the target value data and the actual value data and the sum of the actual income is the numerator, and the target value data is the ratio of the denominator, namely:

Earnings parameter = (target value data - actual value data + actual income) / target value data

After simplifying the above formula, it can be obtained that income parameter=1+(actual value data-actual income)/target value data. Among them, 1 is a fixed value, then (actual value data-actual income)/target value data is directly proportional to the final income. In the embodiment of the present invention, the ratio of the difference between the actual value data and the actual income to the target value data may be used as the first value data. Obviously, the larger the value of the first value data, the higher the income of the corresponding traffic. Of course, the specific content contained in the first value data can be set according to requirements before this step, or before any step before this step, which is not limited in this embodiment of the present invention.

Among them, the revenue calculation model is obtained according to the target value data of each category of traffic and the characteristics of each traffic in the historical traffic records, as well as the corresponding actual value data and consumption revenue data. Moreover, only traffic with successful bidding can bring revenue, so each traffic in the historical traffic records in the embodiment of the present invention is the traffic that has received a successful bidding notification from the advertising transaction server.

Then, the characteristics of each flow in the historical flow records can be used as input, and the ratio of the difference between the consumption income data and actual value data of the corresponding flow and the target value data can be used as the output to train the income calculation model.

Based on the principle that the simpler the better, in the embodiment of the present invention, you can use Bayes as a one-period traffic filtering model.

Model input: DSP Cube is directly used in the first phase, and all DSP original logs of a certain period of time (such as a week) may be used later.

Model output: the prior probability of traffic that brings clicks and the independent probability distribution on each feature dimension; the prior probability of traffic that does not bring clicks and the independent probability distribution on each feature dimension.

When the DSP receives a flow, to decide whether to filter it out, it first needs to extract it from the flow, that is, express the flow as (x1, x2, ..., xn), and xi represents the value of the i-th dimension.

Next, combined with the trained The Bayes model predicts the click-to-listen rate (CLR) according to the following formula:

Here, event 1 (positive example) indicates that a listener brought a click, and event 0 (negative example) indicates that a listener did not bring a click. P(1) represents the prior probability that one monitoring can bring about one click, P(0)=1-P(1). p(1|x1,x2,...,xn) represents the probability that a traffic (x1, x2,...,xn) will eventually bring a click. Assuming that x1 represents the dimension value of the second-level domain name, then p(x1|1) represents the proportion of all traffic that brings clicks under the domain name x1, and p(x1|0) represents the proportion of all traffic that does not bring clicks , belonging to the proportion under the x1 domain name.

It can be expected that the click-to-listen rates predicted by the model will be distributed on [0,1], and most of them are less than 0.1%. if fully used Bayes classifier, almost all traffic will be classified as negative examples, so we do not directly use Bayesian classifier. If what we expect is to calculate the probability of traffic filtering according to the CLR, we need to map it more compactly to the [0,1] interval.

Step 104, judging whether the first value data is smaller than a first threshold.

Since CLR itself is a continuous value, its data is mainly distributed in a small sub-interval of [0,1], as shown in Figure 1D, which shows the traffic on all second-level domain names of a website on January 11, 2014 The cumulative distribution of click listen rates for . The average click-to-display ratio is 1.3/10000. It can be seen that the traffic below 0.5/10000 accounts for about 50%, the traffic below 1/10000 accounts for nearly 60%, the traffic below 6/10000 accounts for more than 80%, and more than 90% are In the interval [0,1/1000].

Therefore, we only need to set a threshold a. When the CLR is higher than a, we keep all the values between [0, a], and directly use the linear function to make it more compactly distributed in [0,1] superior. FIG. 1E shows the function graph of calculating the traffic filtering rate according to the click monitoring rate. It can be seen that the smaller the ratio a, the greater the probability of being filtered.

The selection of a here will be directly related to the strictness of traffic filtering, the higher the value is, the stricter it is, and the lower it is, the looser it will be. At this time, it is necessary to use the traffic revenue as the basis for selecting a. Figure 1F shows the distribution of gross profit brought by every 10,000 monitoring times in a certain website under different monitoring ratios. The unit of the vertical axis is yuan. The line is the quantile line with a monitor ratio of 1/10000. It can be seen that in the traffic with a click monitoring rate of 1/10000, we basically lose money, and those with a rate higher than 1/10000 basically lose money. Therefore, this 1/10000 value can actually be used as an important reference for selecting the threshold a.

Figure 1G shows the cumulative distribution of media net profit in the overall net profit under different click monitoring rates. It can be seen that traffic with a click monitoring rate of second-level domain names below 1/10000 (accounting for about 60% of the total) is a loss Yes, this part of the lost money will not be filled until 6/10000. Then starting from 6/10000, it officially started to make money. According to the data of a certain website for this day, we can set the threshold a to 1/10000.

If the critical point is purely based on whether to make money, sometimes it is not necessarily accurate. For example, if the premium is too high, the overall profit of DSP may become negative. At this time, the critical point of making money will be very high, even higher than the average click-to-monitor rate. Therefore, an online parameter μ=[0,1] needs to be provided, so that we can decide in real time how much to filter traffic.

