JP2012098975A - Item recommendation method and device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide an information recommendation method for recommending an unexpected item to a user using a difference between action and taste of the user, and a system for the same.SOLUTION: An item recommendation device includes: an item feature space calculation unit 112 for using an item attribute 102 to structure an item feature space 105; an action model generation unit 113 for using an item selection history 103 of a user to obtain an action model 106 of the user; a taste model generation unit 114 for using item taste 104 of the user to obtain a taste model 107 of the user; an action prediction value calculation unit 116, a taste prediction value calculation unit 117, and an unexpectedness coefficient and taste prediction value combined recommendation index calculation unit 118 for calculating a recommendation index that becomes smaller for an item having a larger taste prediction value obtained and a smaller action prediction value obtained; and a recommendation item determination unit 115 for generating a list with an unselected item having a small recommendation index set as a recommendation item.

Description

  The present invention relates to a method for recommending an item such as moving image content, and more particularly to a technique for recommending an unexpected item while following a user's preference. Here, “item recommendation” is a function for presenting information items to the user according to the user's preference, and “unexpected item recommendation” means that the user can easily follow the user's preference. A function that presents information that is difficult to predict.

  As a system for recommending information to be distributed when distributing information such as content, refer to Non-Patent Document 1 in which a “collaborative filtering” method is conventionally known. This collaborative filtering is a method for estimating a user profile and what type a user belongs to from the similarity of past user histories and determining the next recommended item.

  Non-Patent Document 2 describes a method in which items are clustered, and other items in the cluster to which the user's favorite item belongs are used as an unexpected recommendation result.

  Further, in Patent Document 1, the keyword vector of the customer profile and the keyword vector of the unpurchased book are obtained for each category, the similarity of the vectors is calculated between different categories, and the similar item obtained is determined as an unexpected recommendation result. The method to do is described.

  Patent Document 2 describes a method of analyzing an item selection distribution and recommending non-selected items that are close to the center of the distribution and away from the distribution shape.

JP 2001-265808 “Information Retrieval System and Information Retrieval Method” Japanese Patent Application No. 2009-249672 “Information Recommendation Method and Device” JP 2004-194107 “Information processing apparatus and information processing method, information processing system, recording medium, and program”

G. Linden et al .: “Amazon.com Recommendations, Item-to-Item Collaborative Filtering”, IEEE Internet Computing, 2003. Yuka Kato et al .: “Recommendation System Considering Balance between Accuracy and Unexpectedness for Online Shopping”, IPSJ Journal, Vol.46, No. SIG 13 TOD 27, Sep. 2005.

  The collaborative filtering method represented by Non-Patent Document 1 is mainly a method of determining a recommended item with reference to an unrecommended part in the entire history of another user whose history is partially similar. It is difficult to recommend items that are of interest to the user, i.e., items that are of interest to the user but cannot be easily assumed.

  In addition, in the recommendation method described in Non-Patent Document 2, since a cluster is formed using items, the cluster is fixed, and a recommendation result is always output from the same cluster unless the user's preference changes, and Since items in the same cluster as the user's favorite items are recommended, there is a problem that they are similar and not necessarily surprising.

  In addition, the recommendation method described in Patent Document 1 has a problem that it is not possible to recommend an unexpected item that is in the same category. Further, the present invention cannot be applied to items that are not categorized.

  In addition, in the recommendation method described in Patent Document 2, the center of the action and the center of the preference are the same in the selection of the recommended item, and the item is recommended by the Euclidean distance from the center of the action. There was a problem that could not be said to fit.

  An object of the present invention is to provide an information recommendation method capable of making a recommendation that is unexpected for a user, a system thereof, and the like.

