JP2006302097A - Cooperative filter device - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a cooperative filter device capable of reducing a calculation volume to shorten a processing time. <P>SOLUTION: Logs including user identification data and item identification data are accumulated as history data in a history data accumulation part 21, and history data is processed by a cooperative filter part 23. A processing object data extraction part 29 extracts logs to be processed in the cooperative filter part out of many logs accumulated in the history data accumulation part. In the processing object data extraction part 29, fundamental data of a relation between the numbers of logs and processing times of cooperative filtering is stored in a fundamental data storage part, and a processing time calculation part calculates a predictive processing time corresponding to the inputted target number of logs from fundamental data. A log extraction part of the processing object data extraction part 29 extracts logs accumulated in the history data accumulation part 21, of which the number is the inputted target number of logs. <P>COPYRIGHT: (C)2007,JPO&INPIT

Description

  The present invention relates to a collaborative filter device that obtains an item such as a product to be recommended to a user by calculation processing, and more particularly to a collaborative filter device that can reduce the amount of calculation and shorten the processing time.

  Conventionally, a collaborative filter device has been known as a representative technique used for providing recommendation information of products such as contents. Collaborative filtering is a system that estimates the preference of a recommendation target user based on the preference information of a user who has taken a behavior similar to that of a recommendation target user using data in which the user's preference is recorded in the form of past behavior. It is.

  The collaborative filter device includes a database of product purchase histories by a plurality of users. In collaborative filtering, the similarity between users is defined from the product purchase history. As the similarity, for example, the Jackard coefficient and the cosine are known. The similarity between the recommendation target user and the comparison target user is calculated from the purchase history data. From the similarity, a recommended value of a product that has not yet been purchased by the recommended user is calculated. Then, product information with a high recommendation value is provided to the recommendation target user.

In the collaborative filter device, the greater the number of users and the greater the purchase history data, the better the filtering accuracy. Therefore, it is desirable that the collaborative filter device includes a database of a large amount of purchase history data by a large number of users and is configured to perform the above-described collaborative filtering process on the history data of this database.
The collaborative filtering technique is disclosed in Patent Document 1, for example.
JP 2004-326227 A (page 5-8, FIG. 1)

  However, the conventional collaborative filter device has a problem that when a large amount of purchase history data is processed, the amount of calculation becomes enormous and the processing time becomes too long. In other words, in the past, there was no way of knowing in advance how long the processing time was, and when collaborative filtering was started, there was a situation where the processing would not end indefinitely and the processing would be interrupted It was easy. In particular, since the collaborative filter device operates within the limits of limited hardware resources such as memory and HDD, the processing time is significantly increased when the computer memory is full and swapping to the HDD device occurs. Furthermore, there is a problem that the processing cannot be performed when the HDD device is full.

  The present invention has been made to solve the conventional problems, and an object of the present invention is to provide a collaborative filter device capable of reducing the amount of calculation and shortening the processing time.

  The collaborative filter device of the present invention stores a log including user identification data and item identification data generated each time an item is acquired by a large number of users as history data, and accumulates in the history data accumulation unit. A collaborative filter unit that obtains recommended items for recommendation users by collaborative filtering processing using the history data, an output unit that outputs a processing result by the collaborative filter unit, and a large number stored in the history data storage unit A processing target data extracting unit that extracts a log to be processed by the collaborative filter unit from the log, and the processing target data extracting unit stores basic data on the relationship between the number of logs and the processing time of collaborative filtering Target data storage and target number of logs to be processed by collaborative filtering A processing time calculation unit that reads the basic data from the basic data storage unit and obtains a predicted processing time corresponding to the target log number input by the log number input unit from the basic data; A processing time output unit that outputs the predicted processing time calculated by the processing time calculation unit, and a log stored in the history data storage unit based on the target log number input by the log number input unit. A log extracting unit that extracts a target number of logs. Here, the “log” is item acquisition history data, and is data including user identification data and item identification data. The item is, for example, a product such as content.

