KR101453623B1 - Method and system for inferencing item preference using search keyword of user in recommendation system - Google Patents

Abstract

An item preference inferring method and system using a user's search keyword in a recommendation system are disclosed. A method for inferring a user's item preference includes: inputting a search keyword from a user by an item preference reasoning system; providing information about a searched item based on the search keyword; And calculating the user's preference for each item using information on the item selected by the user.

Description

TECHNICAL FIELD [0001] The present invention relates to an apparatus and method for estimating an item preference using a user's search keyword in a recommendation system,

Embodiments of the present invention relate to a method and system for inferring a user's preference for an item using a search keyword used by a user in a recommendation system.

In general, the recommendation system is a system that finds items that users prefer from among various items. Most of the existing recommendation systems have found users who have similar scores to the ratings left by the target user, and have recommended items to the target user based on the ratings given to the items by the users.

For example, Korean Patent Laid-Open Publication No. 10-2010-0092852 (published on August 23, 2010) entitled " Product recommendation system based on preference and method thereof " The method comprising: providing information about a product and receiving an evaluation of the product provided to the user to determine whether the product is different from the plurality of products based on at least one of the user's evaluation, the user's preference, It is disclosed that information on one product is provided.

However, since most users are bothered to leave a rating on an item, it is difficult to grasp the taste score because only a small number of items leave the rating score or rating. Therefore, the existing recommendation systems can only estimate the rating to be given to the item by the target user using only a small number of ratings left by the users, and as a result, the target user can not predict the rating to be left on the item, I have a cold start problem.

Therefore, there is a need for a way for a user to infer a rating to be given to an item even if the user does not leave or leave a low rating on the item.

Disclosed is a method and system for reasoning an item preference using a user's search keyword in a recommendation system in which a user can infer a rating to be given to an item even if the user does not leave or leave a small number of ratings on the item.

Disclosed is a method and system for reasoning an item preference using a user's search keyword in a recommendation system in which a user can accurately recommend an item preferred by a target user even if the user does not leave or leave a small number of ratings on the item.

A method for inferring a user's item preference includes: inputting a search keyword from a user by an item preference reasoning system; providing information about a searched item based on the search keyword; And calculating the user's preference for each item using information on the item selected by the user.

According to one aspect, the step of calculating may include calculating a user's preference for each item by giving a predetermined score to the item based on information about the item selected by the user.

According to another aspect, there is provided a method of searching for similar users, comprising the steps of: calculating similarity between users based on the calculated preference after the calculating; searching for a similar user of the target user based on the calculated similarity; And predicting a preference of the target user for the target item based on at least one of a preference for the target item and a rating given to the target item.

According to yet another aspect, there is provided a method of calculating similarity, comprising: calculating a similarity between items based on the calculated preference after the calculating; retrieving a similar item of the target item based on the calculated similarity; And predicting the preference of the target user for the target item based on the preference for the item and the rating given to the similar item.

According to another aspect, the method may further include recommending an item to the target user based on the predicted target item preference after the predicting step.

A method for inferring a user's item preference includes a step of the item preference reasoning system receiving a search keyword from a user, a step of providing information on an item searched based on the search keyword, and a step of providing information And calculating the user's preference for each item on the basis of the number of times provided.

The system for inferring a user's item preference includes a search keyword input unit for inputting a search keyword from a user, an item information providing unit for providing information about an item searched based on the search keyword, And a preference calculating unit for calculating the preference of the user for each item based on at least one of the information about the item selected by the user and the number of times the information about the item is provided to the user for each item.

The user's preference is calculated for each item by using the information about the item selected by the user or the number of times the information about the item is provided to the user for each item among information about the searched item based on the search keyword, Even if you do not leave or leave a few, you can deduce a particular user's rating for a specific item.

The degree of similarity between the users or the similarity between the items is calculated using the user's preference calculated for each item, and by recommending the item using the similarity, the user can accurately select the item preferred by the target user even if the user does not leave a rating on the item, can do.

