JP5941875B2 - Recommendation device, method and program - Google Patents

Description

  The present invention relates to a recommendation apparatus, method, and program, and in particular, to recommend a combination of items such as products, software, and services highly evaluated in an E-Commerce (electronic commerce) service to a user group composed of a plurality of users. The present invention relates to a recommendation apparatus, method and program.

  The technology for recommending user groups consisting of multiple users is effective in improving user satisfaction by applying it to services that require use by multiple people, such as restaurant recommendation, travel destination recommendation, and BGM recommendation. It is a technology (see, for example, Non-Patent Document 1). As a recommendation technique for the user group of Non-Patent Document 1, when the product rating information of users belonging to the group is available, each product is evaluated using one of the following four methods using the rating of each user: There is a method to calculate the group score.

Method 1: Average score in the group:
Method 2: Minimum score in the group:
Method 3: The weighted sum of the average value of the scores in the group and the variance value (the larger the average value, the larger the value, and the greater the variance, the smaller the value):
Method 4: Weighted sum of average value and absolute deviation (the larger the average value, the larger the value, and the larger the absolute deviation, the smaller the value):
Recommend a product with a high group score based on the calculated group score. Table 1 shows an example of group score information consisting of users A, B, and C, and statistics used for group score calculation.

表１では各ユーザが入力した評点は０〜５の範囲の実数値としているが、評点の範囲は異なるものでもよく、値も実数値、離散値のいずれでもよい。表1における上記の手法１の各商品のグループ評点の値は平均値の行に対応し、手法２による値は最小値の行に対応する。手法３によるグループ評点は平均値の行から分散値の行の値を重み付けて（例えば0.5を乗算）引いた値、手法４は同様に平均値の行から絶対偏差の行の値を重み付けて引いた値となる。

In Table 1, the score input by each user is a real value in the range of 0 to 5, but the range of the score may be different, and the value may be either a real value or a discrete value. In Table 1, the value of the group score of each product of Method 1 above corresponds to the average value row, and the value according to Method 2 corresponds to the minimum value row. The group score by method 3 is the value obtained by weighting the value of the variance row from the average value row (for example, by multiplying by 0.5), and the method 4 is also weighted by subtracting the value of the absolute deviation row from the average value row. Value.

  Which method to use for group score calculation is appropriately selected according to the recommended policy. For example, if the calculation method using the average value of the method 1 is a method aiming at “maximizing satisfaction within the group” or a calculation method using the minimum value of the method 2, someone in the group is extremely It can be regarded as a method aimed at recommending "minimizing dissatisfaction within the group" in the sense of avoiding products with low scores. The method using the weighted sum of the methods 3 and 4 is a product having a large average and a small variance (or absolute deviation) (= small variation in user's score), that is, a product having the same high score for the entire user. It can be said that it is a method aiming at "balance between satisfaction maximum and dissatisfaction minimization".

  Also, assuming that group action history data is available in addition to user product rating information, a method for creating a recommendation list that takes into account the characteristics of the group, that is, the power relationship between the users of the group and the degree of preference bias (For example, refer to Patent Document 1).

JP 2012-43036 A

Recommendation to Groups, A. Jameson and B. Smyth. The Adaptive Web, 2007.

  However, the techniques of Non-Patent Document 1 and Patent Document 1 described above cannot maintain user fairness when a group wants to select a plurality of products from a recommendation list (a specific example will be described later). This is because, in the method of Non-Patent Document 1, the group score to be calculated is calculated independently for each product, and because the combination of products is not considered, a recommendation list composed of products with a high group score is There is a possibility that only a product with a high score of a specific user will be occupied.

  This is specifically shown using Table 1 above and Table 2 below.