  In order to solve the above-described problems, in the present invention, a server including a processing unit and a storage unit is used to store an item attribute, a history of selecting an item by the user, and a preference for the item selected by the user, An information recommendation method and an information recommendation system for making a recommendation based on a history and a preference, wherein a processing unit constructs an item feature space from item attributes and analyzes a user's behavior distribution in the space from the history And analyzing the user's preference distribution in the space from the preference, and providing an information recommendation method and an information recommendation system using an unselected item having a high preference prediction value and a low behavior prediction value as a recommended item .

  The processing unit creates an action model from the item selection history, creates a preference model from the preference for the item, calculates an action prediction value and a preference prediction value for an unrecommended item, and the value is larger when the action prediction value is larger. The present invention provides an information recommendation method and an information recommendation system in which an unexpected coefficient having a larger preference prediction value is smaller and a smaller value is obtained, and an unrecommended item having a smaller coefficient and a larger preference prediction value is used as a recommended item.

  ADVANTAGE OF THE INVENTION According to this invention, the item close | similar to preference and distant from a user's action can be recommended on item feature space.

  Since items that are close in the feature space are similar, items that are close to the taste are considered to match the user's preference, and it is difficult to predict if they are away from the behavior, which is surprising. High nature. Therefore, it is possible to make a recommendation that is surprising to the user.

  Further, according to the present invention, item category data is not necessary, and it is possible to make a recommendation that matches the user's preference and is surprising to the user by using the user's item selection history and preference.

  According to the present invention, it is possible to make a recommendation that matches the user's preference and is surprising to the user in the item feature space.

It is a figure which shows an example of the system configuration | structure of an information recommendation apparatus. It is a figure of an example of the flow of an information recommendation method. It is a figure which shows an example of the data structure of an item attribute. It is an example of the type | formula which calculates | requires the axis | shaft of item feature space. It is a figure which shows an example of the data structure of an item selection log | history and item preference. It is an example of the type | formula which calculates | requires action space feature matrix. It is an example of the type | formula which calculates | requires a preference space feature matrix. It is an example of the type | formula which calculates | requires a recommendation parameter | index. It is a figure of an example of the flow of the recommendation item determination method. It is a figure which shows another example of the recommendation item list production | generation method. It is an example of the type | formula which calculates | requires the range of a recommendation item. It is an image figure which shows the range of a recommendation item. It is a figure which shows an example of the system configuration | structure of a program recommendation apparatus. It is a figure of an example of the flow of a program recommendation method. It is a figure of an example of program information. It is an example of the type | formula which calculates | requires the axis | shaft of program feature space. It is a figure which shows an example of the data structure of a program preference. It is a figure of an example of the flow of the recommendation program determination method. It is a figure of an example of the method of acquiring a program preference.

  Hereinafter, embodiments for carrying out the present invention will be described with reference to the drawings.

  First, terms used in this specification are defined.

  In this specification, the “item” is a recommendation target, for example, an EC site for a recommended product at the time of purchase, and a program on a television that presents a recommended program at the time of viewing. The “item attribute 102” is information related to the item. For example, in the case of a food item, the item attribute 102 indicates the calorie, sugar content, and the like of the food.

  In this specification, “item selection history 103” is data in which the type and number of items selected by the user are recorded, and “selected item” is a recommendation result list output to the input / output unit 121. The item actually selected by the user, the EC site where the item is a product, and the like are products for which detailed information has been viewed.

  Further, in this specification, “item preference 104” is a preference for a user's item input from the input / output unit 121. If the item is a product, the satisfaction level after purchasing the product, and the item is food. In some cases, it is the degree of likes and dislikes for the food.

  A system configuration example of the present invention will be described with reference to FIG. This system includes a processing unit 111, a storage unit 101, and an input / output unit 121.

  The processing unit 111 includes a feature space calculation unit 112, a behavior model creation unit 113, a preference model creation unit 114, a recommended item determination unit 115, an item behavior prediction calculation unit 116, an item preference prediction value calculation unit 117, an unexpectedness coefficient / preference prediction. And a value combination recommendation index calculation unit 118.