As mentioned above, in the embodiment of the present invention, the first value data is directly proportional to the recognition income, then the greater the first value data, the higher the advertiser’s income, and if the first value data is lower than the first threshold, then It indicates that the advertiser's income is low or has no income at this time, so the advertiser does not need to bid for the corresponding traffic at this time. The first threshold may be set before this step or any step before this step according to requirements, which is not limited in this embodiment of the present invention. Moreover, for different advertisers, the value of the first threshold may be different or the same, which is not limited in this embodiment of the present invention.

Step 105, when the first value data is less than a first threshold, filter the traffic.

Then, if the first value data is less than the first threshold, it indicates that the advertiser does not need to participate in bidding for the corresponding traffic, and at this time, the traffic can be filtered out and not sent to the corresponding advertiser; and if the first value data is not used If it is smaller than the first threshold, the traffic will not be filtered out. However, if for the same traffic, the first value data corresponding to another advertiser is not less than the advertiser's first threshold, then the traffic will not be filtered out for the advertiser.

After the introduction of traffic filtering, the life process of a traffic after it reaches the DSP is mainly divided into three scenarios:

First, the DSP receives ADX traffic and passes it to the backend, and finally initiates the bidding.

Second, the DSP receives ADX traffic and passes it to the backend, but does not participate in the bidding in the end.

Third, the traffic is directly filtered by the DSP and is not sent to the backend.

The current monitoring traffic statistics include all three situations. When calculating the click-to-listen rate, there will be a problem. If the third scenario is more, the click-to-listen rate (CLR) will be smaller. According to the above-mentioned , CLR itself is the factor that decides whether a traffic should be filtered or not, so it will iteratively make the third scenario more, and the final result is that a certain traffic will become worse and worse after the occasional effect is poor on a certain day. Even if other factors improve, there is still a high probability that it will be regarded as low-quality traffic. This seriously affects the robustness and generalizability of the model.

In order to solve this problem, it is necessary to define a new concept, valid listen (Valid Listen). For DSP, only when a flow passes through the back-end computing module, it can be regarded as a valid listen, otherwise it can only be regarded as a false listen. monitor. Therefore, when calculating the CLR, only Scenario 1 and Scenario 2 should be considered. This way, the CLR will not be affected by the results of traffic filtering. In order to count effective monitoring, an additional field needs to be added to the DSP log to mark whether a traffic is valid monitoring, and if not, why it is filtered. At the same time, such a dimension needs to be added to the cube of the dsp, so that the input of the model can be obtained directly from the cube.

The existing DSP Cube does not have the concept of effective monitoring. It is necessary to add a dimension to the Cube based on the log information to mark whether the monitored traffic is valid.

In the embodiment of the present invention, it is possible to receive each traffic sent by the advertising exchange server; extract the traffic characteristics from each traffic; use the preset revenue calculation model to calculate the first value data corresponding to the traffic characteristics; the revenue calculation The model is trained based on the target value data of each category of traffic and the characteristics of each traffic in the historical traffic records, as well as the corresponding actual value data and consumption revenue data; the traffic in the historical traffic records is the traffic that received the successful bidding notification from the advertising exchange server ; Judging whether the first value data is less than a first threshold; when the first value data is less than the first threshold, filter the traffic. In this way, the advertising trading server can filter the traffic before sending it to each DSP, thereby reducing the waste of DSP system resources and improving the cost performance.

Embodiment two

A traffic filtering method provided by an embodiment of the present invention is introduced in detail.

Referring to FIG. 2 , it shows a flowchart of steps of a traffic filtering method in an embodiment of the present invention.

Step 201, train the revenue calculation model according to the target value data of each category of traffic, the characteristics of each traffic in the historical traffic records, and the corresponding actual value data and consumption revenue data.

Since the revenue calculation model needs to be used to calculate the first value data of traffic in the subsequent steps, the revenue calculation model needs to be trained first. In the embodiment of the present invention, specifically, the revenue calculation model can be trained according to the target value data of each category of traffic, the characteristics of each traffic in the historical traffic records, and the corresponding actual value data and consumption revenue data. Moreover, in order to improve the accuracy of the revenue calculation model obtained through training, the target value data of different types of traffic, the characteristics of each traffic in the historical traffic records, and the corresponding actual value data and consumption revenue data are respectively selected to train the revenue calculation model. The specific traffic used to train the revenue calculation model can be selected according to requirements, which is not limited in this embodiment of the present invention.

It should be noted that the specific dimensions of the characteristics of each traffic used in training the revenue calculation model need to be consistent with the traffic feature dimensions extracted from the traffic received from the advertising exchange server, or the subsequent The traffic feature dimension extracted from the traffic sent by the advertising exchange server includes the characteristics of each traffic used when training the revenue calculation model.

Optionally, in the embodiment of the present invention, the step 201 may further include:

Sub-step 2011, for each flow, take the flow characteristic of the flow as the independent variable parameter, and take the difference between the consumption revenue data and the actual value data of the flow and the target value data of the category to which the flow belongs as the ratio Dependent variable parameters.