1 is a flowchart illustrating an item preference reasoning method according to an exemplary embodiment of the present invention.
2 is an exemplary diagram showing an item of interest to a user.
3 is a flowchart illustrating a method of recommending an item by calculating similarity between users based on an item preference deduced through a search keyword in an embodiment of the present invention.
4 is a flowchart illustrating a method of recommending an item by calculating similarity between items based on an item preference deduced through a search keyword according to an exemplary embodiment of the present invention.
5 is a diagram illustrating an item recommendation method according to an embodiment of the present invention.
6 is a block diagram illustrating an item preference reasoning system in one embodiment of the present invention.
7 is a graph illustrating the accuracy of an item recommendation method and an existing user-based recommendation method according to an embodiment of the present invention.
FIG. 8 is a graph illustrating the number of items that can not be recommended for an item recommendation method and an existing user-based recommendation method according to an embodiment of the present invention.

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS Hereinafter, embodiments of the present invention will be described in detail with reference to the accompanying drawings.

FIG. 1 is a flowchart illustrating an item preference reasoning method according to an exemplary embodiment of the present invention, and FIG. 2 is an exemplary diagram illustrating an item that a user is interested in.

Existing item recommendation methods have grasped how much a particular user likes a particular item, that is, a user 's taste, based on ratings given by users. However, since most users leave only a small number of ratings, existing item recommendation methods suffer from a cold start problem. Therefore, the item preference inferring method according to the present invention uses a search keyword used by a user to grasp a user's favorite item, that is, to grasp users' preferences and use them instead of a rating, To be resolved.

To this end, the item preference reasoning system according to the present invention can receive a search keyword from a user as shown in FIG. 1 (110). In addition, information on the searched item may be provided to the user based on the input search keyword (120).

Then, the user's preference can be calculated 130 for each item using the information about the item selected by the user among the information about the provided item (130). At this time, the item preference reasoning system can calculate the user's preference for each item by giving predetermined score to the item based on the information about the item selected by the user among the information about the item provided to the user.

In calculating the preference, the item preference reasoning system may calculate a user's preference for each item based on the number of times information about the item is provided to the user for each item. However, in this case, since the difference between the preferences for items associated with a small number of search keywords and the preferences for items associated with a plurality of search keywords can be calculated to be very large, It is also possible to reduce the difference in preference between items by recomputing using a sigmoid function or recalculating using a log function.

Accordingly, since the existing item recommendation method uses only the ratings assigned by the users in the calculation of the preferences, it is difficult to find out the items preferred by the user when the user is given a low rating to the item. However, The preference degree can be calculated not only for the items to which the user has been given the rating but also for the items to which the rating is not given.

For example, when users who use the online shopping mall search for an item of interest through a search, each user can input a search keyword indicating an item of interest, as shown in FIG. Hereinafter, four methods for inferring preferences will be described with reference to FIG.

First preference calculation method

A preference degree is calculated by assigning predetermined scores to selected items among the retrieved items through a search keyword used by a user. Because the items associated with the search keyword can be thought of as items that the user is interested in.

Referring to FIG. 2, a first user (User 1) uses a first search keyword (Query 1) and a second search keyword (Query 2), through which a first item (Item 1), a second item 2), the third item (Item 3), the fifth item (Item 5), and the sixth item (Item 6). Therefore, the first user can think that the first item, the second item, the third item, the fifth item and the sixth item have a preference of "1" and the remaining items have a preference of "0"Quot; 1 "points can be given to the preferences of the first user for the first item, the second item, the third item, the fifth item, and the sixth item.

How to calculate second preference

And determines how many times each item is searched through search keywords used by the user. If the same item appears as a search result for a plurality of search words used by the user, it can be determined that the user is more interested in the item. Therefore, the preference for each item of the target user can be determined by the number of search keywords related to the item among the search keywords used by the user.

