KR101308279B1 - Method for recommendation using category expert - Google Patents

Abstract

A method of determining recommended items using category experts is provided. The score that the recommended user gives to the candidate item is predicted. The predicted score may be calculated based on the scores given by the experts in the category to which the candidate item belongs to the items belonging to the category, the similarity between the recommended user and the expert, and the preferences of the recommended user in the category.

Description

Recommended ways to use category experts {METHOD FOR RECOMMENDATION USING CATEGORY EXPERT}

The present invention relates to a recommendation system. A method of determining a recommended item using a category expert is disclosed.

The recommendation system provides a function of automatically selecting a product satisfactory to a specific user and recommending the selected product to the user. The performance of the recommendation system can be assessed with how accurate the user can select products that are sufficiently satisfactory. In general, since the recommendation system recommends products to users in real time online, the speed of selection and recommendation can also be an indicator of performance.

The collaboration-based recommendation method may be classified into a memory-based recommendation method and a model-based recommendation method. As a memory-based recommendation method, there are k-Nearest Neighbor (NN) based recommendation method and item based recommendation method. Model-based recommendation methods include clustering, classification, or Bayesian networks.

The k-NN-based recommendation method finds k users most similar to the recommended users, and then predicts a score to be given to each of the items by using the scores assigned to the items by the k users. The k-NN based recommendation method calculates online the similarity between each of the recommended user and all other users whenever there is a request for recommendation for a particular recommendation. The k-NN-based recommendation method may have an advantage that an item may be more accurately recommended by reflecting a score that each user has recently given to the item. However, the k-NN-based recommendation method may have a drawback in that it consumes a lot of time in searching for a user similar to the recommended user.

Previously studied collaboration-based recommendation methods, including k-NN based recommendation methods, have performance or accuracy problems. The problem of accuracy may mean that the recommendation is not accurate, especially due to failure to reflect recent evaluations left by the user.

In the following, a recommendation method capable of overcoming a performance problem and an accuracy problem occurring in existing collaboration-based recommendation methods is proposed.

One embodiment may provide a method of predicting a score to be given to an item by a recommended user based on scores awarded to items by an expert in the category.

One embodiment may provide a method of predicting a score to be assigned to an item by a recommended user based on the similarity between the recommended user and the expert of the category.

An embodiment may provide a method of predicting a score to be given to an item by a recommended user based on a preference that the recommended user has in a category.

According to an aspect, a method of determining, by an electronic device, a recommendation item to be recommended to a recommended user among one or more users of the recommendation system, the method comprising: selecting one or more experts in a category among one or more users, and determining a candidate item Estimating a score to be recommended by the recommended user to the candidate item based on scores assigned by the experts to the candidate item, whether to select the candidate item as a recommended item based on the predicted score. Determining a recommendation item and recommending the candidate item to the recommended user when the candidate item is selected as a recommendation item.

The one or more experts may assign a score to the candidate item of the one or more users, and may be the top k users selected in descending order of number of times the items belong to the category, where k is an integer of 1 or more. Can be.

The score to be assigned to the candidate item by the recommended user is an average value of scores awarded to items belonging to the category by the recommended user, a score granted to the candidate item by each of the one or more experts, and It may be determined based on the difference between the average value of the scores given to items belonging to the category.

The score that the recommended user assigns to the candidate item may be determined based on the similarity between the recommended user and each of the one or more experts.

There may be a plurality of categories.

The score to be assigned to the candidate item by the recommended user may be an average value of scores predicted to be assigned to the candidate item by each of the plurality of categories.

The score to be given to the candidate item by the recommended user may be calculated based on the preferences of the recommended user for each of the plurality of categories.

The preferences of the recommended user for each of the plurality of categories may be determined based on the number of items that the recommended user has scored among one or more items belonging to each of the plurality of categories.

A method is provided for predicting a score that a featured user will assign to an item based on the scores that the category's expert assigns to the items.

A method is provided for predicting a score that a recommended user will award an item based on the similarity between the recommended user and the expert in the category.

A method is provided for predicting a score that a recommended user will give to an item based on a preference that the recommended user has in the category.