  The storage unit 101 includes a database of item / attribute 102, item selection history 103, item preference 104, item feature space 105, user behavior model 106, and user preference model 107.

  The input / output unit 121 is a user interface using an input device such as a touch panel, a keyboard, a remote controller, and an image recognition, and an output device that presents information such as a display, a light, a printer, and sound.

  FIG. 2 is a diagram illustrating an example of the flow of the information recommendation method.

  First, the feature space calculation unit 112 uses the item attribute 102 to construct an item feature space in which items are spatially arranged with the item attribute as an axis, and stores the item feature space in the storage unit 101 (201). The item feature space 105 is constructed when the item set is determined. Here, the “time when the item set is determined” refers to the time when the assortment of products is determined in the case of an EC site, for example, and the time when the contents of a broadcast program are determined in the case of a television. . However, each time an item is input, it may be newly created and updated.

  Next, triggered by the input of a recommendation request from the user via the input / output unit 121, or automatically, the behavior model creation unit 113 uses the item selection history 103 to determine the user's behavior model 106. 202 to create. Further, the preference model creation unit 114 creates 203 the user's preference model 107 using the item preference 104.

  Next, a non-selected item in the item is specified 204 from the item selection history 103. For the non-selected item, a value indicating the degree of possibility that the user will take an action of selecting an item based on the behavior model is obtained 115 and set as an item behavior prediction value 205.

  Similarly, a predicted value of the degree of likes and dislikes for the user's item is obtained 116 based on the preference model 116 and is set as an item preference predicted value 206.

  Next, based on the obtained behavior prediction value and preference prediction value, the unexpectedness coefficient / preference prediction value combined recommendation index calculation unit 118 calculates a recommendation index, and based on the index, the recommended item determination unit 115 determines the recommended item. A list is generated 207.

  Next, the item / attribute 102 will be described with reference to FIG. 3 showing the data structure of the item attribute.

  The item / attribute is represented by a matrix having items 311 to 314 as rows and attributes 301 to 304 as columns. For example, the value of attribute j of item i is value ij321. For example, in the case of a food item, if the item i is “carrot potage” and the attribute j is “calorie”, the value ij is “150 kilocalories”, for example.

  Next, the item feature space construction process 201 performed by the feature space calculation unit 112 will be described with reference to FIG. 4 showing an expression for obtaining the axis of the item feature space.

  Each item is represented by a point feature point in the space, and the feature space is constructed so that the distance between feature points corresponding to similar items is reduced. The feature space may be a space around each attribute of the item as an axis from a simple point of view. However, although the distance between feature points of items with similar attribute values is small, the size and variation of values differ for each attribute, and there may be correlation between attributes, so it may not always be the best feature space. is there. For example, when the item is food, the attribute “calorie” has a value of several tens to several hundreds, while the attribute “salt” has a value smaller than about 2. Even if the Euclidean distance between the feature points is calculated as it is, the difference in salinity is only a slight difference, so the similarity between items is almost determined by calories. Further, when there is a correlation between the attribute “calorie” and the attribute “sugar content”, the magnitude of both values has a greater influence on the similarity than the other attributes.

  Therefore, in this embodiment, principal component analysis using a correlation matrix is performed on the item attribute 102. Principal component analysis is independent of items in the linear combination 401 of attribute values 403 shown in FIG. 5 for data in which a plurality of items such as the item attribute 102 are represented by a plurality of attribute values. To obtain a set of coefficients 402 such that the variance becomes large. The analysis is known to result in the eigenvalue problem of the correlation matrix of attribute values, written by Shoichi Ueda, “Principal Component Analysis, Statistics Based on Course Information 8”, Asakura Shoten, March 28, 2003. The coefficient 402 and the eigenvalue 412 obtained by the obtained eigenvector component represent the linear combination value, that is, the variance of the principal component score 401.