After the traffic used to train the revenue calculation model is selected, the traffic characteristics of each traffic can be set as independent variable parameters for each traffic, that is, the input of the revenue calculation model, and the consumption revenue data and actual value data of each traffic can be used The ratio of the difference of the corresponding flow to the target value data of the category to which the corresponding flow belongs is used as the dependent variable parameter, that is, the output of the revenue calculation model.

In sub-step 2012, the respective variable parameters and dependent variable parameters are used to construct the training feature vector of the flow.

In order to conveniently and quickly complete the training process of the revenue calculation model, in the embodiment of the present invention, the independent variable parameters and dependent variable parameters corresponding to each flow may also be used to construct a training feature vector of the corresponding flow.

Sub-step 2013, using the machine training model to train the training feature vector to obtain a revenue calculation model.

After the training feature vector is obtained, the selected machine training model can be trained by using the training feature vector to obtain the revenue calculation model. The machine training model wherein can be any available machine learning model such as neural network training model, fuzzy mathematics training model, can be before this step in the embodiment of the present invention, or before any step before this step according to requirement The specific type of the machine training model, such as specific parameters in the machine training model, etc., are not limited in this embodiment of the present invention.

Step 202, receiving each flow sent by the advertising exchange server.

Step 203, extract traffic features from each traffic.

Step 204: Substituting the flow characteristics of the flow into the income calculation model to calculate the first value data representing the ratio between the difference between the consumption income and the expected actual value and the target value of the flow.

Step 205, judging whether the first value data is smaller than a first threshold.

Step 206, when the first value data is less than a first threshold, determine the category to which the traffic belongs.

As mentioned above, in the implementation of the present invention, it may be determined whether to filter the traffic for each traffic. For example, if it is confirmed that the first value data of a certain flow is smaller than the first threshold, the flow may be filtered out. Then, the number of judgments at this time is more, so in the embodiment of the present invention, each traffic can be classified for the first time, and after confirming that the first value data of a certain traffic is less than the first threshold value, the first value data of other traffic of the same category can be deduced. A value of data may be less than the first threshold.

However, in actual situations, there may be two flows of the same category, the first value data of one flow is less than the first threshold, but the first value data of the other flow is not less than the first threshold. Then, for traffic whose first value data is less than the first threshold, if all traffic belonging to the same category as this traffic is filtered out, users may not be able to obtain traffic of this category, and the first value data calculated by the revenue calculation model Not all of them are necessarily accurate, so if all of them are filtered out, the misjudged traffic will affect the revenue of the corresponding advertiser. Therefore, in order to avoid the above-mentioned adverse effects, in the embodiment of the present invention, it can be set that the corresponding advertiser can still receive Part of the traffic whose first value data is less than the first threshold is detected, and the specific filtering ratio may be related to the category to which the traffic belongs.

In the embodiment of the present invention, the category to which the traffic belongs can be divided according to all or part of the traffic characteristics of each traffic, and the correspondence between the specific traffic characteristics and the category to which the traffic belongs can be before this step, or any prior to this step. The first step is set according to requirements, which is not limited in this embodiment of the present invention. For example, the category of the traffic can be divided according to the region corresponding to each traffic in the traffic characteristics, or the advertising space size corresponding to each traffic in the traffic characteristics, the media context corresponding to the traffic, the user information corresponding to the traffic, and the The combination of the time and the region corresponding to the traffic divides the category to which the traffic belongs, which is not limited in this embodiment of the present invention.

Moreover, in the embodiment of the present invention, it is also possible not to classify the categories of traffic according to traffic characteristics, but to classify the categories of traffic according to preset parameters, wherein the preset parameters can be before this step according to requirements, or this step The setting is performed before any step before the step, which is not limited in this embodiment of the present invention. For example, the traffic category can be divided according to the first value data of each traffic, then the first value data threshold corresponding to each category can be set, and then the category of each traffic can be determined.

Step 207, filtering the traffic of the category according to a specified ratio.

After the category of the traffic is determined, the traffic can be filtered according to the specified ratio corresponding to the category. Wherein, the designated ratio corresponding to each category may be set according to requirements before this step, or before any step before this step, which is not limited in this embodiment of the present invention.

For example, if the first value data of the current traffic is less than the first threshold, and the specified ratio corresponding to the category of the current traffic is 0.7, then 7 out of every 10 traffics of this category may be filtered out.

In addition, in the embodiment of the present invention, the sequence logic of specific filtering traffic may be preset according to requirements. For example, to filter out 7 out of every 10 traffic of the same category mentioned above, you can set to filter out the first 7 out of 10 traffic of the same category, or filter out the last 7, or other rules, etc. Wait.

Step 208: Use the remaining traffic after filtering the traffic of the category according to the specified ratio as the sampling traffic.

In the embodiment of the present invention, in order to further perform sampling correction on the revenue calculation model, the remaining traffic after filtering according to a specified ratio may be used as the sampling traffic. For example, after filtering out 7 of the above-mentioned 10 traffic, the other three traffic will be sent to the corresponding advertiser, then the remaining traffic after filtering is the unfiltered 3 traffic. Certainly, in the embodiment of the present invention, if the 7 filtered flows can be retrieved again, then the 7 filtered flows can also be used as sampling flows at this time, which is not limited in this embodiment of the present invention.