1 is a flowchart illustrating a recommendation item determining method according to an exemplary embodiment.
2 illustrates methods of predicting a score to be given to a candidate item according to an example.

In the following, embodiments will be described in detail with reference to the accompanying drawings. Like reference symbols in the drawings denote like elements.

1 is a flowchart illustrating a recommendation item determining method according to an exemplary embodiment.

The following steps 110 to 170 may be performed by an electronic device.

The electronic device that determines the recommended item may manage one or more users. One or more users may each be a user of the recommendation system. The electronic device may manage one or more items. One or more items may each be an item provided by the recommendation system. For example, if the recommendation system is an e-commerce site or part of an e-commerce site, one or more users may each be a seller or buyer of the e-commerce site. One or more items may each be an item sold on an e-commerce site. Hereinafter, one or more users of the recommendation system are outlined as users. Name one or more items of the recommendation system as items.

The users may each assign a score to the item (s) of some of the items. The electronic device may use the score given to each item of each of the users in order to select a recommendation item to be described later.

In the following steps, a recommendation item to be recommended to a recommended user among users of the recommendation system may be determined by the electronic device.

In step 110, a category expert may be selected from one or more users of the recommendation system. Here, the category may be a category to which the candidate item determined in operation 130 to be described later belongs. Alternatively, in step 110, an expert in each of one of the categories provided by the recommendation system may be selected. There may be more than one expert selected.

Each item may belong to a specific category among categories provided by the system. The above category may be classified by items or the like. An item may belong to a plurality of categories in common. That is, according to the type of the item or the like, a specific item may belong to a plurality of categories in common. For example, certain athletic shoes may belong to the "clothing" category and the "sporting goods" category. An expert in a category may mean a user having expertise in the above category.

Step 110 may be performed after a candidate item to be described below is determined. That is, after the candidate item is determined in step 130, one or more experts of the category to which the candidate item belongs among one or more users of the recommendation system may be used to predict a score to be given to the candidate item by the recommended user to be described later. Can be selected.

In step 120, the recommended user is determined.

The recommended user may mean a target to which an item is recommended among users of the system. In other words, the recommended user may be named a target user. The recommended item may mean an item that is recommended to the recommender among the items of the system. Also, the recommended item may mean an item that is predicted to be highly preferred by the recommender among the items of the system.

“Scores that a Recommended User Gives to a Candidate Item” may mean “estimated value of the recommended user's preference for the candidate item” or “value predicted by the recommended user for the candidate item”. have.

In step 130, candidate items are determined.

The candidate item may be one of the items of the system. In particular, the candidate item may be an item of which the recommended user has not given a score among the items of the system.

The candidate item may mean an item of which items are checked to be recommended to a recommended user. In general, an item for which the recommended user has not given a score may be determined as a candidate item.

In step 140, the score of the candidate item is determined. Here, the score of the candidate item may be a score that the recommended user assigns to the candidate item. The score of the candidate item may be predicted based on scores assigned to the candidate item by one or more experts in the category to which the item belongs. That is, the score to be given to the candidate item by the recommended user may be predicted based on the scores assigned by the one or more experts to the candidate item.

Specific methods of determining the score of the candidate item are described in detail with reference to FIG. 2 below.

In step 150, it may be checked whether candidate items remain.

The candidate item determined in operation 130 may be one or more. The steps 140-150 may be performed repeatedly to predict the score of each of the one or more candidate items.

If the candidate item remains, step 140 may be performed. If no candidate item remains, step 160 may be performed. If the candidate items belong to different categories, one or more experts in the category to which the new candidate item belongs may be selected before step 140 is repeated.

In step 160, it may be determined whether to select the candidate item as the recommended item based on the predicted score of the candidate item.

If there is more than one candidate item, some candidate items of the one or more candidate items may be selected as recommended items based on the predicted scores of the one or more candidate items.

In operation 170, when the candidate item is selected as the recommended item, the candidate item may be recommended to the recommended user.

If there is more than one candidate item, the candidate items selected in step 160 among one or more candidate items may be recommended to the recommended user.

2 illustrates methods of predicting a score to be given to a candidate item according to an example.