  Here, the linear combination corresponding to a small eigenvalue, that is, the principal component, is not useful as a value for measuring the similarity / difference of items because the value difference between items is small. Therefore, only the principal component corresponding to the eigenvector having an eigenvalue of 1 or more is selected by a criterion generally called the Kaiser criterion, and the principal component score is divided by the square root of the eigenvalue so that the items are uniformly distributed in the feature space. The value 411 is a coordinate value on the feature space. In this case, the number of dimensions of the feature space is equal to one or more eigenvalues.

  Next, the item selection history 103 will be described with reference to FIG. 5 showing the data structure of the item selection history and item preference.

  The item selection history is represented by a matrix having the users 511 to 514 as rows and the items 501 to 504 as columns. For example, when the user k selects or purchases the item i three times, the value ki is “3”. When the user k has never selected an item, the value ki is “0”.

  The item preference 104 is similarly represented by the matrix in FIG. For example, if the degree that the user k likes the item i is “4” in the five-step evaluation, the value ki is “4”. When there is no preference information for the item i of the user k, the value ki is “0”.

  Next, a user behavior model creation process 202 performed by the behavior model creation unit 113 will be described with reference to FIG. 6 showing an expression for obtaining a behavior space feature matrix. The selected item distribution for each user is represented by an action space feature matrix 601. The matrix is an inverse matrix of the variance-covariance matrix 611 of the attribute value of the item selected by the user, and the variance-covariance matrix is calculated by the formula shown in FIG. Here, 612 is the item selection history 103, which is an element in FIG. 5 and is a value corresponding to the user k and the item i, 621 is an average vector of the selected items of the user k and represents the action center, and 622 is the item feature space The coordinate value 411 on the p-th axis of the item i obtained in the construction process 201 is a position vector 613 on the feature space of the item i having the coordinate value as a component.

  Next, the user preference model creation process 203 performed by the preference model creation unit 114 will be described with reference to FIG. 7 showing an expression for obtaining a preference space feature matrix. The item preference distribution for each user is represented by an action space feature matrix 701. The matrix is an inverse matrix of the variance-covariance matrix 711 of the attribute values of items whose user preferences are known, and the variance-covariance matrix is calculated by the formula shown in FIG. Here, 712 is the item selection history 104, the element of FIG. 5, which is a value corresponding to the user k and the item i, 721 is an average vector of the attributes of the preference item of the user k and represents the preference center, and 722 item features The coordinate value 411 of the p-th axis of the item i obtained in the space construction process 201 is a position vector on the feature space of the item i having the coordinate value as a component.

  Next, the calculation method in the item behavior prediction value calculation unit 116, the item preference prediction value calculation unit 117, and the unexpectedness coefficient / preference distance combined recommendation index calculation unit 118 will be described with reference to FIG. . The recommendation index 801 is represented by the product of the unexpectedness coefficient 802 and the inverse of the preference prediction value 805. The unexpectedness coefficient 802 is expressed as a ratio between the behavior prediction value 803 and the preference prediction value 805. The predicted behavior value 803 is represented by the reciprocal of the behavior distance 804, and the behavior distance 804 corresponds to the Mahalanobis distance for the variance-covariance matrix of the selected item attribute value. The preference prediction value 805 is indicated by the reciprocal of the preference distance 806, and the preference distance 806 corresponds to the Mahalanobis distance for the variance-covariance matrix of preference item attribute values. Since an item with a small preference distance 806 is close to the center of gravity of the user's favorite items in the past, it is an item similar to the favorite item and can be expected to be an item that the user likes. In addition, since an item having a large action distance 804 is far from the distribution of a plurality of items selected by the user in the past, it can be expected that the item is difficult to predict unlike the action distribution of the user. Therefore, an item having a small unexpectedness factor 802 that is a ratio of these distances can be expected to be an item that matches the user's preference and difficult for the user to predict.