Step 209 , acquiring the sampling records of each sampling flow that receives the bidding success notification from the advertising exchange server every specified time period.

We know that no matter how good the model is, there may be misjudgments. A good traffic may be filtered out as bad, and a bad traffic may be diagnosed as good traffic. For example, a traffic itself is actually good, but its effect is not good (low display ratio or click-through rate), maybe not because of poor traffic, but because of some other factors that we have not yet counted. This requires our model to provide an error correction mechanism.

Considering that this phenomenon may only appear after going online, from the perspective of simplicity, effectiveness and timeliness, in addition to the traffic filtering model, there will also be independent blacklist and whitelist functions. When good traffic is filtered out due to misjudgment by the model, it needs to be added to the whitelist; on the contrary, when bad traffic is not filtered out due to misjudgment, it needs to be added to the blacklist. Of course, the smaller the blacklist and blacklist, the better. Every entry or exit requires strong data support. And it can only be used as a temporary solution, and it still needs to be solved by the optimization of the model.

The future effect evaluation of DSP will be more and more inclined to two data: the number of clicks and the amount of effective monitoring. Therefore, the most direct basis for judging the quality of traffic filtering is that the number of clicks drops very little when the effective monitoring traffic decreases. At the same time, platform margins will improve due to fewer bids on loss-making media.

If the traffic of a media is not good in the later stage, after being filtered by the traffic, the solution with the specified media delivery will be affected, and the advertiser may find that the various data on this media are reduced. But this situation should be, we cannot cause ourselves to lose money because of the needs of advertisers. Moreover, generally designated media will designate multiple media at the same time, and less is just an amount that is not effective. In addition, the traffic of this poor media is not completely filtered out, and a part of it will be left as exploration (for example, 5%). If the advertiser particularly wants the traffic of this media, as long as the bid is sufficient, then when the creative selection is made Naturally, they can all win, and grabbing this amount is enough for them to consume.

After confirming the sampling traffic, you can perform sampling correction on all the sampling traffic, but in order to reduce the repeated switching between sampling correction and normal filtering process, you can set a specified time period. Sampling record of each sampled traffic notified by the server bid successfully.

Step 210, extracting the flow characteristics of each sampled flow from the sampling records.

In the embodiment of the present invention, in order to recalculate the first value data of each sampled flow, it is correspondingly necessary to extract the flow feature of each sampled flow again. The specific collection process and traffic characteristics are similar to the aforementioned step 102, and will not be repeated here.

Step 211, using a preset revenue calculation model to calculate the first value data corresponding to the traffic characteristics.

This step is similar to the aforementioned step 103 and will be repeated here.

Step 212, judging whether the mean value of multiple first value data is smaller than a first threshold.

At this time, in order to sample and correct the filtered traffic, if the judgment is still made on the first value data of each individual traffic, the result may still not be accurate enough. Therefore, in the embodiment of the present invention, in order to improve the accuracy of sampling deviation correction, it may be determined whether the average value of the first value data of multiple flows belonging to the same category is smaller than the first threshold. The number of specific first value data can be set according to requirements, which is not limited in this embodiment of the present invention. Alternatively, whether the average value of the first value data of each sampled flow belonging to the same category in the current sampled record is smaller than the first threshold is not limited by this embodiment of the present invention.

Step 213, if the average value of the plurality of first value data is not less than the first threshold, stop filtering the traffic of the category.

As mentioned above, if the first value data of a certain type of traffic is less than the first threshold, it means that the traffic brings relatively little revenue to the corresponding advertiser at this time, and the possibility of the advertiser participating in the traffic bidding is very low. Then in the process of deviation correction, if the average value of the first value data of the above-mentioned multiple flows of the same category is not less than the first threshold, then the filtering of the flow of this category can be stopped correspondingly, that is, if the above-mentioned multiple first value data of the same category If the average value of the value data is not less than the first threshold, then the filtering of traffic of this category may be stopped correspondingly.

For example, assuming that for traffic A belonging to category 1, the first value data of traffic A calculated by using the revenue calculation model is 4, and for the current advertiser, the set first threshold value is 5, then for the current advertiser , traffic A will be filtered out, and all traffic contained in category 1 will be filtered according to the specified ratio. Assuming that the currently acquired sampling records include traffic B and traffic C belonging to category 1, and the first value data of traffic B is calculated again using the revenue calculation model, and the first value data of traffic C is 6, then the corresponding If the average value of the three first value data of the category is 5, which is not less than the first threshold, then at this time, it is possible to stop filtering the traffic of this category and send it to the current advertiser.

In the embodiment of the present invention, it is possible to receive each traffic sent by the advertising exchange server; extract the traffic characteristics from each traffic; use the preset revenue calculation model to calculate the first value data corresponding to the traffic characteristics; the revenue calculation The model is trained based on the target value data of each category of traffic and the characteristics of each traffic in the historical traffic records, as well as the corresponding actual value data and consumption revenue data; the traffic in the historical traffic records is the traffic that received the successful bidding notification from the advertising exchange server ; Judging whether the first value data is less than a first threshold; when the first value data is less than the first threshold, filter the traffic. In this way, the advertising transaction server can filter the traffic before sending it to each DSP, thereby reducing the waste of DSP system resources and improving the cost performance.