表２は表1の評点を元に商品の組合せに対して評点を記述したものである。表１において平均値をグループ評点として二つの商品の推薦を行う場合、表中の平均値行の値の上位２つの「商品１」、「商品４」が推薦されることとなる。しかし、この組み合わせは望ましい組み合わせではない。なぜなら、表２を見ると｛商品１、商品４｝の組み合わせと｛商品1、商品２｝の組み合わせの評点の平均値は同じ値（6.9）でありながら、分散の値が｛商品１、商品４｝の組みの方がはるかに大きな値となっているためである。表２における各ユーザの評点を見ても｛商品1、商品２｝の組み合わせではユーザＡとユーザＣの差は1.2程度であるが、｛商品1、商品４｝の組ではユーザＡの評点がユーザＣの評点よりも３以上も小さな値となっており、ユーザ間に大きな不公平が生じてしまっている。これはグループに対する複数商品の推薦において、商品の組み合わせに対してグループ評点を定義すべきであることを示唆している。

Table 2 describes scores for combinations of products based on the scores in Table 1. In the case of recommending two products using the average value as a group score in Table 1, the top two “product 1” and “product 4” in the average value row value in the table are recommended. However, this combination is not a desirable combination. Because, when Table 2 is viewed, the average score of the combination of {product 1, product 4} and {product 1, product 2} is the same value (6.9), but the variance is {product 1, product 4} This is because the group 4} has a much larger value. Looking at the score of each user in Table 2, the difference between user A and user C is about 1.2 in the combination of {product 1, product 2}, but the score of user A is in the combination of {product 1, product 4}. The value is 3 or more smaller than the score of the user C, resulting in a large unfairness among the users. This suggests that a group score should be defined for a combination of products in the recommendation of multiple products to the group.

  In the method of Patent Document 1, unfairness is not regarded as an important measure in the first place. Although it is an important approach to focus on the characteristics of the group rather than on fairness, the method of Patent Document 1 cannot be applied to a newly created group without group action history data. Therefore, it is a natural approach to recommend a new group with attention to fairness. However, as described above, the method of Non-Patent Document 1 that does not use group action history data selects multiple products. Fairness among users at the time is not maintained.

As described above, in the prior art, for example, when it is desired to recommend K products out of N products,
(1) For each product, find the average value of the scores for simply selecting the items that all users have a high evaluation for by the users in the group;
(2) Find the minimum score for the user to avoid and choose what was rated too low;
(3) Obtain a weighted difference between the average value of the scores and the variance value for selecting the one having a higher evaluation of all the users and a lower variation in the evaluation of the users;
(4) Obtain a weighted difference between the average value of the scores and the absolute deviation for selecting the one with the highest evaluation of all the users and the lower variation in the evaluation of the users;
As a result, K products are recommended in the order of increasing objective function values. However, this method does not take into consideration the possibility that the average value and the variance value of the individual items have different tendencies. Therefore, for example, when K products were recommended using the objective function of (3), the average value was high and the variance value was high for some products, but as a result, the value of the objective function was high, For some products, even if the average value is not a high value, the variance value may be small, resulting in a high objective function value. As a result, when looking at products selected as having a high objective function, one product is highly rated only by a specific user, and another product is rated moderately by each user. It can be said that it is unsuitable when you want to recommend a combination of multiple products.

  The present invention has been made in view of the above points, and in an E-Commerce service, it is possible to recommend items such as products and services in consideration of both satisfaction and fairness to a group of a plurality of users. It is an object to provide a recommendation device, method, and program.

According to one aspect, a recommendation device for recommending an item assigned an ID for each item capable of electronic commerce,
For a user group consisting of a plurality of users, by presenting an item with an ID to each user in the user group, a score information storage means for storing a score for the thing input from the user for each matter ID;
Group information storage means for storing user IDs belonging to the group;
The objective function to be obtained uses the weighted difference between the average value of the scores and the variance value, or the weighted difference between the average value of the scores and the absolute deviation, and a combination of a plurality of things includes a thing. Expressed by a variable (w) that is a positive real value, and based on the scores of the score information storage means corresponding to users included in the group of the group information storage means, the variables (w ) To change the objective function calculating means for calculating the objective function f (w) including the score value and the goal score value in the combination of things,
A recommendation device is provided that includes a recommendation unit that outputs a variable (w) that maximizes the difference between the score value and the goal score value as an optimal solution.

  According to one aspect, it is possible to recommend a product considering both satisfaction and fairness to a user group composed of a plurality of users, which is impossible with the conventional technology.

  In addition, reducing unfairness promotes longer-term use and purchases of services, leading to significant revenue growth.