  With this configuration, a basic data storage unit that stores basic data on the relationship between the number of logs and the collaborative filtering processing time is provided, the predicted processing time corresponding to the input target log number is calculated, and the input target The number of logs is extracted from the history data storage unit. Since the prediction processing time is known in advance, the number of logs can be set so that the prediction processing time is appropriate. In addition, by extracting logs of the set number of logs from the history data, the amount of calculation can be reduced so as to achieve the calculated predicted processing time. In this way, the amount of calculation can be reduced and the processing time can be shortened.

  In the collaborative filter device of the present invention, the log extraction unit is based on a theoretical similarity value determined from the number of personal logs of the recommendation target user, the number of personal logs of the comparison target user, and the number of logs of the common acquisition items of both users. The number of personal logs of the comparison target user is limited according to the number of personal logs of the recommendation target user, and the log of the comparison target user corresponding to the limited number of personal logs is extracted from the history database. Has been.

  With this configuration, by extracting logs from historical data based on theoretical similarity values, it is possible to leave logs of users who are close to the recommendation target user and the number of items to be acquired, and the number of logs can be increased while maintaining higher filtering accuracy. Can be reduced.

  In the collaborative filter device of the present invention, the log extraction unit includes a means for setting a similarity threshold value of the similarity theory value and a comparison target corresponding to the similarity theory value equal to or higher than the similarity threshold value. Means for setting a threshold reference personal log number that is the number of personal logs of the user, and a count value of the number of logs of the comparison target user corresponding to the threshold reference personal log number from the logs of the history data storage unit The threshold reference individual when the similarity threshold value is set to the maximum within a range in which the count value obtained by the count acquisition unit achieves the target number of logs. It is configured to limit the number of personal logs of the comparison target user of the log extraction source to the number of logs.

  With this configuration, the similarity threshold value of the similarity theory value is set, and by using this similarity threshold value to extract the log from the history data, the log of the user whose acquisition target number is close to the recommendation target user The number of logs can be reduced while maintaining higher filtering accuracy.

  Moreover, in the collaborative filter device of the present invention, the basic data storage unit stores basic data on the relationship between the number of logs and the collaborative filtering processing time when the recommendation target user is divided into a plurality of items, and the processing time calculating unit Calculates the prediction processing time using the basic data when the recommendation target user is divided. With this configuration, the amount of calculation can be reduced and the processing time can be shortened.

  Moreover, in the collaborative filter device of the present invention, the basic data storage unit stores basic data on the relationship between the number of logs and the collaborative filtering processing time when the log is divided into a plurality of logs, and the processing time calculation unit includes: The predicted processing time is calculated using the basic data when the log is divided. With this configuration, the amount of calculation can be reduced and the processing time can be shortened.

  Further, in the collaborative filter device of the present invention, the basic data storage unit stores basic data on the relationship between the number of logs and the collaborative filtering processing time when the recommendation target user and the log are each divided into a plurality of pieces, and the processing The time calculation unit calculates the predicted processing time using the basic data when the pre-recommended user and the log are divided. With this configuration, the amount of calculation can be reduced and the processing time can be shortened.

  In the collaborative filter device of the present invention, the processing target data extraction unit further reads a setting processing time input unit that inputs a setting processing time of collaborative filtering and the basic data from the basic data storage unit, and A target log number calculation unit that obtains a target log number corresponding to the set processing time input by the processing time input unit from the basic data, and the log extraction unit, when the set processing time is input, The log of the target log number obtained by the target log number calculation unit is extracted from the logs stored in the history data storage unit.

  With this configuration, the target number of logs corresponding to the input set processing time is calculated according to the input of the set processing time, and logs of the calculated target log number are extracted from the history data storage unit. Therefore, even with this configuration, it is possible to suitably reduce the amount of calculation and shorten the processing time.