Moreover, in the embodiment of the present invention, when filtering the traffic, the traffic of the corresponding category may also be filtered according to a specified ratio according to the category of the traffic. Thus, the waste of DSP system resources can be further reduced.

In addition, in the embodiment of the present invention, after the step of filtering the traffic of the corresponding category according to the specified ratio, the remaining traffic after filtering the traffic of the category according to the specified ratio can also be used as the sampling traffic; every specified time Periodically, obtain the sampling record of each sampled traffic that has received the bidding success notification from the advertising exchange server; extract the traffic characteristics of each sampled traffic from the sampling records; use the preset revenue calculation model to calculate the first value data; judging whether the average value of the multiple first value data is less than the first threshold; if the average value of the multiple first value data is not less than the first threshold, stop filtering the traffic of the category. Therefore, it is possible to sample and correct the filtered traffic, avoid false filtering of traffic, and ensure that each DSP can receive the traffic required by the advertiser.

It should be noted that, for the steps similar to those in the first embodiment above, reference may be made to the first embodiment, which will not be repeated here.

For the method embodiment, for the sake of simple description, it is expressed as a series of action combinations, but those skilled in the art should know that the embodiment of the present invention is not limited by the described action order, because according to the embodiment of the present invention , certain steps may be performed in other order or simultaneously. Secondly, those skilled in the art should also know that the embodiments described in the specification belong to preferred embodiments, and the actions involved are not necessarily required by the embodiments of the present invention.

Embodiment three

A traffic filtering device provided by an embodiment of the present invention is introduced in detail.

Referring to FIG. 3 , it shows a schematic structural diagram of a flow filtering device in an embodiment of the present invention.

The flow receiving module 301 is configured to receive each flow sent by the advertising transaction server.

The first traffic feature acquisition module 302 is configured to extract traffic features from each traffic.

The first value data acquisition module 303 is configured to use a preset revenue calculation model to calculate the first value data corresponding to the traffic characteristics; The characteristics of the traffic and the corresponding actual value data and consumption revenue data are obtained through training; each traffic in the historical traffic record is the traffic that received the successful bidding notification from the advertising exchange server.

The first judging module 304 is configured to judge whether the first value data is smaller than a first threshold.

A filtering module 305, configured to filter the traffic when the first value data is less than a first threshold.

In the embodiment of the present invention, it is possible to receive each traffic sent by the advertising exchange server; extract the traffic characteristics from each traffic; use the preset revenue calculation model to calculate the first value data corresponding to the traffic characteristics; the revenue calculation The model is trained based on the target value data of each category of traffic and the characteristics of each traffic in the historical traffic records, as well as the corresponding actual value data and consumption revenue data; the traffic in the historical traffic records is the traffic that received the successful bidding notification from the advertising exchange server ; Judging whether the first value data is less than a first threshold; when the first value data is less than the first threshold, filter the traffic. In this way, the advertising transaction server can filter the traffic before sending it to each DSP, thereby reducing the waste of DSP system resources and improving the cost performance.

Embodiment four

A traffic filtering device provided by an embodiment of the present invention is introduced in detail.

Referring to FIG. 4 , it shows a schematic structural diagram of a flow filtering device in an embodiment of the present invention.

The revenue calculation model training module 401 is used to train the revenue calculation model according to the target value data of each category of traffic, the characteristics of each traffic in the historical traffic records, and the corresponding actual value data and consumption revenue data.

Optionally, in the embodiment of the present invention, the income calculation model training module 401 includes:

The parameter confirmation sub-module 4011 is used to, for each flow, take the flow characteristic of the flow as an independent variable parameter, take the difference between the consumption revenue data and the actual value data of the flow, and the target value of the category to which the flow belongs The ratio of the data is used as the dependent variable parameter.

The training feature vector construction submodule 4012 is used to construct the training feature vector of the traffic by using the respective variable parameters and dependent variable parameters.

The revenue calculation model training sub-module 4013 is configured to use the machine training model to train the training feature vector to obtain a revenue calculation model.

The flow receiving module 402 is configured to receive each flow sent by the advertising exchange server.

The first traffic feature acquisition module 403 is configured to extract traffic features from each traffic.

The first value data acquisition module 404 is configured to use a preset revenue calculation model to calculate the first value data corresponding to the traffic characteristics; The characteristics of the traffic and the corresponding actual value data and consumption revenue data are obtained through training; each traffic in the historical traffic record is the traffic that received the successful bidding notification from the advertising exchange server.

Optionally, in this embodiment of the present invention, the first value data acquisition module 404 includes:

The first value data calculation sub-module 4041 is used to substitute the flow characteristics of the flow into the income calculation model, and calculate the first value representing the ratio between the difference between the consumption income and the expected actual value and the target value of the flow. A value data.

The first judging module 405 is configured to judge whether the first value data is smaller than a first threshold.

A filtering module 406, configured to filter the traffic when the first value data is less than a first threshold.

Optionally, in the embodiment of the present invention, the filtering module 406 includes:

The traffic category judging sub-module 4061 is configured to judge the category to which the traffic belongs.