It is a flowchart of outline | summary operation | movement of the user group goods recommendation apparatus in one embodiment of this invention. It is an example of composition of a user group goods recommendation device in one embodiment of the present invention. It is an example of the score information table in one embodiment of this invention. It is an example of the group information table in one embodiment of this invention. It is an example of the group recommendation list in one embodiment of the present invention. It is a flowchart of the process at the time of the rating information update in one embodiment of this invention. It is a flowchart of the process at the time of group information update in one embodiment of this invention. It is a flowchart at the time of the group recommendation list creation in one embodiment of this invention. It is a flowchart of the detailed process of step 330 of FIG. 8 in one embodiment of this invention. It is a flowchart of the recommendation process in one embodiment of this invention. It is an example of the recommendation list | wrist in one embodiment of this invention.

  Hereinafter, embodiments of the present invention will be described with reference to the drawings.

  The present invention recommends a product to an E-Commerce service in which a user group consisting of a plurality of users exists, and in order to eliminate unfairness among users who recommend a plurality of products based on the conventional technology, Instead of calculating the group score independently, a new group score considering the combination of products (hereinafter referred to as “combination score”) is defined, and a recommendation list is generated based on this. Specifically, the method based on the weighted sum of the average value and the variance (Method 3) and the method based on the weighted sum of the average value and the variance (Method 4) of Non-Patent Document 1 described above for the combination of products. This is an extended method that can be applied. In the present invention, when a product recommendation is made to a user group, a score is calculated based on a combination of products.

  First, the general operation of the present invention will be described.

  FIG. 1 is a flowchart of an outline operation of a user group product recommendation device according to an embodiment of the present invention.

  Step 1) Select a plurality of products and generate a combination of products. In the following, if the number of all products is N, the purpose is to limit the number of products to be finally recommended to K (N ≧ K).

  Step 2) The score given by the user to each product in the product combination is taken out and applied to a new objective function considering the product combination for evaluation. In addition, a plurality of users assigns a grade with a value in a certain range to each product in advance (for example, 5 levels of 1 to 5).

  Step 3) Determine objective functions for all product combinations.

  Step 4) A combination of products having the maximum objective function is set as a recommended product.

  FIG. 2 is a configuration example of a user group product recommendation device according to an embodiment of the present invention.

  The user group product recommendation device 1 shown in FIG. 1 includes a grade information processing unit 10, a group information processing unit 20, a group recommendation list creation unit 30, a recommendation processing unit 40, a storage unit 50, a communication processing unit 60, and an input / output unit 70. The communication processing unit 60 is connected to the network 2, and the input / output unit 70 is connected to the external device 4 such as an input device or a display device. The network 2 is connected to the external terminal 3.

  The group recommendation list creation unit 30 includes a fairness-considered group score calculation unit 31.

  The storage unit 50 includes a grade information table 51, a group information table 52, and a group recommendation list 53.

  Each table will be described below.

<Grade information table 51>
As shown in FIG. 3, the grade information table 51 has a user ID field, a product ID field, and a grade field. In the user ID field, an identifier for specifying a user who performs the evaluation is set by the score information processing unit 10. In the product ID field, an identifier for specifying which product is related to the product scored by the rating information processing unit 10 is set. In the rating field, the rating information processing unit 10 sets a rating value for the product by the user.

<Group information table 52>
As shown in FIG. 4, the group information table 52 has a group ID field and a user ID field. In the group ID field, an identifier for specifying a group composed of a plurality of users is set by the group information processing unit 20. In the user ID field, the group information processing unit 20 sets identifiers of users included in the group.

<Group recommendation list 53>
The group recommendation list 53 includes a group ID field, a product ID field, and a group recommendation score field as shown in FIG. In the group ID field, an identifier for identifying a group is set by the group recommendation list creation unit 30. In the product ID field, an identifier for specifying a product by the group recommendation list creation unit 30 is set. In the group recommendation score field, a group recommendation score for the product of the group calculated by the group recommendation list creation unit 30 is set.

  The operation in the above configuration will be described.

  In the present embodiment, it is considered that a recommendation list is generated based on the combination group score.