Step 140 described above with reference to FIG. 1 may include one or more of steps 210-230.

Methods (or algorithms) for predicting a score to be given to the candidate item are named "CatExpert", "CatExpert + Sim" and "CatExpert + Pref", respectively.

In step 210, the score to assign to the candidate item is predicted by using CatExpert.

CatExpert may refer to a method of predicting a score to be given to a candidate item by a recommended user based on scores assigned to the candidate item by one or more experts in the category.

CatExpert can predict scores based on Equation 1 below.

Where u represents the recommended user. i represents a candidate item. c indicates a category to which the candidate item belongs.

는 피추천 사용자 u가 시스템의 아이템들 중 카테고리 c에 속하는 아이템들 각각에게 부여한 점수들의 평균 값이다. p _{u, i, c} represents a score predicted to be given to the candidate item i by the recommended user u belonging to category c .

Is an average value of scores given by the recommended user u to each of the items of the system in the category c .

E _c is a collection of one or more experts of category c . E _c may be the top k users selected in descending order of 1) scoring a candidate item i , and 2) scoring the items belonging to category c of the users of the system. Here k is an integer of 1 or more. That is, the one or more experts may be the top k users selected in descending order of numbering the candidate items among the users of the system and scoring the items belonging to the category to which the candidate item belongs.

는 전문가 v가 시스템의 아이템들 중 카테고리 c에 속하는 아이템들 각각에게 부여한 점수들의 평균 값이다. r _{v, i} denotes a score assigned to candidate item i by expert v included in E _c .

Is the average value of the points given by expert v to each of the items in the system in categories c .

That is, according to Equation 1, the score to be given to the candidate item by the recommended user is 1) the average value of the scores given to items belonging to the category to which the candidate item belongs by the recommended user and 2) each of the one or more experts. May be determined based on a difference between a score given to the candidate item and an average value of scores given to items belonging to a category to which the candidate item belongs. That is, it can be predicted that an item that has been given a high score by other experts than other items in the category will be given a relatively higher score by the recommended user.

According to the characteristics of the candidate item, one candidate item may belong to a plurality of categories in common. When a candidate item belongs to a plurality of categories, it is a question of which category of experts' opinions are reflected in predicting the score of the candidate item.

When the candidate item belongs to a plurality of categories, scores given by experts of each of all categories to which the candidate item belongs may be considered as shown in Equation 2 below.

는 후보 아이템 i가 속하는 복수 개의 카테고리들의 개수이다. p _u,i 는, 후보 아이템 i가 속하는 모든 카테고리들이 고려된, 피추천 사용자 u가 후보 아이템 i에 부여할 것으로 예측되는 점수이다.Here, C is a set of a plurality of categories to which the candidate item i belongs.

Is the number of categories to which the candidate item i belongs. p _{u, i} is a score predicted to be given to the candidate item i by the recommended user u , taking into account all categories to which the candidate item i belongs.

That is, according to Equation 2, the score to be given to the candidate item by the recommended user may be an average value of scores predicted to be assigned to the candidate item by each of a plurality of categories to which the candidate item belongs.

CatExpert can reduce the time required for score prediction compared to the existing k- NN based recommendation method. The k- NN based recommendation method should find k users similar to the recommended user for each of the recommended users. In order to find k similar users, the similarity between each of the recommended users and the total users of the system must be calculated. When cosine similarity or correlation coefficient is used for the similarity calculation, much time is consumed.

In CatExpert, on the other hand, for all users who require scores of each of different items belonging to the same category, the same users (ie, experts in the above categories) can be specified for prediction of the scores.

Thus, CatExpert can pre-select one or more experts for each category and can incrementally update one or more experts each time new data is entered. Here, the input of new data may indicate 1) addition of an item, 2) addition of a user, 3) addition of a score assigned to the item by the user. The expert of the category may be determined by the number of scores the user has awarded to each of the items. The above selection and readjustment can easily determine the expert of the category when the user has additionally scored. The operation of selecting an expert in a category may be relatively simple. Thus, the process of finding one or more experts in each of all categories may also be performed in a short time.