The filtering submodule 4062 is configured to filter the traffic of the category according to a specified ratio.

The sampling traffic acquisition module 407 is configured to use the remaining traffic after filtering the traffic of the category according to a specified ratio as the sampling traffic.

The sampling record obtaining module 408 is configured to obtain the sampling records of each sampled flow that receives the bidding success notification from the advertising transaction server every specified time period.

The second flow feature acquisition module 409 is configured to extract the flow feature of each sampled flow from the sampling record.

The sampling first value data acquisition module 410 is configured to calculate the first value data corresponding to the traffic characteristics by using a preset revenue calculation model.

The second judging module 411 is configured to judge whether the average value of the plurality of first value data is smaller than the first threshold.

The filtering stop module 412 is configured to stop filtering the traffic of the category if the average value of the plurality of first value data is not less than the first threshold.

In the embodiment of the present invention, it is possible to receive each traffic sent by the advertising exchange server; extract the traffic characteristics from each traffic; use the preset revenue calculation model to calculate the first value data corresponding to the traffic characteristics; the revenue calculation The model is trained based on the target value data of each category of traffic and the characteristics of each traffic in the historical traffic records, as well as the corresponding actual value data and consumption revenue data; the traffic in the historical traffic records is the traffic that received the successful bidding notification from the advertising exchange server ; Judging whether the first value data is less than a first threshold; when the first value data is less than the first threshold, filter the traffic. In this way, the advertising exchange server can filter the traffic before sending it to each DSP, thereby reducing the waste of DSP system resources and improving the cost performance.

Moreover, in the embodiment of the present invention, when filtering the traffic, the traffic of the corresponding category may also be filtered according to a specified ratio according to the category of the traffic. Thus, the waste of DSP system resources can be further reduced.

In addition, in the embodiment of the present invention, after the step of filtering the traffic of the corresponding category according to the specified ratio, the remaining traffic after filtering the traffic of the category according to the specified ratio can also be used as the sampling traffic; every specified time Periodically, obtain the sampling record of each sampled traffic that has received the bidding success notification from the advertising exchange server; extract the traffic characteristics of each sampled traffic from the sampling records; use the preset revenue calculation model to calculate the first value data; judging whether the average value of the multiple first value data is less than the first threshold; if the average value of the multiple first value data is not less than the first threshold, stop filtering the traffic of the category. Therefore, it is possible to sample and correct the filtered traffic, avoid false filtering of traffic, and ensure that each DSP can receive the traffic required by the advertiser.

As for the device embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and for related parts, please refer to the part of the description of the method embodiment.

The algorithms and displays presented herein are not inherently related to any particular computer, virtual system, or other device. Various generic systems can also be used with the teachings based on this. The structure required to construct such a system is apparent from the above description. Furthermore, the present invention is not specific to any particular programming language. It should be understood that various programming languages can be used to implement the content of the present invention described herein, and the above description of specific languages is for disclosing the best mode of the present invention.

In the description provided herein, numerous specific details are set forth. However, it is understood that embodiments of the invention may be practiced without these specific details. In some instances, well-known methods, structures and techniques have not been shown in detail in order not to obscure the understanding of this description.

Similarly, it should be appreciated that in the foregoing description of exemplary embodiments of the invention, in order to streamline this disclosure and to facilitate an understanding of one or more of the various inventive aspects, various features of the invention are sometimes grouped together in a single embodiment, figure, or its description. This method of disclosure, however, is not to be interpreted as reflecting an intention that the claimed invention requires more features than are expressly recited in each claim. Rather, as the following claims reflect, inventive aspects lie in less than all features of a single foregoing disclosed embodiment. Thus, the claims following the Detailed Description are hereby expressly incorporated into this Detailed Description, with each claim standing on its own as a separate embodiment of this invention.

Those skilled in the art can understand that the modules in the device in the embodiment can be adaptively changed and arranged in one or more devices different from the embodiment. Modules or units or components in the embodiments may be combined into one module or unit or component, and furthermore may be divided into a plurality of sub-modules or sub-units or sub-assemblies. All features disclosed in this specification (including accompanying claims, abstract and drawings) and any method or method so disclosed may be used in any combination, except that at least some of such features and/or processes or units are mutually exclusive. All processes or units of equipment are combined. Each feature disclosed in this specification (including accompanying claims, abstract and drawings) may be replaced by alternative features serving the same, equivalent or similar purpose, unless expressly stated otherwise.

Furthermore, those skilled in the art will understand that although some embodiments described herein include some features included in other embodiments but not others, combinations of features from different embodiments are meant to be within the scope of the invention. and form different embodiments. For example, in the following claims, any of the claimed embodiments may be used in any combination.

The various component embodiments of the present invention may be implemented in hardware, or in software modules running on one or more processors, or in a combination thereof. Those skilled in the art should understand that a microprocessor or a digital signal processor (DSP) may be used in practice to implement some or all functions of some or all components in the traffic filtering device according to the embodiment of the present invention. The present invention can also be implemented as an apparatus or an apparatus program (for example, a computer program and a computer program product) for performing a part or all of the methods described herein. Such a program for realizing the present invention may be stored on a computer-readable medium, or may be in the form of one or more signals. Such a signal may be downloaded from an Internet site, or provided on a carrier signal, or provided in any other form.

