JP6566515B2 - Item recommendation system and item recommendation method - Google Patents

Description

  The present invention relates to an item recommendation system and an item recommendation method. Specifically, in collaborative filtering (CF) typified by a user base or item base, against a shilling attack, that is, an attack that illegally inputs a fake user to intervene in the process of determining an item recommended to the user by the system The present invention relates to a robust item recommendation system and item recommendation method.

The user-based CF is a system that recommends an item to a user by using, for example, a k-nearest neighbor method for similarity calculation and referring to past evaluation values of other users with similar preferences for the item. In other words, k other users who are similar in how to give an evaluation value to an item are selected, an evaluation value related to the item is predicted, and an item with a high evaluation value is recommended to the user.
However, for example, when the item is a product and the evaluation value is a preference level, a fake user designed to have an average preference level is inserted into a user-based CF system (shilling attack called average attack). Since the fake user becomes a hub user that shows a high degree of similarity with any user, that is, an influencer, there is a possibility that a product that the fake user likes is always recommended.
A shilling attack called segment attack or popular attack is effective for item-based CF, that is, a recommendation system and a recommendation method for determining an item to be recommended to a user by referring to the user's past evaluation of other similar items. have.

On the other hand, the inventors have proposed a method of reducing the hub by the k-nearest neighbor method. That is, a method of applying a Laplacian-based kernel to a similarity measure for a large-scale high-dimensional data set (see Non-Patent Document 1), a method of applying centering (see Non-Patent Document 2), and local centering Has proposed a method (see Non-Patent Document 3).
By applying these hub mitigation methods to user-based CF or item-based CF, even if a fake user is introduced by an attacker to illegally increase the evaluation of the target item, the evaluation of the target item It is expected that an item recommendation system and an item recommendation method that prevent fluctuations can be provided.

“Ikumi Suzuki, Kazuo Hara, Masashi Shimbo, Yuji Matsumoto, Marco Saerens,“ Investigating the Effect of Produced-Lanced Intensive In Kern. ” 26th AAAI Conference on Artificial Intelligence, pp. 1112-1118, 2012 Satomi Suzuki, Kazuo Hara, Hitoshi Shinho, “Similarity Measure to Reduce Hubs with k-Nearest Method”, Information Processing Society of Japan Research Report, Natural Language Processing Study Group, 2012-NL-209, No. 11, pp. 1-8, 2012 Kazuo Hara, Ikumi Suzuki, Masashi Shimbo, Kei Kobayashi, Kenji Fukumizu, Milos Radovanovic, “Localized Centering: Reducating HubLense. 29th AAAI Conference on Artificial Intelligence, pp. 2645-2651, 2015

When a user-based CF receives an attack called average attack, that is, an attack in which a large number of fake users showing high similarity with any user are injected, the product preferred by the fake user may be recommended at any time. There is.
In addition, the item-based CF has an attack called a segment attack or a popular attack, that is, a target item having a high degree of similarity to a popular item in a specific topic (for example, a topic such as an action movie or a horror movie). For this reason, when an attack is performed in which a large number of fake users who give high evaluation values to both the target item and the popular item are input, the target item is easily recommended to a user who likes an item belonging to the topic.
The recommendation as described above is unnatural and has a problem of hindering the original function of the recommendation system.

The present invention uses a similarity measure in which the appearance of a hub is suppressed, or converts a given similarity measure so that the hub is less likely to appear. By suppressing the appearance of items serving as sensors and reducing their influence, the evaluation value of the target item is not changed as intended by the attacker.
An object of the present invention is to provide an item recommendation system and an item recommendation method that are less affected by an attack as a result even when subjected to an attack that introduces a fake user.

  In order to solve the above-described problem, the item recommendation system 1 according to the first aspect of the present invention is an evaluation in which an evaluation value R (u, i) related to an item i of the user u is entered as shown in FIG. An evaluation matrix storage unit 21 that stores the matrix R, a first similarity calculation unit 131 that calculates similarity between users using a similarity measure that suppresses the appearance of a hub, and a first similarity calculation unit 131 Are extracted by the first neighborhood data extraction unit 141 and the first neighborhood data extraction unit 141 that extract k users from the ones with higher similarity to the target user, using the similarity degree calculated in (1). A first evaluation value predicting unit 151 that predicts an evaluation value to be entered in an unfilled cell related to the target user using an evaluation value related to the items of k users, and a first evaluation value predicting unit 151 A high evaluation value predicted by It extracts an item to be recommended to the target user from Temu, and a item recommendation unit 16 to be recommended to the target user.

Here, the item is typically a product or service. Furthermore, conditions such as limiting the type of product or service, the provision period, the provision region, and the price range (summer fruits, movies released in X month, etc.) may be defined. However, it is not limited to goods or services, and may be animals and plants, mountains, cities, architecture, paintings, music, theater, martial arts, academics, productivity, and effects as long as they can be evaluated.
The matrix R is typically an evaluation matrix of the number of users × the number of items. As the evaluation value R (u, i), typically, the preference degree related to the item i of the user u is used. However, it is not limited to the degree of preference, and it is sufficient if it can be quantitatively evaluated. For example, the degree of contribution to health, the value of real estate, or the time required for the destination may be used. Further, quantitative evaluation may be expressed by rank and level.

  The similarity measure includes all those that can be used as a measure for measuring the similarity between two data. Typically, dot product, cosine, Pearson correlation, distance are used. The inner product is a scalar product of two vector data, and the cosine is an inner product of vector data normalized to length 1. The Pearson correlation is an inner product of vector data normalized to length 1 after subtracting the element sum from each element value so that the element sum becomes zero. It also includes kernels (various that are mainly called in the field of machine learning) that can be regarded as generalizations of inner products. A typical distance is the Euclidean distance (L2 norm) between two vector data, but also includes a generalized distance (Manhattan distance, Lp norm, etc.). Further, the similarity level output by a similarity score calculation method (BLAST or the like) appropriately determined by a person having domain knowledge according to the purpose of each task is also included in the similarity scale here. These are called general similarity measures.

In addition, as a similarity measure that suppresses the appearance of the hub, for example, a similarity measure that is converted so that all data objects are equally similar to the data center, that is, a similarity measure that does not have Spatial Centrality is applicable. For example, for the above general similarity measure, transformed by applying “(global) centering” that moves the origin to the average (global centroid) of the data set, with the origin as the center of the local subset And applying “local centering” that moves to the local centroid. In addition, Laplacian-based kernels, such as “commuted time kernels” (Marco Saerens, Francois Focuses, Luh Yen, and Pierre du ce nt and sir ent es s s s s s s s and s i s e s s s s s s s e s s s s s e s s s s s s s s s s s s s s s s s s s s s s s s s s s s s s s s s s s s s s s s s s s s s s This corresponds to the one converted by applying Machine Learning (ECML), pp. 371-383 (2004).
In addition to the similarity measure that is converted so that all data objects are equally similar to the data center, there are mutual proximity, local scaling, and the like as the similarity measure that suppresses the appearance of the hub.