  Further, in the collaborative filter device of the present invention, the processing target data extraction unit has a theoretical similarity value determined from the number of personal logs of recommendation target users, the number of personal logs of comparison target users, and the number of logs of common acquisition items of both users. A similarity threshold value input unit for inputting a set similarity threshold value, a target log number acquisition unit for obtaining a target log number corresponding to the set similarity threshold value, and a predicted processing time corresponding to the target log number When the set similarity threshold value is input, the target log number acquisition unit obtains a value equal to or higher than the similarity threshold value from the log of the history data storage unit. The count value of the number of logs of the comparison target user log corresponding to the theoretical similarity value is obtained as the target log number, and the processing time calculation unit reads the basic data from the basic data storage unit Obtains the estimated processing time corresponding to the number of target logs acquired by the target log number obtaining unit from the basic data.

  With this configuration, the target log number is calculated from the input set similarity threshold, and the predicted value of the processing time is calculated from the target log number. Therefore, the log can be reduced so that the processing time is appropriate by adjusting the set similarity threshold. By using the similarity threshold of the similarity theoretical value, the amount of calculation can be reduced while maintaining high filtering accuracy, and the processing time can be shortened.

  According to another aspect of the present invention, there is provided a history data storage unit that stores user identification data generated each time an item is acquired by a large number of users and a log including the item identification data as history data, and the history data storage unit. From a large number of logs stored in the history data storage unit, for a collaborative filter device comprising a collaborative filter unit that seeks a recommended item for a recommendation target user by collaborative filtering processing using the history data stored in A collaborative filter support apparatus that extracts a log to be processed by the collaborative filter unit. This collaborative filter support device includes a basic data storage unit that stores basic data on the relationship between the number of logs and the processing time of collaborative filtering, a target log number input unit that inputs a target log number to be processed by collaborative filtering, By reading the basic data from the basic data storage unit and calculating a predicted processing time corresponding to the target log number input by the log number input unit from the basic data, and by the processing time calculation unit Based on the target log number input by the log number input unit, a processing time output unit that outputs the calculated predicted processing time, and logs of the target log number from the logs accumulated in the history data storage unit And a log extracting unit for extracting. This configuration also provides the above-described advantages of the present invention.

  According to another aspect of the present invention, there is provided a history data storage unit that stores user identification data generated each time an item is acquired by a large number of users and a log including the item identification data as history data, and the history data storage unit. From a large number of logs stored in the history data storage unit, for a collaborative filter device comprising a collaborative filter unit that seeks a recommended item for a recommendation target user by collaborative filtering processing using the history data stored in A collaborative filter support processing method for extracting a log to be processed by the collaborative filter unit by computer processing. This collaborative filter support processing method inputs a target log number to be processed by collaborative filtering from a target log number input unit, and from a basic data storage unit that stores basic data on the relationship between the number of logs and the processing time of collaborative filtering. Reading basic data, obtaining a predicted processing time corresponding to the target log number input by the log number input unit from the basic data read from the basic data storage unit, and outputting the calculated predicted processing time as a processing time The log having the target number of logs is extracted from the log accumulated in the history data accumulation unit based on the target log number output from the log unit and input by the log number input unit. This method also provides the above-described advantages of the present invention.

  According to another aspect of the present invention, there is provided a history data storage unit that stores user identification data generated each time an item is acquired by a large number of users and a log including the item identification data as history data, and the history data storage unit. From a large number of logs stored in the history data storage unit, for a collaborative filter device comprising a collaborative filter unit that seeks a recommended item for a recommendation target user by collaborative filtering processing using the history data stored in A collaborative filter support processing program for causing a computer to execute processing for extracting a log to be processed by the collaborative filter unit. This collaborative filter support processing program inputs a target log number to be processed by collaborative filtering from a target log number input unit, and from a basic data storage unit that stores basic data on the relationship between the number of logs and the processing time of collaborative filtering. Reading basic data, obtaining a predicted processing time corresponding to the target log number input by the log number input unit from the basic data read from the basic data storage unit, and outputting the calculated predicted processing time as a processing time Based on the target log number output from the log number input unit and causing the computer to execute processing for extracting the log of the target log number from the log stored in the history data storage unit. This configuration also provides the above-described advantages of the present invention.