  Next, the recommended item determination method in the recommended item determination unit 115 will be described with reference to FIG. 9 showing the flow of the recommended item determination method. First, the recommended index is calculated 902 by the unexpectedness coefficient / preference distance combined recommendation index calculation unit 118 for the item i having the value ki = 0 in FIG. Next, the items are sorted 903 in ascending order of the obtained recommendation index. The recommended items are set in the descending order of recommendation index 904. At this time, the system administrator or the user can specify the recommended item number n. In that case, the top n items are determined in ascending order of recommendation index.

  Next, another embodiment of the item list generation 207 will be described with reference to FIGS. 1, 8, and 10 to 12. FIG. 10 is a diagram illustrating another example of the recommended item list generation method, FIG. 11 is an example of a formula for obtaining the range of recommended items, and FIG. 12 is an image diagram illustrating the range of recommended items.

  First, the preference predicted value lower limit 1102 and the behavior predicted value upper limit 1112 are input 1001 through the input / output unit 121. Next, in the unexpectedness coefficient / preference prediction value combined recommendation index calculation unit 118, the preference prediction value 805 and the behavior prediction for the item i whose value ki = 0 in FIG. 1002 to calculate the value 803.

  Next, the recommended item determination unit 115 determines 1003 an item having a preference prediction value 805 of 1102 or more as a lower limit 1101 and an action prediction value 803 of 1112 or less as a recommended item.

  The feature space constructed in the item feature space construction processing 201 is schematically represented in two dimensions as shown in FIG. In this embodiment, an item is represented by a flat circle, a selection item with a user's item selection history is a white circle 1201 and a double-line white circle 1202, and a user's favorite item is a double-line white circle 1202. The range in which the preference prediction value 1102 or more is set is inside the distribution shape represented by 1212.

  In addition, the range of the input action predicted value upper limit 1112 or less is outside 1211 representing the distribution shape of the selected item. Therefore, an item 1203 represented by a black circle included in a range 1213 surrounded by a thick line which is a common part of the two ranges is a recommended item.

  Next, an example in which the present invention is applied to program recommendation in a television 1331 having a program reception / playback unit with a recording function will be described with reference to FIGS.

  FIG. 13 is a diagram showing the system configuration of the program recommendation device, FIG. 14 is a flowchart showing the program recommendation method, FIG. 15 is a diagram showing program information, and FIG. 17 is a diagram showing a data structure of program preferences, FIG. 18 is a diagram showing a flow of a recommended program determining method, and FIG. 19 is a diagram showing a method for acquiring program preferences.

  EPG data 1302 relating to the television program is fetched from the program receiving unit 1332.

  When all programs are captured, 1401 program vector generation unit 1312 calculates program vector 1306 using program EPG data 1302 for all programs 1402. A number of methods for generating a program vector from 1501 are known as EPG data examples. Next, the feature space calculation unit 1313 uses the program vector 1306 to construct a program feature space 1307 1403. The processes 1401 to 1403 are executed each time program EPG data is added or changed. When the user views the program at the program playback unit 1333, the program viewing history 1304 is updated. When the user makes a recording reservation at the program recording unit 1334, the program recording reservation history 1303 is updated. The program preference calculation unit 1317 calculates a program preference 1305 using the program recording reservation history 1303 and the program viewing history 1304.

  Next, when the user requests program recommendation, the behavior model creation unit 1314 uses the program viewing history 1304 to create the user behavior model 1308 1404, and the preference model creation unit 1315 uses the program preference 1305. The user's preference model 1315 is created 1405. Next, for each unviewed program 1406, a value indicating the degree of possibility that the user views the program based on the behavior model is obtained 1318, a behavior prediction value is 1407, and the user's program based on the preference model is obtained. A prediction value of the degree of likes and dislikes is obtained 1319 and set as a program preference prediction value 1408. Next, based on the obtained behavior prediction value and preference prediction value, the unexpectedness coefficient / preference prediction value combined recommendation index calculation unit 1302 calculates a recommendation index, and based on the index, the recommended program determination unit 1316 Generate 1409 list.