In addition, “extracting k users from those with a high degree of similarity to users” typically means extraction using the k-nearest neighbor method. An arbitrary numerical value can be used as k. For example, in a movie lens data set, 30 <= k <= 100 is preferable and 40 <= k <= 70 in order to increase prediction accuracy as a result of supervised learning. Is more preferable, and k = 50 is most preferable (see FIG. 7).
Moreover, when predicting an evaluation value, an average value of k people can be used. Furthermore, when using the average value, it is preferable to use a weighted average value as described later. For weighting, for example, a similarity between users, a coefficient depending on the season (the quality of the fruit is influenced by the season), and the like can be used.

  The user-based CF is not yet referred to the past evaluation values of the k nearest users (k data from the closest one extracted by the k-nearest neighbor method (kNN) and the data having the highest similarity). It is the recommendation system of the form which estimates the user's evaluation value with respect to an evaluation item. A possible drawback of user-based CF is vulnerability to shilling attacks. Shilling attacks throw fake users into the recommendation system to force the system to make biased (arbitrary) recommendations by attackers. The user-based CF does not make a recommendation based on the feature of the item, but makes a recommendation based on a past evaluation value of another user for the item. For this reason, the user-based CF is vulnerable to an attack that attempts to change the decision of the recommended item by the system by designing a fake user to be similar to any user.

On the other hand, it was found that hub data is likely to appear in high-dimensional data sets as a result of so-called “curse of dimension” (Milos Radovanovic, Alexandros Nanopoulos, Nirjana Ivanovic. High-Dimensional Data ", Journal of Machine Learning Research, pp. 2487-2531, 2010). That is, in a high dimension, a small number of data called hubs frequently appear in the kNN of other data. In a user-based CF system, when kNN is calculated, each user is represented as a vector having a dimension of the number of items, but since there are generally many items, the vector is a high-dimensional vector. Therefore, data that serves as a hub (hub data) appears. Since the hub user contributes to the recommendation process as an influencer, it greatly affects the determination of recommended items by the recommendation system.
A shilling attack can be seen as an attack using a hub. In the attack on the user-based CF, an influencer, that is, a fake user serving as a hub is introduced for the purpose of intentionally controlling the determination of recommended items by the system. Specifically, a fake user is designed and put in a manner similar to the data center related to the user.

  Therefore, in order to avoid the impact of the attack, the similarity measure used to determine kNN is transformed so that all users are equally similar to the data center, so that the appearance of the hub user, ie the influencer itself We propose to disable attackers' attempts to send fake users to the system as influencers. Several methods for suppressing the appearance of hubs have been proposed, but can be achieved, for example, by calculating a commute time kernel from a given similarity matrix, or more simply by centering the similarity matrix. Using a movie lens data set, it was confirmed that a fake user tends not to be a hub user after applying this method (see FIGS. 8 and 9). As a result, such similarity scale conversion provides a system that is resistant to attack without degrading the accuracy of item recommendation (see FIG. 7).

  When configured in this manner, the similarity is calculated using a similarity scale that suppresses the appearance of the hub, so the appearance of the hub can be suppressed. It is possible to provide an item recommendation system that is rarely received.

  In order to solve the above-mentioned problem, the item recommendation system 1 according to the second aspect of the present invention is an evaluation in which an evaluation value R (u, i) related to an item i of a user u is entered, as shown in FIG. An evaluation matrix storage unit 21 that stores the matrix R, a second similarity calculation unit 132 that calculates the similarity between items using a similarity measure that suppresses the appearance of a hub, and a second similarity calculation unit 132 Are extracted by the second neighborhood data extraction unit 142 that extracts k items from the one with the higher similarity to the target item, and the second neighborhood data extraction unit 142 using the similarity calculated in A second evaluation value predicting unit 152 that predicts an evaluation value to be entered in an unfilled cell related to the target user using the evaluation values of the target user related to the k items, and a second evaluation value predicting unit Evaluation predicted at 152 Extracts an item to be recommended to the target user from the highest item comprises an item recommendation unit 16 to be recommended to the target user.

In the first aspect, the evaluation value is obtained based on the similarity between users, but in this aspect, the evaluation value is obtained based on the similarity between items. In the first aspect, an average value of k people is used, but in this aspect, an average value of k items is used. However, other system configurations are the same as in the first aspect, and as in the first aspect, the appearance of a hub can be reduced. Therefore, even if a fake user is entered, the recommended item is determined as a result of the fake user. It is possible to provide an item recommendation system that is not easily affected by input, that is, robust against attacks.
With this configuration, since the similarity measure that suppresses the appearance of the hub is used, it is possible to provide an item recommendation system that is less affected by an attack even when subjected to an attack that introduces a fake user. .

In addition, the item recommendation system 1 according to the third aspect of the present invention includes a similarity measure storage unit 22 that stores a similarity measure that suppresses the appearance of a hub in the first or second aspect.
If comprised in this way, since similarity is calculated using the similarity scale which suppresses the appearance of the hub memorize | stored in the system, the appearance of a hub can be suppressed and the influence by a fake user at the time of item recommendation can be reduced. .

The item recommendation system according to the fourth aspect of the present invention converts the similarity based on a general similarity scale into a similarity based on a similarity scale that suppresses the appearance of the hub in the third aspect. A similarity scale conversion unit 135 is provided.
Here, for similarity conversion, for example, a general similarity scale formula is converted to a similarity scale formula that suppresses the appearance of the hub, and the similarity is calculated, or the general similarity scale is used. A method of converting the similarity into a similarity based on a similarity measure that suppresses the appearance of a hub by a matrix is used.
If comprised in this way, the appearance of a hub can be suppressed by converting and using a general similarity scale to the similarity scale which suppresses the appearance of a hub, and the influence by a fake user at the time of item recommendation can be reduced. .

The item recommendation system according to the fifth aspect of the present invention should be filled in when predicting an evaluation value to be written in an unfilled cell related to the target user in any of the first to fourth aspects. A weighted average value is used as the evaluation value.
Here, for example, similarity between users or items, a coefficient by season (the quality of the fruit is influenced by the season), and the like can be used for weighting. When configured as in this aspect, prediction accuracy can be improved.

  In order to solve the above-described problem, the item recommendation method according to the sixth aspect of the present invention fills in an evaluation value R (u, i) related to the item i of the user u as shown in FIG. An evaluation matrix storage step (S104) for storing an evaluation matrix R to be performed, a first similarity calculation step (S107) for calculating a similarity between users using a similarity measure for suppressing the appearance of a hub, a first A first neighborhood data extraction step (S108) for extracting k users from those having a higher similarity to the target user using the similarity calculated in the similarity calculation step (S107) of the first, First evaluation value prediction that predicts an evaluation value to be entered in an unfilled cell of the target user using the evaluation value of the k users' items extracted in the neighborhood data extraction step (S108) Step (S109) From high prediction evaluation value item in the evaluation value prediction step (S109) extracting the item to be recommended to the target user of, and a item recommendation process to be recommended to the target user (S110).