  According to another aspect of the present invention, there is provided a history data storage unit that stores user identification data generated each time an item is acquired by a large number of users and a log including the item identification data as history data, and the history data storage unit. From a large number of logs stored in the history data storage unit, for a collaborative filter device comprising a collaborative filter unit that seeks a recommended item for a recommendation target user by collaborative filtering processing using the history data stored in A collaborative filter support apparatus that extracts a log to be processed by the collaborative filter unit. The collaborative filter support apparatus includes a basic data storage unit that stores basic data of a relationship between the number of logs and a processing time of collaborative filtering, a setting processing time input unit that inputs a setting processing time of collaborative filtering, and the basic data storage The basic data is read from the unit, and the target log number calculation unit that obtains the target log number corresponding to the setting processing time input by the setting processing time input unit from the basic data is calculated by the target log number calculation unit A target log number output unit that outputs the target log number, and a log of the target log number from logs accumulated in the history data accumulation unit based on the target log number calculated by the target log number calculation unit And a log extracting unit for extracting. With this configuration, the target number of logs corresponding to the input set processing time is calculated according to the input of the set processing time, and logs of the calculated target log number are extracted from the history data storage unit. Therefore, even with this configuration, it is possible to suitably reduce the amount of calculation and shorten the processing time.

  According to another aspect of the present invention, there is provided a history data storage unit that stores user identification data generated each time an item is acquired by a large number of users and a log including the item identification data as history data, and the history data storage unit. From a large number of logs stored in the history data storage unit, for a collaborative filter device comprising a collaborative filter unit that seeks a recommended item for a recommendation target user by collaborative filtering processing using the history data stored in A collaborative filter support processing program for causing a computer to execute processing for extracting a log to be processed by the collaborative filter unit. This collaborative filter support processing program inputs collaborative filtering setting processing time from a setting processing time input unit, and reads out basic data from a basic data storage unit that stores basic data on the relationship between the number of logs and the collaborative filtering processing time. The target log number corresponding to the set processing time input by the set processing time input unit is obtained from the basic data read from the basic data storage unit, and the calculated target log number is calculated from the target log number output unit. And outputting a log of the target number of logs from the log stored in the history data storage unit based on the target log number calculated by the target log number calculation unit. This configuration also provides the above-described advantages of the present invention.

  According to another aspect of the present invention, there is provided a history data storage unit that stores user identification data generated each time an item is acquired by a large number of users and a log including the item identification data as history data, and the history data storage unit. From a large number of logs stored in the history data storage unit, for a collaborative filter device comprising a collaborative filter unit that seeks a recommended item for a recommendation target user by collaborative filtering processing using the history data stored in A collaborative filter support apparatus that extracts a log to be processed by the collaborative filter unit. The collaborative filter support device includes a means for inputting a similarity threshold of a theoretical similarity value determined from the number of personal logs of a user to be recommended, the number of personal logs of a user to be compared, and the number of logs of a common acquisition item of both users; Means for setting a threshold reference personal log number that is the number of personal logs of a comparison target user corresponding to a similarity theoretical value equal to or greater than the similarity threshold; and a comparison target user of the threshold reference personal log number Means for extracting a log from the history data storage unit. With this configuration, the similarity threshold value of the similarity theory value is set, and by using this similarity threshold value to extract the log from the history data, the log of the user whose acquisition target number is close to the recommendation target user The number of logs can be reduced while maintaining higher filtering accuracy, and the processing time can be reduced by reducing the amount of calculation.

  The present invention provides a basic data storage unit that stores basic data on the relationship between the number of logs and the collaborative filtering processing time, calculates a predicted processing time corresponding to the target number of logs, and stores a log of the target number of logs as a history data storage unit Thus, it is possible to provide a collaborative filter device that has the effect of reducing the amount of calculation and shortening the processing time.