  In the present invention, similar to the concept of “weighted sum of average value and variance value” of Method 3 of Non-Patent Document 1 of the prior art or “weighted sum of average value of score and absolute deviation” of Method 4. However, by using the combination group score, the objective function is evaluated to be small for a set of products for which no combination is taken into account, and is easily excluded as a product recommendation.

  A specific operation will be described below.

  (1) If the number of all products is N, N and K are set (N ≧ K) for the purpose of limiting the number of products to be recommended to K in the end.

  (2) A plurality of users give a grade to each product with a value within a certain range in advance, collect this in the grade information processing unit 10, and store it in the grade information table 51. For example, the score is a value of 5-level evaluation of 1-5.

  The group information processing unit 20 stores users belonging to the group in the group information table 52.

  (3) The group recommendation list creation unit 30 calculates a combined group score by the following method and stores it in the group recommendation list 53 of the storage unit 50.

Let w = (w _j ) _j be a variable w representing a combination of K items for N items. Where w is

を満たす、次元数が全商品数となるベクトルである。jは商品に付与したID番号であり、ｊ＝1,…,Nである。w_jは０または1の値をとり、w_j＝1の場合はその商品ｊが組み合わせとして選択されたことを示す。同時にw_j=１となるのはK個とする。

This is a vector whose number of dimensions is the total number of products. j is an ID number assigned to the product, and j = 1,. w _j takes a value of 0 or 1, and when w _j = 1, it indicates that the product j is selected as a combination. At the same time, it is assumed that w _j = 1 is K.

  (4) The group recommendation list creation unit 30 reads the user score for each group in the group information table 52 from the score information table 51, and sets the combined group score (objective function f (w)) at w as follows. Ask. The objective function f (w) is as follows. The objective function f (w) shown in Equation (1) has a large user average of the score sum (the first term on the right side of Equation (1)) and the square error from the user average of each user (Equation (1) on the right side) It is defined as that which is maximized when the second term) is small, and a combination of products with a large combination group score has high user satisfaction and low unfairness.

但し、

However,

は、商品の組み合わせに対する評点和のユーザ平均である。λ（＞０）は有限の値をとる不公平度調整パラメータであるr_ijはグループ内ユーザiの商品jへの評点であり、Mはグループ内ユーザ数である。式（１）、式(２)におけるiに関する総和は、グループに所属しているユーザに関する総和であり、jに関する総和は全商品に関する総和である。

Is the user average of the score sum for the combination of products. λ (> 0) is an unfairness adjustment parameter that takes a finite value, r _ij is a score for the product j of the user i in the group, and M is the number of users in the group. The sum relating to i in Equations (1) and (2) is the sum relating to users belonging to the group, and the sum relating to j is the sum relating to all products.

  Equation (2), which corresponds to the first term on the right side of Equation (1) above, is the average value of the sum of the scores given to the entire combination of products represented by w by users (M people) in the user group. Yes, it is the score value in the objective function f (w). In addition, the second term of the formula (1) is a variance value indicating the variation in the scores of the users (M people), which is a goal value for the score value. When the value of this term increases, the value of the objective function f (w) Becomes lower. Here, the squared term of the second term of the formula (1) is obtained by subtracting the user's score sum in the product combination w from the group sum of the score sums, and the squared term is calculated for a certain user in the group. When the score sum is too low or too high from the average score sum in the group, it is evaluated as a large value deviating from the average. In order to obtain this as the sum of all users in the group, as the objective function f (w), in the product combination w, there are frequent cases where the user score sum in the group is too high or too low than the average in the group. When the average value within the group of the score sum is large and any user score sum of the members takes a large value close to the average within the group, the objective function can take a high value.

  In other words, the optimal combination w under the above-mentioned equation (1) as an objective function and the number of combination products limited to K can be obtained by solving the discrete optimal problem of equation (2). .

(5) The above-mentioned (3) to (4) are calculated by changing the variable w _j so that all K product combinations can be taken.

(6) Finally, in the recommendation processing unit 40, as the objective function f (w) (formula (1)), “w _j = 1” when “the difference between the score value and the goal score value” is maximized is obtained. The combination of K products is output as K products with the optimal solution according to the purpose.

  The optimal solution is obtained by solving the discrete optimal problem shown in Equation (3).