This aspect is an item recommendation method corresponding to the item recommendation system according to the first aspect.
By configuring as in this aspect, it is possible to reduce the appearance of hub users, and thus it is possible to provide an item recommendation method that is less affected by an attack as a result even when subjected to an attack that introduces a fake user.

  In order to solve the above-described problem, the item recommendation method according to the seventh aspect of the present invention fills in the evaluation value R (u, i) related to the item i of the user u as shown in FIG. An evaluation matrix storage step (S104) for storing an evaluation matrix R to be performed, a second similarity calculation step (S207) for calculating a similarity between items using a similarity measure that suppresses the appearance of a hub, and a second A second neighborhood data extraction step (S208) for extracting k items from the higher similarity to the target item using the similarity calculated in the similarity calculation step (S207) of the second, A second evaluation value for predicting an evaluation value to be entered in an unfilled cell relating to the target user, using the evaluation value of the target user relating to the k items extracted in the neighborhood data extraction step (S208) Prediction process (S20 And an item recommendation step (S110) for extracting an item to be recommended to the target user from the items having a high evaluation value predicted in the second evaluation value prediction step (S209) and recommending it to the target user. .

This aspect is an item recommendation method corresponding to the item recommendation system according to the second aspect.
By configuring as in this aspect, it is possible to reduce the appearance of hub items, and thus it is possible to provide an item recommendation method that is less affected by an attack as a result even when subjected to an attack that introduces a fake user.

The item recommendation system according to the eighth aspect of the present invention is the sixth or seventh aspect, wherein the similarity based on the general similarity measure is changed to the similarity based on the similarity measure that suppresses the appearance of the hub. A similarity scale conversion step (S106) for conversion is provided.
With this configuration, it is possible to suppress the appearance of the hub by converting the similarity based on the general similarity scale into the similarity based on the similarity scale that suppresses the appearance of the hub, and the fake user at the time of item recommendation Can reduce the influence of.

  A program according to the ninth aspect of the present invention is a program for causing a computer to execute the item recommendation method according to any of the sixth to eighth aspects.

  A recording medium according to the tenth aspect of the present invention is a computer-readable recording medium recording the program according to the ninth aspect.

  ADVANTAGE OF THE INVENTION According to this invention, even if it receives the attack which throws in a fake user, it can provide the item recommendation system and the item recommendation method which are hardly influenced by an attack as a result.

It is a figure which shows the example of the evaluation matrix R. It is a figure which shows the correlation between users. 2A is a diagram for explaining the correlation between the user u and the user v, FIG. 2B is a scatter diagram when the correlation is strong, and FIG. 2C is a scatter diagram when the correlation is weak. . Is a diagram showing an N ₁₀ distribution in low-dimensional and higher-dimensional. 3 (a) is low-dimensional, 3 (b) is a histogram of the _{N 10} in high level. It is a scatter diagram showing a similarity relationship to N ₁₀ value and the data center. FIG. 4A is a diagram in a low dimension, and FIG. 4B is a diagram in a high dimension. It is a figure which shows the structural example of the item recommendation system 1 in Example 1 and Example 2. FIG. It is a figure which shows the example of a processing flow of the item recommendation method in Example 1 and Example 2. FIG. FIG. 6A is a diagram illustrating an example of a processing flow in the first embodiment, and FIG. 6B is a diagram illustrating an example of a processing flow in the second embodiment. It is a figure which compares the item recommendation system which used the nearest neighbor parameter k as a horizontal axis | shaft, made the average absolute error (MAE) the vertical axis | shaft, and used a different similarity measure. When general Pearson correlation is used as a measure of similarity between users, it is a diagram (part 1) showing that the input fake user is a hub, including all sincere users (users other than fake users) it is a scatter diagram relating to the similarity of the N ₅₀ and the data center about the user. When general Pearson correlation is used as a similarity measure between users, it is shown that the input false user becomes (a) a hub, and (b) and (c) it becomes difficult to become a hub by conversion of the similarity measure. It is a figure (the 2). FIG. 9A is a histogram regarding N ₅₀ when a general Pearson correlation is used as the similarity measure between users. FIG. 9B is a histogram regarding N ₅₀ after conversion by centering, and FIG. 9C is a histogram regarding N ₅₀ after conversion by the commuting time kernel.

  Embodiments of the present invention will be described below with reference to the drawings. In the drawings, the same or corresponding parts are denoted by the same reference numerals, and redundant description is omitted.

[User-based CF]
The matrix R of the number of users N _user × the number of items N _{item is} a data set composed of user's past responses (evaluation values) to the items. R (u, i) represents an evaluation value for the i-th item of the u-th user. The matrix R contains a blank with no value called nil. This nil value means that the user's item has not yet been rated. In general, the matrix R is blank and most is nil. The user-based CF (and the item-based CF described later) predicts these values using the k-nearest neighbor method.

式（１）はｕ番目のユーザのｉ番目のアイテムへの評価値Ｒ（ｕ，ｉ）を予測する予測関数Ｐｒｅｄ（ｕ，ｉ）を示す。ユーザｕとユーザｖ間の類似度をＳｉｍ（ｕ，ｖ）、類似度Ｓｉｍのもとでユーザｕと最近傍となるｋ人のユーザの集合をＵとし、Ｕに属するユーザｎについて使用する評価値は、Ｒ（ｎ，ｉ）≠ｎｉｌを満たす。

Expression (1) represents a prediction function Pred (u, i) for predicting an evaluation value R (u, i) for the i-th item of the u-th user. Assume that the similarity between the user u and the user v is Sim (u, v), the set of k users closest to the user u under the similarity Sim is U, and the user n belonging to U is used for evaluation The value satisfies R (n, i) ≠ nil.

である。

はユーザｕが評価したアイテムに対する平均の評価値である。δ〔．〕は〔〕内の条件が満たされれば１、それ以外は０となる指示関数である。 further

It is.

Is an average evaluation value for the items evaluated by the user u. δ [. ] Is an indicator function that is 1 if the condition in [] is satisfied, and 0 otherwise.

  FIG. 1 shows an example of the matrix R. FIG. 1A shows an example before the fake user is input, and FIG. 1B shows an example after the fake user is input. The item i is arranged in the column direction and the user u is arranged in the row direction, and the evaluation value R (u, i) is entered in the cell (column) that is the intersection. Here, the evaluation value R (u, i) is evaluated by an integer of 5 levels from 1 to 5. In the average attack for the purpose of raising the evaluation value of the target item, as shown in the lower part of FIG. The fake user who gives a value close to the average of the evaluations given by the users who gave the evaluations is input.

A method of measuring the similarity between users by the Pearson correlation of evaluation values given to items will be described with reference to FIG. FIG. 2A is a diagram for explaining the correlation between the user u and the user v. Item 2 (R (u, 2) = 1, R (v, 2) = 1) and item 3 (R (u, 3) = 5, R (v, 3) shown in FIG. = 4) is plotted in FIG. Figure 2 (b), (c) the total item i (R (u, i) , R (v, i)) is a scatter diagram plotting, FIG. 2 (b) of the user u and user v (positive (C) is an example when the correlation between the user u and the user v is weak. When the correlation is strong, the plots of all items are on a straight line, and when the correlation is weak, they are at random.