It should be noted that the above-mentioned embodiments illustrate rather than limit the invention, and that those skilled in the art will be able to design alternative embodiments without departing from the scope of the appended claims. In the claims, any reference signs placed between parentheses shall not be construed as limiting the claim. The word "comprising" does not exclude the presence of elements or steps not listed in a claim. The word "a" or "an" preceding an element does not exclude the presence of a plurality of such elements. The invention can be implemented by means of hardware comprising several distinct elements, and by means of a suitably programmed computer. In a unit claim enumerating several means, several of these means can be embodied by one and the same item of hardware. The use of the words first, second, and third, etc. does not indicate any order. These words can be interpreted as names.

This application discloses A1, a traffic filtering method, comprising:

Receive all traffic sent by the advertising exchange server;

Extract flow features from each flow;

Using the preset revenue calculation model to calculate the first value data corresponding to the traffic characteristics; the revenue calculation model is based on the target value data of each category of traffic and the characteristics of each traffic in the historical traffic records and the corresponding actual value data and Obtained through consumption revenue data training; each traffic in the historical traffic records is the traffic that received the successful bidding notification from the advertising exchange server;

judging whether the first value data is smaller than a first threshold;

When the first value data is less than a first threshold, the traffic is filtered.

A2. The method as described in A1, the step of filtering the traffic includes:

determine the category to which the traffic belongs;

Filter the traffic of the category according to the specified ratio.

A3. The method as described in A1, after the step of filtering the traffic of the category according to a specified ratio, further includes:

Use the remaining traffic after filtering the traffic of the category according to the specified ratio as the sampling traffic;

Obtain the sampling records of each sampling traffic that receives the successful bidding notification from the advertising exchange server at a specified time period;

extracting flow characteristics of each sampled flow from the sampling records;

Using a preset revenue calculation model to calculate the first value data corresponding to the traffic characteristics;

judging whether the mean value of multiple first value data is less than a first threshold;

If the mean value of the plurality of first value data is not less than the first threshold, stop filtering the traffic of the category.

A4. The method as described in A1, before the step of receiving the traffic sent by the advertising exchange server, it also includes:

According to the target value data of each category of traffic and the characteristics of each traffic in the historical traffic records, as well as the corresponding actual value data and consumption revenue data, the revenue calculation model is trained.

A5. The method as described in A4, the step of training the revenue calculation model according to the target value data of each category of traffic and the characteristics of each traffic in the historical traffic records and the corresponding actual value data and consumption revenue data, including:

For each flow, the flow characteristic of the flow is used as an independent variable parameter, and the difference between the consumption income data of the flow and the actual value data, and the ratio of the target value data of the category to which the flow belongs are used as the dependent variable parameter;

Construct the training feature vector of the traffic with the respective variable parameters and dependent variable parameters;

A machine training model is used to train the training feature vector to obtain a revenue calculation model.

A6. The method as described in A5, the step of using the preset revenue calculation model to calculate the first value data corresponding to the traffic characteristics includes:

Substituting the flow characteristics of the flow into the income calculation model to calculate the first value data representing the ratio between the difference between the consumption income and the expected actual value and the target value of the flow.

The application also discloses B7, a flow filtering device, comprising:

The flow receiving module is used to receive each flow sent by the advertising exchange server;

A first traffic feature acquisition module, configured to extract traffic features from each traffic;

The first value data acquisition module is used to calculate the first value data corresponding to the traffic characteristics by using a preset revenue calculation model; the revenue calculation model is based on the target value data of each category of traffic and each traffic in the historical traffic record The characteristics and the corresponding actual value data and consumption revenue data are obtained through training; each flow in the historical flow record is the flow that received the successful bidding notification from the advertising exchange server;

A first judging module, configured to judge whether the first value data is smaller than a first threshold;

A filtering module, configured to filter the traffic when the first value data is less than a first threshold.

B8, the device as described in B7, the filter module, comprising:

A traffic category judging submodule, configured to judge the category to which the traffic belongs;

The filtering sub-module is configured to filter the traffic of the category according to a specified ratio.

B9, the device as described in B7, after the filter module, also includes:

A sampling traffic acquisition module, configured to use the remaining traffic after filtering the traffic of the category according to a specified ratio as the sampling traffic;

The sampling record acquisition module is used to obtain the sampling records of each sampling flow received from the advertising exchange server for bidding success notification every specified time period;

A second traffic feature acquisition module, configured to extract the traffic feature of each sampled flow from the sampling record;

Sampling the first value data acquisition module, configured to use a preset revenue calculation model to calculate the first value data corresponding to the traffic characteristics;

The second judging module is used to judge whether the mean value of multiple first value data is smaller than the first threshold;

A filtering stop module, configured to stop filtering the traffic of the category if the average value of the plurality of first value data is not less than the first threshold.

B10, the device as described in B7, before the flow receiving module, also includes:

The revenue calculation model training module is used to train the revenue calculation model according to the target value data of each category of traffic, the characteristics of each traffic in the historical traffic records, and the corresponding actual value data and consumption revenue data.