  Next, the program feature space construction process 1403 performed by the feature space calculation unit 1313 will be described with reference to FIG. A principal component analysis using a correlation matrix is performed on the program vector 1603. In the principal component analysis, a vector 1602 is obtained which is an inner product of the program vector 1603 and the vector 1602 and is independent from each other and has a large variance for each element of the program vector. The analysis results in the eigenvalue problem of the correlation matrix of attribute values, and an eigenvector 1602 is obtained. The eigenvalue 1612 represents the variance of the linear combination value principal component score 1601.

  Here, only the principal component corresponding to the eigenvector having an eigenvalue of 1 or more is selected and generally called the Kaiser criterion, and the value obtained by dividing the principal component score by the square root of the eigenvalue so that the items are uniformly distributed in the feature space. Let 1511 be the coordinate value in the feature space. In this case, the number of dimensions of the feature space is equal to one or more eigenvalues.

  Next, the program preference 1305 will be described with reference to FIG. The data is represented by a matrix having users 1711 to 1714 as rows and items 1701 to 1704 as columns. For example, when the degree that the user k likes the item i is “4” in the five-step evaluation, the value ki is “4”. When there is no preference information for the item i of the user k, the value ki is “0”.

  Next, a recommended program determination method in the recommended program determination unit 1316 will be described with reference to FIG. First, 1801 is calculated for all unviewed programs of the user k, and a recommendation index is calculated 1802 by the unexpectedness coefficient / preference distance combined recommendation index calculation unit 1302. Next, the programs are sorted 1803 in ascending order of the recommended recommendation index. 1804 is a recommended item in ascending order of recommendation index. At this time, the system administrator or the user can specify the recommended item number n. In that case, the top n items are determined in ascending order of recommendation index.

  Next, an example of acquiring the program preference 1305 will be described with reference to FIG. A screen 1901 is an example of a television 1331 having a program reception / playback unit with a recording function, and is operated by the remote controller 1910. A list 1902 of recorded programs is displayed on the screen, and detailed information 1904 such as EPG of the program 1903 selected by the remote controller is displayed. A button 1807 for reproducing a preference 1906 of five levels 1 to 5 for the selected program and a button 1908 for deleting are simultaneously displayed. There is a program search window 1905 by a character string in the screen, and the user inputs a search character string by remote control operation.

  When the user operates the preference 1906 by remote control operation, the preference is acquired as the preference 1305 of the program. In addition, when the user plays the program 1907 and the preference of the program is not the highest 5, it is increased by 1. Also, when the user deletes the program 1908 and there is no viewing record 1204 of the program and the preference is not the lowest one, the preference for the program is reduced by one. Further, if the preference of the program including the character string input in the program search window 1905 in the EPG data is not the highest 5, it is increased by 1. The program reception / playback unit 1331 with a recording function reflects the preferences for these programs and changes to the program preferences 1305 as needed.

111-118 ... processing unit 121 ... input / output units 102, 321, 403 ... item attribute value 103, 521 ... item selection history 104, 521 ... item preference 105 ... item feature space 106 ... behavior model 107 ... preference model 301-304 ... Item attributes 311 to 314, 501 to 504 ... items 511 to 514 ... user 401 ... principal component 402 ... coefficient 411 ... feature space coordinate value 601 ... action space feature matrix 701 ... preference space feature matrix 613, 713 ... feature space position vector 801 ... recommendation index 802 ... unexpectedness factor 805 ... preference prediction value 806 ... preference distance 803 ... behavior prediction value 804 ... behavior distance 1102 ... preference prediction value lower limit 1112 ... behavior prediction value upper limit 1213 ... recommended item range 1201 ... selected item 1202 ... preference Item 1203 ... Recommended item

Claims (11)