[Inter-user similarity measure commonly used for user-based CF]
In the user-based CF, it is important to appropriately select the function Sim (.,.) That gives the similarity between users. This is because the similarity function determines the weight of the user who enters kNN and the user who enters kNN according to Equation (1).
As a general similarity measure function, there is a cosine similarity for calculating a cosine (cos) of an angle formed by a row vector (a vector of evaluation values given by each user) after replacing nil of the matrix R with 0. This is shown in equation (2).

ここに、ｘ_ｕはN_item次元ベクトルで、その成分はＲ（ｕ，ｉ）≠ｎｉｌならばｘ_ｕ（ｉ）＝Ｒ（ｕ，ｉ）、それ以外はｘ_ｕ（ｉ）＝０となる。上記関数使用の１つの欠点は、各々のユーザｕがアイテムに与える平均的評価

の違いに基づくバイアスが無視されるという点である。それ故、上記欠点の修正方法として、各ベクトル成分から

を差し引く方法が一般的に使用される。

Here, x _u is an N _item dimensional vector, and its component is x _u (i) = R (u, i) if R (u, i) ≠ nil, and x _u (i) = 0 otherwise. . One drawback of using the above function is that the average rating that each user u gives to the item

The bias based on the difference is ignored. Therefore, as a correction method for the above-mentioned defects, from each vector component

The method of subtracting is generally used.

を差し引いたベクトルを用いた類似度は、式（３）のように計算され、これはユーザ間のピアソン（Ｐｅａｒｓｏｎ）相関と呼ばれる。

ここに、もし、Ｒ（ｕ，ｉ）≠ｎｉｌかつＲ（ｖ，ｉ）≠ｎｉｌならば、

であり、そうでなければｘ’_ｕ（ｉ）＝０、ｘ’_ｖ（ｉ）＝０となる。 User bias

Similarity using a vector obtained by subtracting is calculated as shown in Equation (3), which is called Pearson correlation between users.

Here, if R (u, i) ≠ nil and R (v, i) ≠ nil,

Otherwise, x ′ _u (i) = 0 and x ′ _v (i) = 0.

[Attack on CF]
A shilling attack called average attack is effective for a user-based CF, that is, a recommendation system and a recommendation method for determining items recommended to a user by referring to past evaluations of other users similar to the user. . If the evaluation value matrix R possessed by the system is altered by adding an illegal evaluation value, the recommended item is changed. An attack that introduces a fake user for this purpose is called a shilling attack, and the average attack is one of them. When subjected to this attack, every user becomes highly similar to a fake user. That is, the fake user is an influencer who has an influence on the determination of the recommended item, that is, a hub user.
The fake user thrown in the average attack behaves like the target item (attack target item), and acts like an honest user (user other than the fake user) on the other items. That is, the fake user gives a high evaluation value point to the target item and gives an average evaluation value to the remaining items. As a result, the fake user likes the target item and becomes more similar to any sincere user. Therefore, when the user-based CF receives an average attack, the user-based CF easily recommends a target item to which a fake user gives a high evaluation to all sincere users.

  A shilling attack called segment attack or popular attack is effective for item-based CF, that is, a recommendation system and a recommendation method for determining an item recommended to a user by referring to a user's past evaluation of other similar items. have. Upon receiving this attack, the target item to be attacked becomes highly similar to a popular item that is highly evaluated by any user. The attacker misuses that the popular item is an influencer that influences the determination of the recommended item, that is, a hub item, and tries to change the evaluation value of the target item by the system to be unreasonably high.

[Hub phenomenon]
The hub phenomenon is one of the phenomena that occurs as a result of the “dimensional curse”. Let D be a set of d-dimensional data, and N _k (x) represents the number of times data xεD in D falls within kNN of other data in D. As the dimension d increases, the distribution shape of _Nk changes to have a long tail on the right (see FIG. 3). Alternatively, a small number of data has a large N _k value. Data showing such a large N _k value is called a hub, and this phenomenon is called hubness (hub phenomenon).

Here, the hub phenomenon will be described using an artificial data set. In the recommendation system, each user generally gives an evaluation value to only a few items, so the evaluation matrix R is a sparse matrix with many blanks, but a sparse data set is artificially generated to mimic this situation. did. The data set consists of 2000 pieces of data, each of which is a d-dimensional vector. The method of generating each data is as follows: First, each dimension i = 1,. . . , D, a positive real number generated according to the Lognomal (5; 1) distribution is rounded to obtain an integer ni. Then, select the n _i pieces randomly from 2000 pieces of data, with respect to each, to generate a uniform random number in the range [0,1], i-th element of each vector it (i dimension Component).

3, the angle between the vectors of the similarities between the data, that is, when measured with cos (cosine similarity) _is a histogram showing the _{N 10} distribution. FIG. 3A is a histogram in the case of low dimensions, and FIG. 3B is a histogram in the case of high dimensions. In order to explain the appearance of the hub phenomenon, N ₁₀ distributions were compared in two cases, when the dimension was low (d = 50) and when the dimension was high (d = 1000). FIG. 3 shows that data having a large N ₁₀ value appears in a high dimension, and as a result, the distribution of N ₁₀ is distorted (not symmetric). The maximum N ₁₀ is 38 in FIG. 3A and 133 in FIG.

Figure 4 is a scatter diagram showing a similarity relationship to N ₁₀ value and the data center. FIG. 4A is a diagram in a low dimension, and FIG. 4B is a diagram in a high dimension. Since a strong correlation is observed in the high dimension between the N ₁₀ value and the similarity to the data center, the hub phenomenon is originated from a bias to the data center occurring in the high dimension, that is, Spatial Centrality. I understand that.

[Relationship between Attack Shilling Attack and Hub Phenomenon]
(A. Nanopoulos, and M. Radovanovic, M. Ivanovic. How does high dimensionality impact collaborative filtering? 29 Proc. 3rd ACM Cons. 3rd ACM. P. Knees, D. Schnitzer, and A. Flexer, “Improving neighborhood-based collaborative filtering by reducing hubness”, In Proc. ICMR'14, pages 161-168. In item-based CF, kNN is because it is calculated at a high level, it reported that the hub phenomenon appears. Usually, since the number of users and the number of items are large, vectors used for calculation of similarity such as cosine similarity and Pearson correlation are high-dimensional, and a hub phenomenon occurs. And hub data that frequently appears in the kNN of other data affects the determination of many recommendations. However, hub data is data that does not have much meaning for many data. This is because hub data frequently occurs in kNN only because it is high-dimensional and similar to the data center, and is not useful for characterizing individual data. In fact, according to Nice et al., The performance of the recommendation system is exacerbated by the presence of hub data. Furthermore, since hub data is data that has a strong influence on the determination of recommended items by the system, if the hub data can be manipulated from outside the system, the system can be effectively attacked.
In fact, the hub phenomenon compromises the recommendation system against attacks. For example, a fake user entered into the system due to an average attack becomes hub data, which greatly affects the system. Therefore, suppressing the occurrence of the hub phenomenon is thought to lead to attack avoidance.