 For example, in FIG. 2, a first user uses a first search keyword and a second search keyword, wherein the first item is a search result of a first search keyword and a second search keyword. Therefore, the preference for the first item of the first user is "2", the preference for the second item, the third item, the fifth item and the sixth item of the first user is "1" Quot; 0 ".

However, in the second preference calculation method, the difference between the preferences for items associated with a small number of search keywords and the preferences for items associated with a plurality of search keywords can be calculated to be very large. For example, if there are an item a associated with 100 search keywords used by the user, and an item b associated with one keyword, the preference of these two items may differ by a maximum of 100 times. However, since the user does not actually like item a more than item b by 100 times, the difference in preference between items can be reduced for more accurate inference reasoning.

Third preference calculation method and fourth preference calculation method

The preference calculated through the second preference calculation method is recalculated using a sigmoid function or a log function to re-adjust the preference determined through the second preference calculation method to give a new preference . For example, the third preference calculation method and the fourth preference calculation method may use a sigmoid function as shown in Equation 1 below and a log function as in Equation 2, respectively.

In each expression, p _{u , i} is the number of search keywords associated with item i among the search keywords used by user u, and p ' _{u , i} is the newly calculated preference by recalculating the preference calculated through the second preference calculation method .

The item preference inferring method according to the present invention can calculate the preference not only for the items to which the user has given the direct rating but also for the other items by inferring the preference through the above-mentioned four methods. In addition, the preference can be calculated not only for the items preferred by the users who have given low rated items but also for the items preferred by the users who have not rated the items.

3 is a flowchart illustrating a method of recommending an item by calculating similarity between users based on an item preference deduced through a search keyword according to an embodiment of the present invention. FIG. 6 is a flowchart illustrating a method of recommending an item by calculating similarity between items based on the item preference deduced through a search keyword; FIG.

The existing item recommendation method grasps the taste of the user through the ratings given to each product by the users, and uses the result to predict the rating of the product that the user has not left yet. Then, the user selects a satisfactory item from the predicted rating and recommends it to the user. One of the most popular methods for recommending items is user based recommendation and item based recommendation.

First, referring to FIG. 3, a method of calculating similarity between users using a user-based recommendation method based on an item preference deduced through a search keyword and recommending an item will be described.

In the user-based recommendation method, k similar users (k is a natural number) having the similar taste to the target user who is the target of the recommendation are searched, and then the scores of the items And a method of predicting the score to be given by the target user to the user.

The first step is to calculate the similarity between the target user and the rest of the users by using the rating given to the item by the users and to find the k most similar users. In the second step, the users similar to the target user It is a step to synthesize the scores and to predict the rating that the target user will give to the target item.

For example, the user-based recommendation method uses Pearson's correlation coefficient or cosine similarity when calculating similarity to find similar users. At this time, the rating p _{u ,} i to be given to the target item u by the target user u is predicted by the following Equation (3). The similarity between the two users is determined by the fact that more information Can be calculated more accurately.

는 사용자 u가 모든 아이템에 부여한 평점의 평균을 각각 나타낸다. sim(u, v) 는 타겟 사용자 u와 사용자 v의 유사도, S_u 는 타겟 사용자 u에 대하여 유사도가 높은 k명의 사용자를 각각 나타낸다.Where r _{v , i} is the rating given by user v to item i,

Represents the average of ratings given to all items by user u. sim (u, v) represents the similarity between the target user u and the user v, and S _u represents the users with high similarity to the target user u.

However, since the user-based recommendation method uses the rating given by the users when calculating the similarity between the users, it is difficult to find the items preferred by the user in the case of a user who has given a low rating to the item. Therefore, if the search keyword is used instead of the rating given by the user, the user's preference can be grasped for a larger number of items when using the rating, so that the similarity between users can be calculated more accurately.

Therefore, the item recommendation method according to the present invention calculates the similarity between users using Equation (4) based on the inferred inference based on the item preference inference method at the time of item recommendation based on the user-based recommendation method as an example (310). The similar user of the target user may be searched based on the similarity calculated through Equation (4) (320).