  Hereinafter, a collaborative filter device according to an embodiment of the present invention will be described with reference to the drawings.

  A collaborative filter device according to an embodiment of the present invention is shown in FIG. In FIG. 1, the collaborative filter device 1 is shown together with a product purchase system 3. In the present embodiment, the product is content. The product purchase system 3 includes a purchase processing unit 5 and a content database 7. The purchase processing unit 5 is composed of an information distribution device including a WWW server, and has a function of communicating with the user terminal device 11 via the Internet 9. The content database 7 stores a large amount of content data to be sold to the user. The user terminal device 11 is a personal computer, a portable terminal, or the like. The purchase processing unit 5 reads content data from the content database 7 and provides it to the user terminal device 11 via the Internet 9. Each time the purchase processing unit 5 provides content data to the user terminal device 11, the purchase processing unit 5 generates history data and provides it to the collaborative filter device 1, and the collaborative filter device 1 accumulates the history data.

  FIG. 2 is a flowchart of the operation of the overall configuration shown in FIG. 1, and shows the flow of content purchase and history data accumulation. When the purchase processing unit 5 is accessed by the user terminal device 11 to search for the content (S1), and the download of the searched content is requested (S3, Yes), the content is downloaded (S5), and then downloaded. History data is generated and stored (S7). If the purchase operation has not ended (No at Step S9), the process returns to Step S1.

  The collaborative filter device 1 includes a history data storage unit 21, a collaborative filter unit 23, a recommendation information distribution unit 25, a product information storage unit 27, and a processing target data extraction unit 29.

  The collaborative filter device 1 has a communication function with the purchase processing unit 5, this communication function functions as a history data input unit, inputs history data provided from the purchase processing unit 5, and the history data accumulates history data. Stored in the unit 21. In the present embodiment, the history data of a single product purchase is referred to as a log.

  The log includes user identification data and product identification data. The user identification data is data for identifying the user who has purchased the product, and the product identification data is data for specifying the purchased product. FIG. 3 shows an example of data. In FIG. 3, the user identification data is a user ID, and the product identification data is a content ID. Note that the content that is a product is an example of an item that should be recommended to the user in the present invention. Further, the history data, that is, the log may include data such as purchase date and time in addition to the above-described user identification data and product identification data.

  The collaborative filter unit 23 is configured to obtain a recommended item for the recommendation target user by collaborative filtering processing using the history data accumulated in the history data accumulation unit 21. The collaborative filter unit 23 is realized by a computer executing a collaborative filtering process program. The collaborative filter unit 23 includes a similarity calculation unit 31 and a recommendation unit 33.

  The recommended information distribution unit 25 includes an information distribution device including a WWW server, and corresponds to an output unit that outputs a processing result from the collaborative filter unit 23. The merchandise information storage unit 27 stores merchandise information about a large number of merchandise. The recommendation information distribution unit 25 receives supply of data specifying a recommended product to be recommended to the recommendation target user from the collaborative filter unit 23. Then, the recommended information distribution unit 25 reads the product information of the recommended product from the product information storage unit 27 and provides it to the user terminal device 11 of the recommendation target user via the Internet 9. The merchandise information (recommended information) may be incorporated into an e-mail, or in another example, printed on paper and sent by mail.

  The processing target data extraction unit 29 is configured to extract a log to be processed by the collaborative filter unit 23 from a large number of logs stored in the history data storage unit 21. The internal configuration itself of the collaborative filter unit 23 may be the same as that of the related art, but the collaborative filter unit 23 uses the log extracted by the processing target data extracting unit 29 as a processing target (input data). That is, this embodiment differs from the prior art in that a processing target data extraction unit 29 is provided in addition to the normal collaborative filter unit 23.