[Hub suppression by reducing bias toward data center]
If a small number of data with high similarity to the data center is a hub, the hub phenomenon is suppressed by converting the similarity scale to a similarity scale that makes all data objects equally similar to the data center. It is considered possible. Such conversion of the degree of similarity (scale) is obtained by calculating a commutation time kernel from a given degree of similarity, and more simply by centering the given degree of similarity.
Let N be the number of data and K be a similarity matrix of size N. The commutation time kernel K ^CT for K is given by equation (4).

K ^CT = L ⁺ (generalized inverse matrix of L) (4)

Here, L = D−K is called a graph Laplacian. D is a diagonal matrix such that D _ii = Σ _j K _ij .

を全要素が１であるＮ次元ベクトルとする。Ｋをセンタリングした類似度行列Ｋ^ＣＥＮＴは式（５）のように計算される。

Next, I is the identity matrix,

Is an N-dimensional vector with all elements being 1. A similarity matrix K ^CENT centered on K is calculated as shown in Equation (5).

FIG. 5 shows a configuration example of the item recommendation system 1 in the first embodiment.
In this embodiment, a user-based CF will be described as the item recommendation system 1. In other words, this is a system that recommends items by using, for example, the k-nearest neighbor method for similarity calculation and referring to past evaluation values of other users whose evaluation is similar to that of the target user. When the item is a product and the evaluation value is the preference level, the system recommends the product with reference to the past preference levels of other users who have similar preferences to the target user.
The configuration in FIG. 5 is applicable to both the first embodiment (user-based CF) and the second embodiment (item-based CF). For this reason, the description common to the user-based CF and the item-based CF will be described in the present embodiment, and in the second embodiment, the difference will be described to an extent.

The item recommendation system 1 includes a personal computer (PC) 10 for processing data and commands, a display unit 18 for displaying data / commands processed or input / output by each unit, and an input / output for inputting / outputting data and commands. The output unit 19 includes a storage unit 20 that stores data commands and the like processed or input / output by each unit.
A personal computer (PC) 10 includes a registration unit 11 for registering users and items, an evaluation unit 12 for entering an evaluation value indicating the degree of evaluation of a user's item in an evaluation matrix, and similarity between users using a similarity scale The similarity calculation unit 13 that calculates the degree and / or similarity between items, the neighborhood data extraction unit 14 that extracts, for example, k target data (users and / or items) from the higher similarity, the evaluation matrix R An evaluation value prediction unit 15 for predicting an evaluation value predicted to be written in the cell based on the evaluation value extracted by the neighborhood data extraction unit 14 when the evaluation value is not written in the cell; An item recommendation system that controls an item recommendation unit 16 that recommends an item for which a high evaluation value is predicted by the prediction unit 15 and the item recommendation system 1 is controlled. And a control unit 17 to function as a.

  Here, the registration unit 11 includes a user registration unit 111 that registers users and an item registration unit 112 that registers items. In the present embodiment, the similarity calculation unit 13 calculates the similarity between users and / or the similarity between items using a similarity scale that suppresses the hub as a similarity scale. Specifically, the similarity calculation unit 13 uses a similarity measure that suppresses the hub, and pays attention to the evaluation value of the user in each row of the evaluation matrix R to calculate the similarity between users. And the 2nd similarity calculation part 132 which calculates the similarity between items paying attention to the evaluation value of the item of each row | line | column. In addition, a similarity scale conversion unit 135 that performs conversion from similarity based on a general similarity scale to similarity based on a similarity scale that suppresses the hub is provided. The neighborhood data extraction unit 14 includes, for example, a first neighborhood data extraction unit 141 that extracts, for example, k users from the higher degree of similarity, and second neighborhood data that extracts, for example, k items from the higher degree of similarity. An extraction unit 142 is included. The evaluation value prediction unit 15 includes a first evaluation value prediction unit 151 that predicts an evaluation value of the target user based on the evaluation value extracted by the first neighborhood data extraction unit 141, and a second neighborhood data extraction unit. Based on the evaluation value extracted in 142, the second evaluation value prediction unit 152 that predicts the evaluation value of the target user is included. The first evaluation value prediction unit 151 and the second evaluation value prediction unit use, for example, an average value as an evaluation value to be entered when predicting an evaluation value to be entered in an unfilled cell related to the target user. it can. In addition, it is preferable to use a weighted average value because recommendation accuracy can be increased.

  In addition, the storage unit 20 stores the evaluation matrix R for storing the evaluation matrix R, the similarity measure for the user calculated by the similarity measure and the first similarity calculation unit 131, and / or the second similarity calculation. The similarity measure storage unit 22 stores similarity data related to the item calculated by the unit 132. The calculation data of the first similarity calculation unit 131 and the second similarity calculation unit 132 are calculated using a similarity measure that makes it difficult for hub data to appear. The similarity scale storage unit 22 may store a general similarity scale. In addition, the storage unit 20 includes a neighborhood data storage unit 23 that stores data (user and / or item) extracted by the neighborhood data extraction unit 14, and a recommended item storage unit that stores items estimated by the evaluation value estimation unit 15. 24. The similarity scale storage unit 22 stores a general similarity scale and / or a similarity scale that suppresses the appearance of a hub in addition to the similarity data. When the similarity measure that suppresses the appearance of the hub is not stored but the general similarity measure is stored, the similarity measure conversion unit 135 suppresses the appearance of the hub with the general similarity measure. The similarity scale is converted into a similarity scale, and the similarity scale storage unit 22 stores the obtained similarity scale that suppresses the appearance of the hub. For similarity conversion, for example, a general similarity scale formula is converted to a similarity scale formula that suppresses the appearance of hubs, and the similarity is calculated. A method of converting to a similarity based on a similarity measure that suppresses the appearance of a hub by a matrix is used. In this case, a general similarity measure before conversion may be left without being deleted. If a general similarity measure and a similarity measure that suppresses the appearance of a hub are stored together, the similarity calculation result and the evaluation value prediction result related to item recommendation can be used in combination or compared. . The neighborhood data storage unit 23 stores the first neighborhood data storage unit 231 that stores the users extracted by the first neighborhood data extraction unit 141 and the items extracted by the second neighborhood data extraction unit 142. A second neighborhood data storage unit 232 is provided. The recommended item storage unit 24 stores contents to be displayed when recommending items to the user. For example, in addition to the item name, a description about the item, a description for using the item, an image of the item, and the like are stored. These contents are displayed on the display unit 18 at the time of recommendation.