In fact, because most users get information about items from the experts, CatExpert can make each user's recommendation more satisfactory than the k- NN-based recommendation method. For example, to see a movie in the science-fiction genre, a user is recommended through the advice of an expert who knows the genre well. Thus, the method of using the expert in the category may be considered to predict the score that each user will give to the item, reflecting the actual process of the user evaluating the item.

In step 220, the score to assign to the candidate item is predicted by using CatExprt + Sim.

CatExpert reflects the user's personal preferences very weakly in score prediction. This reflection is due to the fact that in the prediction of the score of the candidate item, the scores given by the experts of the category to which the candidate item belongs are used in the same manner regardless of the personal taste of the recommended user, so that the score of the candidate item is predicted.

However, each of the users of the system resonates more with the opinions of users with similar tastes as theirs. Therefore, the user may be overpowered to accept the opinions of the experts who have similar tastes among the categories of experts more actively, and less to the opinions of the experts of different categories.

CatExpert + Sim may reflect the propensity of the recommended user as described above in CatExpert. That is, CatExpert + Sim reflects the similarities of each of the featured user and one or more experts in the category in predicting the score that the recommended user will assign to the candidate item.

In order to measure the similarity of taste between the recommended user and the expert, the similarity between the recommended user and the expert in the category may be calculated. The similarity s _u, v between the recommended user u and the expert v may be calculated based on Pearson's correlation coefficient, as shown in Equation 3 below.

Here, I is a set of items that the recommended user u and the expert v commonly score.

는 사용자 u가 모든 아이템에 각각 부여한 점수들의 평균 값이다. r _{u, i} is the score given user u to item i.

Is the average value of the scores that the user u gives to each item, respectively.

CatExpert + Sim can predict scores based on Equation 4 below.

Here, the symbols used in Equation 4 may be the same as the symbols used in Equation 1 described above. Further, p _{u, i, c} in Equation 4 may be applied to Equation 2 described above.

That is, according to Equation 4, the score to be assigned to the candidate item by the recommended user is 1) the average value of the points given to items belonging to the category to which the candidate item belongs by the recommended user and 2) each of the one or more experts. Can be determined based on the difference between the score given to the candidate item and the average value of the scores given to items belonging to the category to which the candidate item belongs and 3) the similarity between each of the featured user and one or more experts.

In step 230, the score to assign to the candidate item is predicted by using CatExprt + Perf.

CatExpert and CatExpert + Sim predict the score by equally reflecting the opinions of experts in each of the plurality of categories when the candidate item belongs to the plurality of categories. However, the degree to which the recommended user is interested in each of the plurality of categories may be different. Recommended users will be more active in accepting feedback from users who have a similar tendency as theirs, and there is a commonality between the recommended users and the experts in the category that the recommended users are interested in. . Thus, it may be assumed that the recommended user will be more actively accepting the opinions of experts of the category in which he or she is more interested.

For example, if the recommended user has a low interest in a specific category, the recommended user has a low interest in the specific category even though the experts in the specific category have given a high score to the candidate item. I will rarely hear feedback. On the other hand, if the recommended user has a high interest in a specific category, the recommended user has a very high interest in the item if the experts of the specific category give a high score to the items belonging to the specific category. Will be seen. Therefore, the accuracy of score prediction can be improved by determining the weight of the opinions of the experts in the category according to the preference (ie, the degree of interest) of the category of the recommended user.

CatExpert + Pref may refer to a method in which the user predicts a score to be given to the candidate item by the recommended user based on a preference for each of a plurality of categories to which the candidate user's candidate item belongs.

The preference for a category can be determined by various methods. For example, a preference for a specific category may be determined based on the number of items that the recommended user scores among one or more items belonging to the category. It may be assumed that the more the item that the recommended user has scored among the one or more items belonging to the category, the more the recommended user prefers the above category.

CatExpert + Pref can predict scores based on Equation 5 below.

Here, the symbols used in Equation 5 may be the same as the symbols used in Equation 2 described above. In addition, f _{u, c} in Equation 5 is a preference that c has in one of a plurality of categories to which the candidate item of the recommended user u belongs.