B11, the device as described in B10, the income calculation model training module, including:

The parameter confirmation sub-module is used for each flow, taking the flow characteristic of the flow as an independent variable parameter, taking the difference between the consumption income data of the flow and the actual value data, and the target value data of the category to which the flow belongs The ratio of is used as the dependent variable parameter;

The training feature vector construction sub-module is used to construct the training feature vector of the traffic with the respective variable parameters and dependent variable parameters;

The income calculation model training sub-module is used to use the machine training model to train the training feature vector to obtain the income calculation model.

B12. The device as described in B11, the first value data acquisition module, including:

The first value data calculation sub-module is used to substitute the flow characteristics of the flow into the income calculation model, and calculate the first value representing the ratio between the difference between the consumption income and the expected actual value and the target value of the flow. value data.

Claims (10)

1. a kind of traffic filtering method, including：

Receive each flow that advertisement trading server is sent；

Traffic characteristic is extracted from each flow；

The corresponding first value data of the corresponding traffic characteristic are calculated using preset income calculation model；The income calculation The feature of each flow and corresponding real price during model is recorded according to the target value data and historical traffic of flow of all categories Value Data and consumption avail data training obtain；Each flow in historical traffic record is bidded to receive advertisement trading server The flow of success notification；

Judge whether the first value data are less than first threshold；

When described first value data be less than first threshold, then the flow is filtered.

2. according to the method described in claim 1, it is characterized in that, described the step of being filtered the flow, including：

Judge the classification belonging to the flow；

The flow of the classification is filtered according to designated ratio.

3. according to the method described in claim 1, it is characterized in that, being carried out according to designated ratio to the flow of the classification After the step of filter, further include：

Rear remaining flow will be filtered to the flow of the classification according to designated ratio as sampling flow；

Every period specified time, obtain receive advertisement trading server bid success notification each sampling flow sampling note Record；

The traffic characteristic of each sampling flow is extracted from the sample record；

The first value data of the corresponding traffic characteristic are calculated using preset income calculation model；

Judge whether the mean value of multiple first value data is less than first threshold；

If the mean value of multiple first value data stops the filtering to the flow of the classification not less than first threshold.

4. the according to the method described in claim 1, it is characterized in that, each flow for receiving advertisement trading server and sending Before step, further include：

The feature of each flow and corresponding real price in being recorded according to the target value data and historical traffic of flow of all categories Value Data and consumption avail data training income calculation model.

5. according to the method described in claim 4, it is characterized in that, the target value data according to flow of all categories and The feature of each flow and corresponding real value data and consumption avail data training income calculation model in historical traffic record The step of, including：

For each flow, using the traffic characteristic of the flow as independent variable parameter, with the consumption avail data of the flow with The difference of real value data, the ratio with the target value data of the flow generic is as dependent variable parameter；

Respective variable parameter and dependent variable parameter are built to the training feature vector of the flow；

The training feature vector is trained using machine training pattern, obtains income calculation model.

6. according to the method described in claim 5, it is characterized in that, described calculate corresponding institute using preset income calculation model The step of traffic characteristic corresponding first is worth data is stated, including：

The traffic characteristic of the flow is substituted into the income calculation model, expression consumption income and expected real price is calculated First value data of the ratio between the difference of value and flow said target value.

7. a kind of traffic filtering device, including：

Flow receiving module, for receiving each flow of advertisement trading server transmission；

First flow feature acquisition module, for extracting traffic characteristic from each flow；

First value data acquisition module, it is corresponding for calculating the corresponding traffic characteristic using preset income calculation model First value data；The income calculation model is each in being recorded according to the target value data and historical traffic of flow of all categories The feature of flow and corresponding real value data and consumption avail data training obtain；Historical traffic record in each flow be Advertisement trading server is received to bid the flow of success notification；

First judgment module, for judging whether the first value data are less than first threshold；

Filtering module is less than first threshold for working as the first value data, is then filtered the flow.

8. device according to claim 7, which is characterized in that the filtering module, including：

Traffic classes judging submodule, for judging the classification belonging to the flow；

Filter submodule, for being filtered according to designated ratio to the flow of the classification.

9. device according to claim 7, which is characterized in that after the filtering module, further include：

Sampling flow acquisition module is made for will be filtered rear remaining flow to the flow of the classification according to designated ratio For sampling flow；

Sample record acquisition module, for every period specified time, acquisition, which receives advertisement trading server and bids, successfully to lead to The sample record of each sampling flow known；

Second flow feature acquisition module, for extracting the traffic characteristic of each sampling flow from the sample record；

Sampling the first value data acquisition module, for calculating the corresponding traffic characteristic using preset income calculation model First value data；

Whether the second judgment module, the mean value for judging multiple first value data are less than first threshold；

Stopping modular is filtered, if the mean value for the multiple first value data is not less than first threshold, is stopped to the class The filtering of other flow.

10. device according to claim 7, which is characterized in that before the flow receiving module, further include：

Income calculation model training module, for according to each in the target value data and historical traffic of flow of all categories record The feature of flow and corresponding real value data and consumption avail data training income calculation model.