  In the present embodiment, k people are extracted from those having higher similarity to the user, and the user evaluation value is predicted as an average of the k evaluation values, so that the similarity calculation unit 13 performs the first similarity calculation. Unit 131, first neighborhood data extraction unit 141 as neighborhood data extraction unit 14, first assessment value prediction unit 151 as assessment value prediction unit 15, and neighborhood user storage unit 231 as neighborhood data storage unit 23 may be used. The second similarity calculation unit 132, the second neighborhood data extraction unit 142, the second evaluation value prediction unit 152, and the neighborhood item storage unit 232 may be omitted. These are used in Example 2.

FIG. 6 shows a processing flow example of the item recommendation method in the first embodiment. FIG. 6A is a diagram illustrating an example of a processing flow in the first embodiment, and FIG. 6B is a diagram illustrating an example of a processing flow in the second embodiment described later.
First, an item i (S101: item registration process) and a user u are registered in the evaluation matrix R (S102: user registration process). Either the item i registration or the user u registration may be performed first or in parallel. Next, an evaluation value R (u, i) indicating the degree of evaluation related to the item i of the user u is registered in the evaluation matrix R (S103: evaluation value registration step). In this embodiment, an item is a product, an evaluation value is a preference level, and an evaluation matrix R is a preference matrix. Evaluation is performed by, for example, a five-step evaluation (an integer of 1 to 5). The user himself / herself may be registered, or may be registered by referring to the behavior related to the item of the past user on the system side. It is not always necessary to fill in the entire matrix R, and there may be blank cells, and most of them are usually blank. The evaluation matrix R is stored in the evaluation matrix storage unit 21 (S104: evaluation matrix storage step).

  Next, an item to be recommended to each user is determined based on the evaluation matrix R. First, similarity calculation is performed to obtain another user similar to the user himself / herself. When performing the similarity calculation, it is assumed that a similarity measure storage unit 22 stores a general similarity measure or a similarity measure that suppresses the appearance of a hub in advance as a similarity measure. First, it is determined whether or not there is a similarity measure that suppresses the appearance of the hub in the similarity measure storage unit 22 (S105: presence / absence determination step of a hub suppression similarity measure). When the similarity measure storage unit 22 does not store the similarity measure that suppresses the appearance of the hub and stores a general similarity measure (in the case of No in S105), the general similarity is stored. The similarity based on the degree scale is converted into the similarity based on the similarity scale that suppresses the appearance of the hub (S106: similarity scale conversion step). As a similarity measure for suppressing the appearance of the hub, for example, a similarity measure in which all data objects are equally similar to the data center, that is, a similarity measure without Spatial Centrality can be used. Specifically, for example, centering or conversion to a commuted time kernel is performed. The similarity measure for suppressing the converted hub is stored in the similarity measure storage unit 22. Next, the similarity between users is calculated using the similarity measure that suppresses the appearance of the hub (S107: first similarity calculation step). Note that when the similarity conversion is performed in a matrix, the similarity scale conversion step (S106) and the first similarity calculation step (S107) are performed in a lump. In this case, the similarity measure does not necessarily remain as an expression, but remains in the similarity data based on the similarity measure that suppresses the hub in the calculation result. When the similarity measure that suppresses the appearance of the hub is already stored in the similarity measure storage unit 22 (Yes in S105), the similarity measure conversion step (S106) is omitted, and the appearance of the hub is determined. The similarity between users is calculated using the similarity measure to suppress (S107: 1st similarity calculation process). The result calculated using the similarity measure that suppresses the appearance of the hub is stored in the similarity measure storage unit 22.

  Next, for example, k users are extracted from the one with the higher similarity stored in the similarity scale storage unit 22 (S108: first neighborhood data extraction step). Then, based on the extracted average evaluation values of k users and the like, an evaluation value that is blank for the target user is predicted (S109: first evaluation value prediction step). In the first evaluation value predicting step (S109), when an evaluation value to be entered in an unfilled cell related to the target user is predicted, it can be predicted using, for example, an average value. It is preferable to use a weighted average value. Finally, an item with a high predicted value is recommended (S110: item recommendation step). For example, when a user visits an Internet site that provides an item, items are presented in descending order of the predicted value for the user. Moreover, you may deliver to the said user with an email.

[Experiment]
Experiments were conducted to ensure that suppressing the appearance of hub data makes user-based CF robust against average attacks. In the experiment, a movie lens 1M data set (m1-1m) used as benchmark data for a recommended task was used. This data set consists of 1,000,209 evaluation values (5-level evaluation of integers 1 to 5) for 6,040 users and 3,706 items. Every user has rated at least 20 items. The evaluation value was predicted using the formula (1) by using commonly used cosine similarity (Cos) and Pearson correlation and shrunken Pearson correlation (Pearson) as the similarity measure between users serving as baselines. . Since it is known that the shrunken Pearson correlation is more accurate than the Pearson correlation, this time, the shrunken Pearson correlation is used as the Pearson correlation. From now on, the shrunken Pearson correlation will be referred to as Pearson correlation (Pearson). The Pearson correlation parameter (Pearson) parameter β was set to β = 100 in accordance with past research reports. As a method for suppressing the appearance of hub data, a method of converting the similarity as a baseline into a commuted time kernel (CT) according to equation (4), or a method of centering conversion according to equation (5) (C _ENT ) Was used. The main purpose of this experiment is to compare the robustness (resistance) of the system to attacks before and after conversion.

[Prediction accuracy when there is no attack]
Before examining the robustness against the attack, it was verified whether or not the CT conversion and the _CENT conversion do not deteriorate the prediction accuracy of the evaluation value when there is no attack. In order to simulate the recommendation work, 1,000,209 evaluation values in the data set are converted into 939,809 training data (data used for prediction of test data) and 60,400 test data ( Divided into data subject to prediction). The average absolute error (MAE) generally used for the evaluation of the CF algorithm was used to compare the similarity measures before and after the conversion.

MAE = 1 / | T | Σ _{(u, i) ∈T} | Pred (u, i) −R (u, i) |

As calculated. Here, T is a set of user-item pairs given as test data (| T | = 60400).

In FIG. 7, the nearest neighbor parameter k is varied between 10 and 100, and the cosine similarity (Cos) and the Pearson correlation (Pearson), which are the similarity measures as a baseline, are converted into a commuted time kernel transform (CT) or a centering transform The average absolute error (MAE) in the case of (C _ENT ) is compared and shown. From FIG. 7, C _ENT decreases MAE in most cases (increases prediction accuracy), and CT decreases MAE in the case of Pearson correlation. From this, it can be seen that the C _ENT transform and the CT transform for Pearson correlation are improved rather than worsening the prediction accuracy when there is no attack.
In the following, when evaluating the robustness against the average attack, k = 50 is set to achieve the best MAE in the above experiment.

[Robustness against attack]
Twenty-one movie items were selected as target items for the average attack. These items were found in Lam et al. (SK Lam and J. Riedl. Shilling. Recommenders Systems for Fun and Profit. In Proc. WWW '04, pages 393-402, 2004. Year) was selected to be as close as possible to the items used in the experiment (in terms of the number of evaluation users and average evaluation values). As an average attack, 100 fake users were introduced, and the fake user was given a high evaluation (that is, 5) for the target item, and an average evaluation with noise added to the remaining other items. That is, for each of the remaining items, a random number according to a normal distribution (μ; σ), where μ = average evaluation value and σ = 1.0, is generated and converted to the nearest integer value 1-5. Granted. A value called a prediction shift, that is, a value that is a difference between prediction evaluation values before and after the attack, was used to compare the similarity measures before and after the conversion. More precisely, except for training data, a prediction shift was calculated for each pair of honest users (all users except fake users) and target items, and the average value was used for comparison.