  The processing target data extraction unit 29 is integrated with the collaborative filter unit 23 and is realized by the same computer as the collaborative filter device 23. That is, a program for realizing the processing target data extraction unit 29 is executed by the computer of the collaborative filter unit 23. The processing target data extraction unit 29 has a function of adjusting processing target data before processing in the collaborative filter unit, and in this respect, it can also be called a collaborative filter support device. From this viewpoint, the collaborative filter support device is integrated and incorporated in the collaborative filter device. The processing target data extraction unit 29 has been described as being realized by the same computer as the collaborative filter device 23, but may be realized by another computer.

  The overall outline of the collaborative filter device 1 according to the present embodiment has been described above. Next, the collaborative filtering process by the collaborative filter unit 23 will be described, and then the processing target data extraction unit 29 characteristic of the present embodiment will be described.

"Collaborative filtering process"
FIG. 4 shows a specific example of the purchase history together with the similarity of the Jackard coefficient. FIG. 4 shows the purchase history data of the users A, B, and C and the similarity. User A purchases content U1, U2, U3, user B purchases content U2, U3, U4, U5, and user C purchases content U2, U3, U5, U6, U7.

When using the Jaccard coefficient, the similarity between users is defined as follows.
Similarity between users A and B: Sim (A, B) = (A∩B) / (A∪B)
(A∩B) is the number of commonly purchased products. Further, (A∪B) is a value obtained by subtracting the number of common purchased products from the total number of purchased products of the users A and B (the number of types of products purchased by the users A and B). Therefore, in the case of FIG. 4, the calculated value of the similarity is as follows.
Similarity between users A and B: Sim (A, B) = 2/5 = 0.4
Similarity between users A and C: Sim (A, C) = 2/6 = 0.33
Similarity between users B and C: Sim (B, C) = 3/6 = 0.5

  FIG. 5A and FIG. 5B show a list of recommended values obtained from the similarity. For example, the user A has not purchased the content U4 yet. The content U4 is purchased by the user B. Therefore, the recommended value of the content U4 for the user A is Sim (A, B) = 0.4.

  Further, the user A has not purchased the content U5 yet. The content U5 is purchased by the users B and C. Therefore, the recommended value of the content U5 for the user A is obtained by the sum of Sim (A, B) and Sim (A, C).

  In this way, as shown in FIG. 5A, the recommended value of the unpurchased product for each user is obtained by adding the similarity of the user who purchased the product. In FIG. 5B, the contents are arranged in descending order of recommended values. The recommended content can be determined from the recommended values in FIG. That is, content with a large recommended value is obtained as recommended content. At this time, the content having the maximum recommended value may be selected. A predetermined number of contents having a large recommended value may be selected. In addition, content whose recommended value is a predetermined value or more may be selected.

  In the collaborative filter unit 23 of FIG. 1, the similarity calculation unit 31 performs the processing shown in FIG. 4 to calculate the similarity between users. And the recommendation part 33 performs the process shown by FIG. 5, calculates a recommendation value, and determines a recommendation content from a recommendation value. The content ID of the recommended content is provided to the recommended information distribution unit 25 together with the user identification data of the recommendation target user. The recommendation information distribution unit 25 reads the product information associated with the content ID from the product information storage unit 27 and transmits it to the recommended user.

  FIG. 6 is a flowchart showing the operation of the collaborative filter unit 23. The similarity between users is calculated by the similarity calculation unit 31 (S11), the recommendation level (recommended value) of the recommended content is calculated by the recommendation unit 33 (S13), and a recommendation list is generated by the recommendation unit 33. (S15).

  The above is collaborative filtering when using the Jackard coefficient. Next, collaborative filtering when using cosine will be described.

  FIG. 7 shows a specific example of purchase history data including evaluation values together with cosine similarity. The purchase history data includes evaluations (1 to 5) for each content of each user. For example, the user A has purchased the content U1, and the evaluation value of the content U1 by the user A is 3. The evaluation value is sent from the user terminal device 11 of FIG. 1 to the purchase processing unit 5, and the purchase processing unit 5 incorporates the evaluation value into the history data.