式（３）の解

Solution of equation (3)

の値を商品ｊのグループ推薦スコアと呼ぶこととする。式（３）により明らかに、

Will be referred to as the group recommendation score for product j. Clearly from equation (3)

の値は０か１であり、かつ、

The value of is 0 or 1, and

となる商品jの数はK個である。これが式（３）の目的関数を最小化する商品ｊの組み合わせである。

The number of products j is K. This is a combination of products j that minimizes the objective function of Equation (3).

  However, the discrete optimal problem equation (Equation (3)) shows a method for obtaining an approximate solution because it is difficult to obtain a solution in a large-scale problem. The group recommendation score is the solution of the linear programming problem (Formula (4))

と定義する。

It is defined as

上記の線形計画問題は、式（４）の3行目、4行目の制約条件により、式（３）では離散値に限定されていた変数ｗの離散値の制約を外すことで近似解を求めている。式（４）の右辺は必ずしも1である必要はなく、任意の有限の正の実数であればよい。）この目的関数はリスク尺度に絶対偏差を用いたポートフォリオ最適化問題と等価であり、線形計画問題として効率的に解くことができる。絶対誤差を用いたことで線形計画問題に帰着することができている。離散最適問題の場合と同じく、

The above linear programming problem solves the approximate solution by removing the constraint on the discrete value of the variable w, which was limited to the discrete value in Equation (3), due to the constraints on the third and fourth rows of Equation (4). Looking for. The right side of Expression (4) does not necessarily have to be 1, and may be any finite positive real number. This objective function is equivalent to a portfolio optimization problem using absolute deviation as a risk measure, and can be efficiently solved as a linear programming problem. The absolute error can be used to reduce the linear programming problem. As with discrete optimal problems,

を商品ｊのグループ推薦スコアと呼ぶこととする。しかし、式（４）においては式（３）と異なり、

Will be referred to as the group recommendation score for product j. However, in Formula (4), unlike Formula (3),

の値は、実数値となる。したがって、この場合に推薦する商品の組み合わせを提示する場合は工夫が必要となるが、

The value of is a real value. Therefore, when presenting the combination of products recommended in this case, it is necessary to devise,

の値は大きい上位K個を選ぶ、もしくは、ある閾値を設けて

Choose a large top K value, or set a certain threshold

がその値を上回る商品ｊを推薦することで、グループ推薦スコアを基に所望の商品の組み合わせが求まる。

By recommending a product j that exceeds that value, a desired combination of products is obtained based on the group recommendation score.

  The processing for each component shown in FIG. 2 is shown below.

<Grade Information Processing Unit 10>
First, the process at the time of updating the score information table 51 will be described.

  FIG. 6 is a flowchart of the process at the time of updating the grade information according to the embodiment of the present invention.

  The update timing of the score information by the score information processing unit 10 may be managed manually by a system administrator, for example, or may be automatically activated when a user evaluates some product.

  Step 110) The score information processing unit 10 inserts a line in which the product ID field, the user ID field, and the score value are set in the score information table 51 in accordance with the evaluated product, the evaluated user, and the score.

<Group Information Processing Unit 20>
Next, processing when the group information processing unit 20 updates group information will be described.

  FIG. 7 is a flowchart of processing when group information is updated according to an embodiment of the present invention.

  The group information update timing by the group information processing unit 20 may be manually managed by, for example, a system administrator, or automatically when a new group is created or a new user is added. You may make it start.

  Step 210) The group information processing unit 20 inserts a row in which the values of the group ID field and the user ID field are set according to the edited group and the user to which the group information table 52 belongs. The group information editing process refers to a process of inserting a row including a new group ID that has not existed in the group information table 52 until now.

<Group recommendation list creation unit 30>
Next, processing when the group recommendation list creation unit 30 creates a group recommendation list will be described.

  FIG. 8 is a flowchart when creating a group recommendation list according to the embodiment of the present invention.

  The recommendation list creation timing by the group recommendation list creation unit 30 may be manually managed by a system administrator, or may be automatically performed periodically. In the embodiment of the present invention, a case is considered in which scores are given to all {user, item} pairs. Even if the {user, item} pair of grades does not exist, the grades are complemented by the method described in Non-Patent Document 2 (Variational Bayesian Approach to Movie Rating Prediction. YJ Lim and YW The., 2007.) The present invention can be applied as it is.