表１はアベレジアタックにより生じた予測シフトと、Ｎ_ｋ分布の歪度を示す。大きいＮ_ｋを持つデータ、すなわちハブデータが存在するほど、Ｎ_ｋ分布の歪度は大きい値となる。Ｎ_ｋ分布の歪度は、Ｓ_Ｎｋ＝Ｅ〔（Ｎ_ｋ−μ_Ｎｋ）^３／σ_Ｎｋ ^３〕（Ｅ〔 〕は期待オペレータ、μ_Ｎｋとσ_ＮｋはそれぞれＮｋ分布の平均と標準分散である）で表される。Ｎ_ｋ分布の歪度、予測シフトのどちらも、Ｃ_ＥＮＴ又はＣＴ変換後に減少した。このことは、変換された類似度尺度の使用は、ハブデータの出現を抑制し、その結果、推薦システムを攻撃に対してロバストにすることを示している。

Table 1 shows the prediction shift caused by the average attack and the skewness of the _Nk distribution. The more the data having a large N _k , that is, the hub data, the greater the skewness of the N _k distribution. Skewness of N _{k distribution, S Nk} = E ^{_{[(N k -μ Nk) 3 /}} σ Nk 3 ] (E [] is the expectation operator, the mu _Nk and sigma _Nk is the average and the standard variance of each Nk distribution ). Both skewness and prediction shift of _Nk distribution decreased after _CENT or CT conversion. This indicates that the use of the transformed similarity measure suppresses the appearance of hub data, thus making the recommendation system robust against attacks.

FIGS. 8 and 9 show that when a general Pearson correlation is used as a similarity measure between users, the input false user becomes a hub, and the conversion of the similarity measure makes it difficult for the false user to become a hub. (No. 1 and No. 2) 8, in order to see the relation between the similarity to the N ₅₀ value and the data center about the user is a scatter diagram plotting the respective user. The horizontal axis represents N ₅₀ , and the vertical axis represents the similarity to the data center. From FIG. 8, it can be seen that the input fake user has a high degree of similarity with the data center and is therefore a hub. Figure 9 (a), (b) , (c) _is a histogram of the user according to the _{N 50} value. Fig. 9 (a) used Pearson correlation (original), Fig. 9 (b) used Pearson correlation after centering transformation, and Fig. 9 (c) used Pearson correlation after commutation time kernel transformation as similarity measures. Is a histogram of the case.
Next, we analyze why _CENT or CT provides robustness in the hub phenomenon.

In the scatter diagram of FIG. 8, it can be seen that there are hub users having N ₅₀ values of up to 961 and that there is a strong correlation between the N ₅₀ values and the similarity to the data center. Authentic users are indicated by 〇, and input fake users are indicated by ×, but since the input fake users are imitated by average sincere users, they have a high degree of similarity with the data center on users. . Therefore, as can be seen from the N ₅₀ distribution in FIG. 9 (a), compared to many sincere users, the input fake user is a hub user (influencer) with a large N ₅₀ value (min 465, max 961). ) On the other hand, the C _ENT or CT conversion suppresses the occurrence of the hub phenomenon. As a result, as shown in FIG. 9B, in C _ENT , the N ₅₀ value of the fake user decreased to a minimum of 101 and a maximum of 156. Further, as shown in FIG. 9 (c), when CT was used, it decreased to a minimum of 0 and a maximum of 4. This is because, compared to using the original Pearson correlation as a user-to-user similarity measure, by using the _CENT or CT-transformed similarity measure, the input fake user can It clearly shows that kNN does not appear so often. In other words, fake users no longer affect the decision of recommended items (no longer fake users are influencers).

  From the above, it can be seen that by converting the similarity measure so as to suppress the occurrence of the hub phenomenon, it is possible to obtain a recommendation system that is robust against attacks and has the same or better prediction accuracy as the original similarity measure. It was.

[Conclusion]
We proposed a method to make collaborative filtering (CF) robust by suppressing the hub phenomenon against attacks aimed at changing recommended items by introducing fake users from outside.
Our approach rests on the basis that the hub phenomenon is one of the main factors used by shilling attacks. We applied two transforms that suppress the appearance of hub data (centering and commutation time kernel transformations) to commonly used similarity measures (cosine similarity and Pearson correlation). Using a movie lens data set, we have shown that these transformations make the recommendation system robust to shilling attacks without degrading recommendation accuracy.

  As described above, according to the present embodiment, it is possible to provide an item recommendation system and an item recommendation method that are less affected by an attack as a result even if an attack in which a fake user is introduced.

  In Example 1 (user base CF), the evaluation value is predicted based on the similarity between users, but in this embodiment, an example in which the evaluation value is predicted based on the similarity between items will be described. That is, in Example 2 (item base CF), k items are extracted from the one having a higher degree of similarity to the target item, and the average of the evaluation values previously given to the k items by the target user is as follows: An example of predicting an evaluation value for a target item of a target user will be described.

  Compared to the first embodiment, the similarity calculation unit 13 replaces the first similarity calculation unit 131 that calculates the similarity between users, and a second similarity calculation unit that calculates the similarity between items. 132 is used. As a similarity measure between items, for example, a measure in which all items are equally similar to the data center related to the item is used. The neighborhood data extraction unit 14 uses the similarity between items calculated by the second similarity calculation unit 132 instead of the first neighborhood data extraction unit 141, and uses the one with the higher similarity to the target item. The second neighborhood data extraction unit 142 is used to extract k items from. In the evaluation value prediction unit 15, instead of the first evaluation value prediction unit 151, the evaluation values for the k items extracted by the second neighborhood data extraction unit 142 are used for the target item of the target user. A second evaluation value predicting unit 152 that predicts an evaluation value to be entered in an unfilled cell is used. Other configurations are the same as those of the first embodiment.

Compared to the first embodiment, in the similarity calculation step, instead of the first similarity calculation step (S107) for calculating the similarity between users, the second similarity calculation for calculating the similarity between items is performed. Step (S207) is performed. As a similarity measure between items, for example, a measure in which all items are equally similar to the data center related to the item is used. In the neighborhood data extraction step, instead of the first neighborhood data extraction step (S108), the similarity with the target item is calculated using the similarity between items calculated in the second similarity calculation step (S207). A second neighborhood data extraction step (S208) for extracting k items from the higher one is performed. In the evaluation value prediction step, instead of the first evaluation value prediction step (S109), the evaluation values of the k items extracted in the second neighborhood data extraction step (S207) are used, and the target user's A second evaluation value prediction step (S209) for predicting an evaluation value to be entered in an unfilled cell for the target item is performed. Other processing flows are the same as those in the first embodiment.
In this way, k items are extracted from the one having a higher degree of similarity to the target item, and the evaluation value of the user is predicted as the (weighted) average of the evaluation values of the k items.