는 사용자 u의 모든 아이템들에 대한 선호도의 평균을 각각 나타낸다.Here, sim (u, v) is the degree of similarity between user u and user v, p _{u , i} is the preference of user u for item i,

Represents the average of the preferences for all items of user u.

Then, in a second step, a rating or a preference for a target item of the target user can be predicted (330) based on the ratings or preferences given to the target item by users similar to the target user.

The existing user-based recommendation method can not predict the rating or preference of the item if the user similar to the target user at this stage does not leave a rating on the target item or show preference. However, the item recommendation method according to the present invention estimates the rating or preference of the target item of the target user based on the item preference deduced through the search keyword, Predictability is possible. However, since the preference may be inaccurate because the user does not express his own opinion, it may be more accurate to use the rating in terms of the accuracy of the prediction. Therefore, the item recommendation method according to the present invention can use either the preference or the rating, or both the preference and the rating. The following Equation (5) means the preference predicted for the target item i of the target user u when the inferred preference through the item preference inferring method according to the present invention is used.

는 사용자 u의 모든 아이템들에 대한 선호도의 평균이고, S_u 는 타겟 사용자 u에 대하여 유사도가 높은 k명의 사용자를 나타낸다.Here, sim (u, v) is the similarity between user u and user v, and p _{u , i} is the preference of user u for item i.

Is the average of the preferences for all items of user u, and S _u represents k users with high similarity to target user u.

If a target user's rating or preference for the target item is predicted through the above process, the user can recommend the item to the target user based on the preference for the predicted target item 340. If the user does not leave a rating The user can accurately recommend a favorite item of the target user.

Next, referring to FIG. 4, a method of calculating similarity between items using an item-based recommendation method based on the item preference deduced through a search keyword and recommending an item will be described.

The item-based recommendation method is a method of predicting the degree of similarity between the items by using the ratings given by the users and estimating the rating of the target item by summing the ratings given by the target user to items similar to the target item. There are two major stages.

The first step is a step of calculating the degree of similarity between items, and the second step is a step of estimating a rating of a target item by summing the ratings given by the target user for items similar to the target item. In other words, the user-based recommendation method and the item-based recommendation method finally recommend items with a high rating to the target user.

For example, the item-based recommendation method calculates the similarity between all the items in advance, and compiles the ratings assigned by the target user for items similar to the target item to predict the rating of the target item. In this method, similarity between items is calculated by using Pearson's correlation coefficient, cosine similarity, etc. based on the rating given to each item by users in order to obtain similarity between items. That is, the degree of similarity is determined by how similar users evaluate the two items. Based on the similarity between the items calculated through the item-based recommendation method, the rating p _{u , i} to be given to the target item i by the target user u is predicted by the following equation (6).

Here, the higher the similarity between the sim (i, j) is the item i and j, S _i represents the set of similar items and the item i.

However, since the conventional item-based recommendation method also uses the rating given by the users when calculating the similarity between items, it is difficult to find the items preferred by the user when the user is given a low rating.

Therefore, the item recommendation method according to the present invention calculates the similarity between items using the following Equation (7) based on the preference inferred through the item preference inference method described in the item recommendation based on the item-based recommendation method (410). Then, items similar to the target item may be searched based on the similarity between the items calculated through Equation (7) (420).

는 사용자 u의 모든 아이템들에 대한 선호도의 평균을 각각 나타낸다.Here, sim (i, j) is the similarity degree between item i and j, p _{u , i} is user u's preference for item i,

Represents the average of the preferences for all items of user u.

Then, in a second step, the rating or preference of the target item may be estimated 430 by summing the ratings or similarities of the target users to the items similar to the target item. Item recommendation methods based on item-based recommendation methods can use both ratings and preferences as well as user-based recommendation methods. The following Equation (8) represents the preference predicted for the target item i of the target user u when the inferred preference through the item preference inferring method according to the present invention is used.