According to Equation 5, the preference for the category of the recommended user is used as a weight for the score predicted for the category. That is, CatExpert + Pref may synthesize the opinions of the experts of one or more categories to which the candidate item belongs by reflecting a score given by the experts of the category in proportion to the preference of the category of the recommended user.

According to Equation 5, the score to be given to the candidate item by the recommended user is 1) an average value of scores assigned to items belonging to the category to which the candidate item belongs by the recommended user and 2) by each of one or more experts. 3) a difference between an average value of scores given to the candidate item and scores given to items belonging to the category to which the candidate item belongs and 3) preferences of the recommended user for each of the plurality of categories.

The principles of score prediction used by CatExpert + Sim and CatExpert + Pref described above may be applied in an overlapping manner. A score prediction method using both 1) similarity of each of the recommended users and experts and 2) preferences for each of the plurality of categories of the recommended users is named CatExpert + Sim + Pref.

CatExpert + Sim + Pref can predict scores based on Equation 6 below.

Here, p _{u, i, c} may be as described in Equation 4 of CatExpert + Sim. Equation 6 may be defined by applying p _{u, i, c} of Equation 4 to Equation 5.

The technical contents described above with reference to FIG. 1 can be applied as it is, so a detailed description will be omitted below.

The technical contents described above may be embodied in the form of program instructions that may be executed by various computer means and may be recorded in a computer readable medium. The computer readable medium may include program instructions, data files, data structures, etc. alone or in combination. The program instructions recorded on the medium may be those specially designed and constructed for the embodiments or may be available to those skilled in the art of computer software. Examples of computer-readable recording media include magnetic media such as hard disks, floppy disks, and magnetic tape, optical media such as CD-ROMs, DVDs, and magnetic disks, such as floppy disks. Magneto-optical media, and hardware devices specifically configured to store and execute program instructions, such as ROM, RAM, flash memory, and the like. Examples of program instructions include not only machine code generated by a compiler, but also high-level language code that can be executed by a computer using an interpreter or the like. The hardware device may be configured to operate as one or more software modules to perform the operations of the embodiments, and vice versa.

As described above, the technical details described in the embodiments can be variously modified and modified. Therefore, the scope of the invention should not be limited to the described embodiments, but should be determined by the equivalents of the appended claims, as well as the appended claims.

Claims (7)

A method for determining, by an electronic device, a recommended item to recommend to a recommended user from among one or more users of a recommendation system,
Determining a candidate item among items of the recommended system that the recommended user does not give a score to;
For each category of a plurality of categories to which the candidate item belongs, among the users who scored the candidate item among the one or more users, a portion of the top k persons in descending order of the number of times that the items belonging to each category are scored. Determining users as one or more professional users of each category;
Predicting a score that the recommended user will assign to the candidate item based on scores given by the one or more professional users to the candidate item;
Determining whether to select the candidate item as a recommended item based on the predicted score; And
Recommending the candidate item to the recommended user when the candidate item is selected as a recommended item.
Lt; / RTI >
K is an integer of 1 or more, and a score to be assigned to the candidate item is calculated according to Equation 1 below.
[Equation 1]

Here, i denotes the candidate item, u denotes the recommended user, p _{u , i} denotes a score to be assigned to the candidate item considering all of the plurality of categories to which the candidate item belongs, and C denotes the Represents the plurality of categories to which a candidate item belongs, c represents one of the plurality of categories, f _{u , c} is the preference for the category represented by the c of the recommended user,

Is an average value of scores given by the recommended user to each of the items belonging to the category represented by the c among the items of the recommendation system, E _c is a set of one or more professional users of the category represented by the c , and v Is a professional user of one of the set of one or more professional users, r _{v , i} is a score given to the candidate item by the professional user represented by the v ,

Is an average value of scores given to each of the items belonging to the category represented by the c among the items of the recommendation system by the professional user indicated by v , and s _{u , v} are the calculated values based on the relationship coefficient of Person Similarity between the recommended user and the expert represented by v . delete delete The method of claim 1,
The similarity is calculated based on a score given by the recommended user and a score given by each of the one or more professional users for each of items that are commonly scored by each of the recommended user and the one or more professional users. Recommended item decision method. delete delete delete

Images