  According to the present embodiment, as in the first embodiment, an item recommendation system and an item recommendation method that are less affected by an attack as a result even when subjected to an attack that introduces a fake user can be provided.

In the above embodiment, an attack in which a fake user is input has been described. However, even when there is no attack, hub data originally tends to exist in a high-dimensional or large-scale data set. Even in this case, if the similarity measure is converted so as to suppress the appearance of the hub, the appearance of the data serving as the influencer can be suppressed. This makes it possible to recommend the item that is originally recommended (see the research report by Nice et al., Described in [Relationship between Attack Shilling Attack and Hub Phenomenon] above).
In the present embodiment, it is possible to provide an item recommendation system and an item recommendation method that are not subject to recommendation bias by an influencer even when there is no attack due to fake user input.

In the above embodiment, an example of recommending an item that suits the taste for each user has been described, but it is assumed that an advertisement is made to a large number of people and the general public. Suppose that it is better to advertise for the average user. In this case, instead of calculating the similarity between the users, a temporary user having an average evaluation value is generated, and the similarity with other users is calculated for the temporary user, and k users If a user is extracted and an evaluation value is predicted, it is preferable to recommend an item for the masses.
Also in the present embodiment, an item recommendation system and an item recommendation method can be provided that are less affected by an attack as a result even when subjected to an attack that introduces a fake user.

  The present invention can also be realized as a program for causing a computer to execute the item recommendation method described in the flowcharts of the above embodiments. The program may be stored and used in the storage unit of the item recommendation system, stored in an external storage device, or downloaded from the Internet and used. Moreover, it is realizable also as a recording medium which recorded the said program.

  Although the embodiment of the present invention has been described above, the embodiment is not limited to the above example, and it is obvious that various modifications can be made without departing from the spirit of the present invention.

  For example, for items and evaluation values, the present invention can be applied if items and evaluation values not listed in the present specification can be evaluated quantitatively. The similarity scale may be adjustable using parameters. For item recommendation, images and explanations can be added to the item name. In addition, the recommendation may be accessible by mail to each user in addition to when each user accesses the web page, and can also be distributed by mail. The dimensions of the evaluation matrix and the parameter k of the k-nearest neighbor method can be appropriately determined according to the purpose and situation.

  The present invention is used in a recommendation system based on collaborative filtering represented by a user base or an item base.

1 Item recommendation system 10 Personal computer (PC)
DESCRIPTION OF SYMBOLS 11 Registration part 12 Evaluation part 13 Similarity calculation part 14 Neighborhood data extraction part 15 Evaluation value prediction part 16 Item recommendation part 17 Control part 18 Display part 19 Input / output part 20 Storage part 21 Evaluation matrix storage part 22 Similarity which suppressed hub data Degree scale storage unit 23 Neighborhood data storage unit 24 Recommended item storage unit 111 User registration unit 112 Item registration unit 131 First similarity calculation unit 132 Second similarity calculation unit 135 Similarity scale conversion unit 141 First neighborhood data Storage unit 142 Second neighborhood data storage unit 151 First evaluation value prediction unit 152 Second evaluation value prediction unit 231 Neighborhood user storage unit 232 Neighborhood item storage unit i Item R Evaluation matrix R (u, i) Evaluation value u User

Claims (10)

An evaluation matrix storage unit for storing an evaluation matrix for entering an evaluation value related to the user's item;
A first similarity calculator that calculates the similarity between users using a similarity measure that suppresses the appearance of a hub;
A first neighborhood data extraction unit that extracts k users from the one having a higher similarity to the target user by using the similarity calculated by the first similarity calculation unit;
A first evaluation value for predicting an evaluation value to be entered in an unfilled cell relating to the target user using an evaluation value relating to an item of k users extracted by the first neighborhood data extraction unit With a predictor;
An item recommendation unit that extracts a recommended item for extracting an item to be recommended to the target user from items having a high evaluation value predicted by the first evaluation value prediction unit and recommends the item to the target user;
Item recommendation system. An evaluation matrix storage unit for storing an evaluation matrix for entering an evaluation value related to the user's item;
A second similarity calculator that calculates the similarity between items using a similarity measure that suppresses the appearance of a hub;
A second neighborhood data extraction unit that extracts k items from a higher similarity with the target item using the similarity calculated by the second similarity calculation unit;
A second evaluation value for predicting an evaluation value to be entered in an unfilled cell relating to the target user, using the user's evaluation value relating to the k items extracted by the second neighborhood data extraction unit With a predictor;
An item recommendation unit that extracts an item to be recommended to the target user from items having a high evaluation value predicted by the second evaluation value prediction unit and recommends the item to the target user;
Item recommendation system. A similarity measure storage unit for storing a similarity measure for suppressing the appearance of the hub;
The item recommendation system according to claim 1 or 2. A similarity scale conversion unit that converts similarity based on a general similarity scale into similarity based on a similarity scale that suppresses the appearance of the hub;
The item recommendation system according to any one of claims 1 to 3. A weighted average value is used as the evaluation value to be entered when predicting an evaluation value to be entered in an unfilled cell of the target user;
The item recommendation system according to any one of claims 1 to 4. An evaluation matrix storage step of storing an evaluation matrix for entering an evaluation value relating to the user's item;
A first similarity calculation step of calculating a similarity between users using a similarity measure that suppresses the appearance of a hub;
A first neighborhood data extraction step of extracting k users from the higher similarity with the target user using the similarity calculated in the first similarity calculation step;
A first evaluation value for predicting an evaluation value to be written in an unfilled cell related to the target user using an evaluation value related to an item of k users extracted in the first neighborhood data extraction step A prediction process;
An item recommendation step of extracting an item to be recommended to the target user from items having a high evaluation value predicted in the first evaluation value prediction step and recommending the item to the target user;
Item recommendation method. An evaluation matrix storage step of storing an evaluation matrix for entering an evaluation value relating to the user's item;
A second similarity calculation step of calculating a similarity between items using a similarity measure that suppresses the appearance of a hub;
A second neighborhood data extraction step of extracting k items from a higher similarity with the target item using the similarity calculated in the second similarity calculation step;
A second evaluation value for predicting an evaluation value to be entered in an unfilled cell relating to the target user, using the user's evaluation value relating to the k items extracted in the second neighborhood data extraction step A prediction process;
An item recommendation step of extracting an item to be recommended to the target user from items having a high evaluation value predicted in the second evaluation value prediction step and recommending the item to the target user;
Item recommendation method. A similarity scale conversion step of converting similarity based on a general similarity scale into similarity based on a similarity scale that suppresses the appearance of the hub;
The item recommendation method according to claim 6 or 7.   The program for making a computer perform the item recommendation method of any one of Claim 6 thru | or 8.   A computer-readable recording medium on which the program according to claim 9 is recorded.

Images