는 사용자 u의 모든 아이템들에 대한 선호도의 평균, S_i 는 타겟 아이템 i와 유사한 k개의 아이템을 각각 나타낸다.Here, sim (i, j) is the similarity degree between item i and j, p _{u , i} is user u's preference for item i,

The mean, S _i of preference for all items of the user u k are respectively the number of items similar to the item and the target i.

If a target user's rating or preference for the target item is predicted through the above process, the user is recommended to recommend the item to the target user based on the preference for the predicted target item (440) The user can accurately recommend a favorite item of the target user.

5 is a diagram illustrating an item recommendation method according to an embodiment of the present invention.

The item preference inference method according to the present invention can obtain various recommendation results as shown in the table according to the recommendation method applied. The item recommendation method according to the present invention is based on a user-based recommendation method and an item-based recommendation method. Based on the preference calculation method used in the similarity calculation of each recommendation method, four results can be obtained, and the rating or preference can be predicted based on the result. At this time, the calculation method of the preference used for calculating the similarity and predicting the preference should be the same. Also, in the case of predicting a rating using a rating, various methods of calculating the preferences used in calculating the similarity can be used in a variety of ways, and thus 16 recommended results can be obtained as shown in FIG. After the rating or preference is predicted, items rated as high ratings or preferences may be recommended to the target user.

6 is a block diagram illustrating an item preference reasoning system in one embodiment of the present invention.

Referring to FIG. 6, the item preference reasoning system 600 according to the present invention may include a search keyword input unit 610, an item information providing unit 620, and a preference calculating unit 630.

The search keyword input unit 610 receives the search keyword from the user and the item information providing unit 620 provides the information about the searched item to the user based on the search keyword inputted through the search keyword input unit 610 .

For example, when the item preference reasoning system 600 is applied to an online shopping mall or the like, the search keyword input unit 610 can receive a search keyword from a user through a web page of an online shopping mall, and the item information providing unit 620 searches As a result, information on the retrieved item can be provided through the web page. At this time, the item preference reasoning system 600 may store information on items provided to the user in a predetermined database, information on items selected by the user, and the like.

The preference calculation unit 630 calculates a preference value based on the information about the item selected by the user or the number of times the information about the item is provided to the user for each item among information on the item provided by the item information providing unit 620 And calculates the user's preference.

For example, when calculating the preference based on the information about the item selected by the user, the preference calculating unit 630 may calculate a user's preference for each item by giving a predetermined score to the item. However, if the preference is calculated based on the number of times the information about the item is provided to the user, the difference between the preferences for items associated with a small number of search keywords and the preferences for items associated with a plurality of search keywords can be calculated to be very large , The preference calculation unit 630 may recalculate the calculated preference based on the number of times the information about the item is provided to the user using the sigmoid function or the log function.

The preferences calculated by the preference calculator 630 may be applied to an item recommendation method such as an existing user-based recommendation method, an item-based recommendation method, and used to recommend an item to a user.

Although not shown in FIG. 6, the item preference reasoning system 600 may further include a similar user searching unit, a preference predicting unit, and an item recommending unit to recommend an item to a user based on a user-based recommendation method.

In this case, the similar user searching unit may calculate similarities between users based on the preferences calculated by the preference calculating unit 530, and search for similar users of the target users based on the calculated similarities.

The preference predicting unit may predict a preference for a target item of a target user based on at least one of a preference for the target item of the similar user retrieved from the similar user retrieval unit and a rating given to the target item, And may recommend the item to the target user based on the preference for the predicted target item.

Meanwhile, although not shown in FIG. 6, the item preference reasoning system 600 may further include a similar item retrieval unit to recommend items to the user based on the item-based recommendation method.

In this case, the similar item search unit may calculate similarities between the items based on the preferences calculated by the preference calculation unit 530, and may search for similar items of the target items based on the calculated similarity between the items.

At this time, the preference predicting unit can predict the preference of the target user for the target item based on the preference of the similar user item of the target user retrieved from the similar item retrieval unit and the rating given to the similar item.