  Step 310) The group recommendation list creation unit 30 refers to the group information table 52 and sets the values of all the group IDs appearing in the group ID field in the group set G of the memory (not shown).

  Steps 320 to 340 are executed for each group g appearing in the acquired group set G.

  Step 320) The group recommendation list creation unit 30 refers to the group information table 52, and stores all values of the user ID field appearing in the same row as the group ID field of the group g in the user set U of the memory (not shown). Set. Let M be the number of elements in U, that is, the number of users belonging to the group.

Step 330) The fairness considering group score calculation unit 31 calculates the group recommendation score gr _gj for each product j of the group g using the evaluation information table 51 and the group information table 52. Specifically, the calculation is performed using any one of the above formulas (3) and (4).

Step 340) The group recommendation list creation unit 30 configures as many rows as the number of products recommended in the group recommendation list 53, and for each recommended product j, the group ID of the group g in the group ID field and the product in the product ID field. The group recommendation score gr _gi is stored in the product ID and group recommendation score field of j. Here, as a criterion for recommending or not recommending a product, a predetermined number of items may be recommended in order from the largest group recommendation score, and a threshold is set for the group recommendation score. A product whose group recommendation score exceeds the value may be recommended.

  The process of step 330 will be described in detail.

  FIG. 9 is a flowchart of detailed processing of step 330 in the embodiment of the invention.

Step 331) The fairness-considered group score calculation unit 31 determines that the temporary variables f (objective function parameters in the linear programming problem), A (variable inequality constraint parameters in the linear programming problem), A _eq (equality constraint parameters), b ( The variable inequality constraint parameter), b _eq (equal constraint parameter), and lb (inequality constraint parameter) in the linear programming problem are initialized according to the equation (5), and the parameter λ is set (J is the total number of items). The parameter λ is set according to a constant given in advance. The temporary variables f, A, A _eq , b, b _eq , and lb are assumed to be stored in a memory (not shown).

ステップ３３２） 公平度考慮グループスコア算出部３１は、評点情報テーブル５１からユーザ集合Uに所属する各ユーザuの各商品（計J個）の評点

Step 332) The fairness-considered group score calculation unit 31 evaluates each product (total of J items) of each user u belonging to the user set U from the rating information table 51.

を取得し、Aを式（６）にしたがって更新する。A_eqを式（７）にしたがって更新する。

And A is updated according to equation (6). Update A _eq according to equation (7).

ステップ３３３） 公平度考慮グループスコア算出部３１は、ｆ，A，A_eq，ｂ，b_eq，lbを用いて式（８）の線形計画問題を解く。線形計画問題の解法には、例えば、単体法が利用できる（非特許文献３：Practical Methods of Optimization, 2nd Edition. R. Fletcher, Wiley, 1987.参照）。式（８）は、線形計画問題の一般形であり、入力ｆ，A，A_eq，ｂ，b_eq，lbを与えることで、式（８）の制約条件下で目的関数f^Txを最小化する解x_optを求める（式（８）はダミー変数を導入することで、式（1）から直接導出される）。

Step 333) The fairness-considered group score calculation unit 31 solves the linear programming problem of Expression (8) using f, A, A _eq , b, b _eq , and lb. For example, a simplex method can be used to solve the linear programming problem (see Non-Patent Document 3: Practical Methods of Optimization, 2nd Edition. R. Fletcher, Wiley, 1987.). Equation (8) is a general form of the linear programming problem. Given inputs f, A, A _eq , b, b _eq , and lb, the objective function f ^T x is minimized under the constraints of Equation (8). solving x _opt for reduction (equation (8) is to introduce a dummy variable is derived directly from equation (1)).