The cosine similarity is represented by the cosine of each user's vector, as shown. In the example of the figure, the similarity between users is as follows.
Similarity between users A and B: Sim (A, B) = 0.648
Similarity between users A and C: Sim (A, C) = 0.531
Similarity between users B and C: Sim (B, C) = 0.589

  FIGS. 8A and 8B show a list of recommended values obtained from the purchase history data of FIG. 7 and the similarity. In this case, the recommendation value is calculated by multiplying the evaluation value by the similarity. For example, the user A has not purchased the content U4 yet. The content U4 is purchased by the user B. The evaluation value of the user B for the content U4 is 3. Therefore, the recommended value of the content U4 for the user A is Sim (A, B) × (B evaluation value) = 0.648 × 3 = 2.59.

  Further, the user A has not purchased the content U5 yet. The content U5 is purchased by the users B and C. The evaluations of the users B and C for the content U5 are 4, 3 respectively. Using these, the recommended value of the content U5 for the user A is obtained as the sum of “Sim (A, B) × (B evaluation value)” and “Sim (A, C) × (C evaluation value)”. It is done. In FIG. 8A, the evaluation values of B and C are simply expressed as B and C.

  In this way, as shown in FIG. 8A, the recommended value of each user's unpurchased product is obtained by adding the “product of similarity and evaluation value” of the user who purchased the product. In FIG. 8B, the contents are arranged in descending order of recommended values. From the recommended values in FIG. 8, the recommended content can be determined as described in FIG.

FIG. 9 is a modification of the processing of FIGS. In the purchase history data in FIG. 7, the evaluation values were 1 to 5. In the example of FIG. 9, the evaluation value is −2 to 2. In this case, as shown in FIG. 9A, the similarity is as follows.
Similarity between users A and B: Sim (A, B) = 0.89
Similarity between users A and C: Sim (A, C) = 0.71
Similarity between users B and C: Sim (B, C) = 0.63

  The recommendation list is as shown in FIGS. 9B and 9C. In FIG. 9C, a list is created only with content whose recommended value is positive (plus).

  Incidentally, although the collaborative filtering process using the Jackard coefficient and the cosine has been described regarding the calculation of the similarity, the collaborative filtering process using a process other than the above may be used.

"Processing data extractor"
Next, the processing target data extraction unit 29 characteristic of this embodiment will be described in detail. The processing target data extracting unit 29 realizes a simulation function for predicting the processing time of collaborative filtering and a function for reducing the amount of data to be processed by collaborative filtering based on the simulation result as follows.

  FIG. 10 shows the configuration of the processing target data extraction unit 29. The processing target data extraction unit 29 includes a basic data storage unit 41 that stores basic data on the relationship between the number of logs and the collaborative filtering processing time, a target log number input unit 43 that inputs the target log number of processing targets for collaborative filtering, The basic data is read from the basic data storage unit 41 and calculated by the processing time calculation unit 45 and the processing time calculation unit 45 which obtains the predicted processing time corresponding to the target log number input by the log number input unit from the basic data. The processing time output unit 47 that outputs the predicted processing time, the permission determination receiving unit 49 that receives permission determination for the predicted processing time output by the processing time output unit 47, and the target log input by the target log number input unit 43 And a log extraction unit 51 that extracts a target number of logs from the logs accumulated in the history data accumulation unit 21 based on the number.

  The basic data in the basic data storage unit 41 is generated by the basic data acquisition experiment processing unit 53. The basic data is the following data.

  FIG. 11A and FIG. 11B show the relationship between the number of logs and the processing time when the number of recommendation target users is fixed. The number of recommended users is, for example, 1,000,000. Ideally, as shown in FIG. 11A, the number of logs is proportional to the processing time, and the relationship between the two is represented by a single straight line.