FIG. 7 is a graph showing the accuracy of an item recommendation method and an existing user-based recommendation method according to an embodiment of the present invention. FIG. 8 is a graph showing the accuracy of an item recommendation method according to an embodiment of the present invention, It is a graph showing the number of items that can not be recommended.

Referring to FIG. 7, a thirteenth embodiment of a method shown in FIG. 5 (a method of calculating similarity between users using a third preference calculation method and predicting a preference for a target item of a target user) Based recommendation methods of the present invention. For the experiments, the actual online data was collected and the accuracy was measured using leave-one-out cross validation. As a result, we confirmed how many answers were recommended. In FIG. 7, the x-axis represents the recommendation of each method in the 100th, 300th, and 500th positions, and the y-axis represents the recommended number of correct answers in each range. User-based recommendation method is superior to 100th place, but since then, the accuracy of the proposed method is better.

FIG. 8 shows a result of comparing the number of correct answers that can not be recommended with the thirteenth embodiment and the existing user-based recommendation method. The x-axis represents each recommendation method, and the y-axis represents the number of correct answers that can not be recommended in each range. As a result, it can be seen that the item recommendation method according to the present invention has fewer correct answers than the user based recommendation method, and the cold start problem is less.

While the present invention has been particularly shown and described with reference to exemplary embodiments thereof, it is to be understood that the invention is not limited to the disclosed exemplary embodiments. For example, it is to be understood that the techniques described may be performed in a different order than the described methods, and / or that components of the described systems, structures, devices, circuits, Lt; / RTI > or equivalents, even if it is replaced or replaced.

Therefore, other implementations, other embodiments, and equivalents to the claims are also within the scope of the following claims.

Claims (12)

delete delete delete delete delete A method for inferring a user's item preference,
An item preference reasoning system receiving a search keyword from a user;
Providing information about the searched item based on the search keyword;
Calculating a user's preference for each item based on the number of times the information about the item is provided to the user for each item; And
And recalculating a user's preference calculated in consideration of the number of search keywords related to the provided item in order to minimize a user's preference difference between items according to the number of times provided to the user,
Wherein the recalculating comprises:
Calculating a difference between the user's preferences calculated for each item according to the number of search keywords related to the item,
After the recalculation step,
Based on the average of the preferences of the items related to the target user and the similarity between the items based on the similar users according to the similarity between the users to predict the preference for the target items of the target user, An item preference inference method using an average of user preferences. delete delete The method according to claim 6,
After the recalculation step,
Calculating a degree of similarity between users based on the calculated preference;
Searching for a similar user of the target user based on the calculated similarity; And
Predicting a preference of the target user for the target item based on at least one of a preference for the target item of the similar user and a rating given to the target item
Wherein the item preference inference method further comprises: The method according to claim 6,
After the recalculation step,
Calculating a degree of similarity between items based on the calculated preference;
Retrieving a similar item of the target item based on the calculated similarity; And
Predicting a preference of the target user for the target item based on a preference of the target user for the similar item and a rating given to the similar item
Wherein the item preference inference method further comprises: 11. The method according to claim 9 or 10,
After the predicting step,
Recommending an item to the target user based on the predicted preference for the target item
Wherein the item preference inference method further comprises: A system for inferring a user's item preference,
A search keyword input unit for inputting one or more search keywords from a user;
An item information providing unit for providing information on the searched item based on the search keyword; And
A preference calculation unit for calculating a preference of the user for each item based on the information about the item selected by the user and the number of times the information about the item is provided to the user for each item,
Lt; / RTI >
The preference calculation unit may include:
The user's preference calculated by considering the number of search keywords related to the provided item is recalculated in order to minimize the user's preference difference between the items according to the number of the search keywords after calculating the user's preference for each item and,
After said recalculation step,
Based on the average of the preferences of the items related to the target user and the similarity between the items based on similar users based on the similarity between the users to predict the preference for the target items of the target user, An item preference reasoning system using an average of preferences given by a user.

Images