ステップ３３４） 公平度考慮グループスコア算出部３１は、各商品ｊに対して解xoptの第M+j+1要素を抜き出し、グループgの商品jに対するグループ推薦スコアgr_giに設定する。これは、第1成分から第M+1成分にはダミー変数を導入しているためである。グループ推薦スコアの商品に関する総和は、

Step 334) The fairness-considered group score calculation unit 31 extracts the M + j + 1th element of the solution xopt for each product j, and sets it as the group recommendation score gr _gi for the product j of the group g. This is because dummy variables are introduced from the first component to the M + 1th component. The total sum of products with group recommendation scores is

を満たす。

Meet.

<Recommendation processing unit 40>
Next, processing during recommendation processing will be described.

  FIG. 10 is a flowchart of processing during recommendation processing according to an embodiment of the present invention.

  The recommendation process may be executed when a recommendation execution request is input from the external terminal 3 to the user group product recommendation device 1 via the communication processing unit 60 and the network 2, for example. The request argument at this time is the group ID to be recommended. Assume that there is a request for recommendation regarding group g.

  Step 410) The recommendation processing unit 40 refers to the group recommendation list 53, acquires (a plurality of) product IDs in a row having the group ID of the group g, and recommends products.

  The number of products to be recommended in the recommendation list may be, for example, recommending the top four items based on the value of the group recommendation score, or setting a threshold value and recommending products exceeding the recommended value. An example of a recommendation list output by the recommendation processing unit 40 is shown in FIG.

  In the above, the method of considering the combination of products based on the objective functions of the methods 3 and 4 of Non-Patent Document 1 of the prior art is shown. However, the method is not limited to this example, and the method 2 of the prior art is used. It is also applicable to “Minimum score”.

  When applying the combination of products to “Minimum In-Group Score” in Method 2 of Non-Patent Document 1,

上記において、b_iは商品iを組み合わせに含む"１"、含まない"０"のフラグであり、r_mはグループ内ユーザmが商品iに付与した評価値のベクトルを表す。

In the above, b _i is a flag of “1” and “0” not including product i in the combination, and r _m represents a vector of evaluation values given to product i by the user m in the group.

The meaning of the above formula is to calculate the score f for a combination of a certain product {b _i } for each user, find the minimum user score f, and then change the product combination {b _i } in various ways Then, the value of the combination {b _i } having the maximum score f is obtained.

  As described above, by applying the combination of “1” including products to “0” not including the products in the method 2, it is possible to recommend a product that is not an excessive evaluation of the user and that is an optimal combination. .

  In the above embodiment, an example of recommending a product is shown. However, the present invention is not limited to this example. For example, an ID number is assigned to each of services, TV programs, software, and the like. Anything that can be identified can be recommended by the device.

  Further, the operation of each component of the user group product recommendation device shown in FIG. 2 of the above embodiment is constructed as a program and installed in a computer used as the user group product recommendation device, or executed. It is possible to circulate through.

  The present invention is not limited to the above-described embodiments, and various modifications and applications are possible within the scope of the claims.

DESCRIPTION OF SYMBOLS 1 User group goods recommendation apparatus 2 Network 3 External terminal 4 External apparatus 10 Rating information processing part 20 Group information processing part 30 Group recommendation list preparation part 31 Fairness consideration group score calculation part 40 Recommendation processing part 50 Storage part 51 Rating information table 52 Group information table 53 Group recommendation list 60 Communication processing unit 70 Input / output unit

Claims (11)