The attribute data of the item stored in the recording medium, the history data that the user selected the item, and the preference data of the user's item are retained, and the recommended item is determined from the attribute data, the history data, and the preference data An information recommendation method for presenting the recommended item to the user,
An item feature space is constructed according to the attribute indicated by the attribute data,
In the item feature space, creating a user behavior model from the history data, creating a user preference model from the preference data,
In the item feature space, a user behavior prediction value is calculated using the user behavior model for an unrecommended item identified from the history data,
In the item feature space, for the unrecommended item specified from the history data, the user preference model is calculated using the user preference model,
From the user behavior prediction value and the user preference prediction value, a recommendation index of an unrecommended item is calculated,
An information recommendation method, wherein an item to be recommended is determined using the index. In the information recommendation method of Claim 1,
When the user behavior prediction value is large, a value is large, and when the user preference prediction value is large, an unexpected coefficient is obtained, and an item having a small unexpected prediction coefficient and a large user preference prediction value is obtained. An information recommendation method characterized by being a recommended item. In the information recommendation method of Claim 2,
A ratio of the user behavior predicted value and the user preference predicted value is used as the unexpectedness coefficient. In the information recommendation method of Claim 2,
An information recommendation method, wherein a product or a linear sum of the unexpectedness coefficient and the reciprocal of the user preference prediction value of the unrecommended item is obtained, and the recommended item is used in ascending order of the value. In the information recommendation method of Claim 2,
An information recommendation method using the reciprocal of the Mahalanobis distance from the center of the user behavior model as the user behavior prediction value, and the reciprocal of the Mahalanobis distance from the center of the user preference model as the user preference prediction value. . A storage unit, a processing unit, and an input / output unit;
The storage unit includes item attribute data, item selection history, and item preference.
The processing unit includes a feature space calculation unit, a behavior model creation unit, a preference model creation unit, an item behavior prediction value calculation unit, an item preference prediction value calculation unit, and an unexpectedness coefficient / preference prediction value combined recommendation index calculation. And a recommended item determination unit,
The feature space part constructs an item feature space according to the attribute indicated by the attribute data of the item in the storage unit, and stores it in the storage unit.
The behavior model creation unit creates a user behavior model from the history data and stores it in a storage unit,
The preference model creation unit creates a user preference model from the preference data and stores it in a storage unit,
The item behavior prediction value calculation unit calculates a user behavior prediction value using the user behavior model for an unrecommended item identified from the history data,
The item preference prediction value calculation unit calculates a user preference prediction value using the user preference model for an unrecommended item specified from the history data,
The unexpectedness coefficient / preference prediction value combined recommendation index calculation unit calculates a recommendation index of an unrecommended item from the user behavior prediction value and the user preference prediction value,
The recommended item determination unit determines an item to be recommended using the index. The information recommendation device according to claim 6,
The recommendation index is small when the user behavior prediction value is large, and the unexpectedness prediction value is small by combining the unexpectedness coefficient and the user preference prediction value that are small when the user preference prediction value is large, An information recommendation device, wherein the value corresponding to an item having a large user preference prediction value is an index. The information recommendation device according to claim 7,
An information recommendation device using a ratio of a user behavior prediction value and a user preference prediction value as the unexpectedness coefficient. The information recommendation device according to claim 7,
The recommended item determining unit obtains a product or a linear sum of an unexpected coefficient and a reciprocal of a user preference prediction value of an unrecommended item, and sets the recommended item in ascending order of the value. The information recommendation device according to claim 7,
The user behavior predicted value calculation unit calculates the user behavior predicted value using the inverse of the Mahalanobis distance from the center of the user behavior model,
The user recommendation prediction value calculation unit calculates the user preference prediction value by using the inverse of the Mahalanobis distance from the center of the user preference model as the user preference prediction value. In the information recommendation method of Claim 1,
An input of an upper limit of the user behavior prediction value and a lower limit of the user preference prediction value, and outputting an item in which the behavior prediction value is less than or equal to the upper limit and the preference prediction value is greater than or equal to the lower limit in the feature space. A featured information recommendation method.