A recommendation device for recommending an item given an ID for each item capable of electronic commerce,
For a user group consisting of a plurality of users, by presenting an item with an ID to each user in the user group, a score information storage means for storing a score for the thing input from the user for each matter ID;
Group information storage means for storing user IDs belonging to the group;
The objective function to be obtained uses the weighted difference between the average value of the scores and the variance value, or the weighted difference between the average value of the scores and the absolute deviation, and a combination of a plurality of things includes a thing. Expressed by a variable (w) that is a positive real value, and based on the scores of the score information storage means corresponding to users included in the group of the group information storage means, the variables (w ) To change the objective function calculating means for calculating the objective function f (w) including the score value and the goal score value in the combination of things,
A recommendation means for outputting a variable (w) that maximizes the difference between the score value and the score loss value as an optimal solution;
A recommendation device characterized by comprising: The objective function calculating means includes
When calculating the score value, the average value of the scores given to each thing by each user is used, and when calculating the goal score, the scores given to each thing by each user are combined. The recommendation apparatus according to claim 1, further comprising means for using a variance value accumulated for all items and all users, or an absolute deviation of each user's expression. The objective function calculating means includes
The recommendation device according to claim 1, further comprising means for deriving a value of the variable (w) that maximizes the value of the objective function by a linear programming problem. The objective function calculating means includes
The range of the variable (w) is not a real value, but expressed as a binary discrete value of “0” when not included, “1” when included, and the number of variables (w) equal to 1. The recommendation apparatus according to claim 1, further comprising means for deriving a variable (w) that maximizes the objective function by a discrete value optimization problem when the value is limited. A recommendation device for recommending an item given an ID for each item capable of electronic commerce,
For a user group consisting of a plurality of users, by presenting an item with an ID to each user in the user group, a score information storage means for storing a score for the thing input from the user for each matter ID;
Group information storage means for storing user IDs belonging to the group;
In the variable (b), the minimum value of the score is used for the objective function to be obtained, and a combination of a plurality of things is “0” when the thing is not included and “1” when the combination is included. Expressing and summing the value obtained by multiplying the variable (b) and the score of the score information storage means for each user belonging to the group information storage means to obtain a score for the combination of things, and changing the variable Minimum value calculating means for obtaining the score f when the minimum value is obtained;
In the minimum value calculation means, a recommendation means for outputting a variable (b) having the maximum score f as an optimum solution;
A recommendation device characterized by comprising: A recommendation method for recommending an item with an ID for each item capable of electronic commerce,
For a user group consisting of a plurality of users, by presenting an item with an ID to each user in the user group, a score information storage means for storing a score for the thing input from the user for each matter ID;
Group information storage means for storing user IDs belonging to the group;
In an apparatus having an objective function calculation means and a recommendation means,
The objective function calculation means uses a weighted difference between the average value of the scores and the variance value or a weighted difference between the average value of the scores and the absolute deviation as the objective function to be obtained, and a combination of a plurality of things Is expressed by a variable (w) that is a positive real value when including an object, and all items are based on the score of the score information storage unit corresponding to a user included in the group of the group information storage unit An objective function calculating step of calculating an objective function f (w) including a score value and a goal value in a combination of things by changing the variable (w) for the combination of
A recommendation step in which the recommendation means outputs a variable (w) that maximizes the difference between the score value and the goal score value as an optimal solution;
A recommendation method characterized by performing. In the objective function calculating step,
When calculating the score value, the average value of the scores given to each thing by each user is used, and when calculating the goal score, the scores given to each thing by each user are combined. The recommendation method according to claim 6, wherein a variance value accumulated for all the users and all users or an absolute deviation of each user's expression is used. In the objective function calculating step,
The recommendation method according to claim 6 or 7, wherein the value of the variable (w) that maximizes the value of the objective function is derived by a linear programming problem. In the objective function calculating step,
The range of the variable (w) is not a real value, but expressed as a binary discrete value of “0” when not included, “1” when included, and the number of variables (w) equal to 1. The recommendation method according to claim 6 or 7, wherein a variable (w) that maximizes the objective function is derived by a discrete value optimization problem when the value is limited. A recommendation method for recommending an item with an ID for each item capable of electronic commerce,
For a user group consisting of a plurality of users, by presenting an item with an ID to each user in the user group, a score information storage means for storing a score for the thing input from the user for each matter ID;
Group information storage means for storing user IDs belonging to the group;
In an apparatus having minimum value calculation means and recommendation means,
The minimum value calculation means uses the minimum value of the score for the objective function to be obtained, and is a binary discrete value of “0” when a combination of a plurality of things does not include a thing and “1” when a combination is included. The value obtained by multiplying the variable (b) and the score of the score information storage means by the user unit belonging to the group information storage means, and representing the score in the combination of things A minimum value calculating step for determining a score f when the variable is changed and the minimum value is obtained;
A recommendation step in which the recommendation means outputs, as an optimal solution, a variable (b) in which the score f is maximized in the minimum value calculating step;
A recommendation method characterized by performing. Computer
The recommendation program for functioning as each means of the recommendation apparatus of any one of Claims 1 thru | or 5.

Images