  However, the actual graph does not become a single straight line as shown in FIG. In a computer that performs collaborative filtering, the capacity of the memory and HDD is finite. As the number of logs increases, the memory becomes insufficient, SWAP occurs at point A in FIG. 11B, and the slope of the processing time line suddenly increases. Further, at point B in FIG. 11B, the capacity of the HDD reaches a limit point, and processing becomes impossible.

  The basic data acquisition experiment processing unit 53 performs the process of measuring the data in FIG. 11B by the basic experiment process of actually causing the computer to execute the collaborative filter process and measuring the processing time.

  FIG. 12 shows the processing of the basic data acquisition experiment processing unit 53. In the basic experiment, the number x of recommended users is fixed, the number y of logs is increased at regular intervals, and the processing time t at each y value is measured. Here, the time actually required for the processing of the sample log by the collaborative filter unit 23 is measured. The sample log may be an actual log stored in the history data storage unit 21. That is, the basic data acquisition experiment processing unit 53 reads the history data of the history data storage unit 21 and performs a basic experiment.

  As shown in FIG. 12, processing times t1 and t2 corresponding to the respective log numbers y1 and y2 are obtained, and points p1, p2, and so on are plotted. When the point p2 is obtained, the slope of the line p1-p2 is obtained. Next, when the point p3 is obtained, the slope of the line p2-p3 is obtained. This process is repeated.

  In the example of FIG. 12, the slope of the line p2-p3 suddenly increases. As the slope increases, the middle point of the line is taken. In the example in the figure, since the slope of the line p2-p3 has increased, y4 is set between y2 and y3, and the processing time t4 for y4 is measured. Thus, the inflection point p4 is obtained, and the graph of FIG. 12 is obtained.

  FIG. 12 is simplified and simplified for easy understanding. Actually, the measurement of the processing time t for a larger number of logs y may be repeated. Then, the inflection point of the inclination may be obtained by performing calculation processing by the bisection method as described above. Further, the Newton method may be applied instead of the bisection method. Further, the graph of FIG. 12 may be obtained by the steepest descent method.

  When the graph of FIG. 12 is obtained, the inflection point log number Y is obtained. The inflection point log number Y is the number of logs corresponding to the inflection point of the graph (= point A in FIG. 11B). The inflection point log number Y can also be said to be the optimal number of logs that use computer resources to the maximum. When the inflection point log number Y is applied, the memory capacity is used below the limit, the log number is increased as much as possible without extremely increasing the processing time, and higher filtering accuracy can be obtained.

  Further, the relationship between the number of logs and the processing time when the number of recommended target users is fixed is obtained from the graph of FIG. t = ay + b (0 <y <Y), t = cy + d (Y <y) (a, b, c, d are coefficients). Data corresponding to this expression is stored in the basic data storage unit 41 as basic data.

  The function of the basic data acquisition experiment processing unit 53 has been described above. Here, a graph is used for easy understanding of the explanation. Of course, the actual calculation process is equivalent to the process of plotting on the graph, and the graph need not be actually used.

  Returning to FIG. 10, the target log number input unit 43, the processing time output unit 47, and the permission determination reception unit 49 are configured by input / output devices such as a keyboard and a display. The input device is operated by the operator to input a desired minimum number of logs. This log minimum value is the minimum number of logs considered necessary for ensuring the necessary accuracy. This log minimum value is input as the target log number to the target log number input unit 43 and acquired by the processing time calculation unit 45.

  The processing time calculation unit 45 calculates a predicted processing time from the target number of logs. The basic data is read from the basic data storage unit 41, the target log number is applied to the basic data, and the prediction processing time is calculated. The basic data is in the form of an expression of the relationship between the log number y and the processing time t as described above. What is necessary is just to substitute the target log number into this formula. The processing time output unit 47 outputs the calculated predicted processing time. The predicted processing time is displayed on the display. If the processing time is appropriate, the operator performs a permission operation on the input device, and the permission determination receiving unit 49 receives a permission determination (satisfaction determination). The permission determination is transmitted from the permission determination reception unit 49 to the log extraction unit 51. The log extraction unit 51 also acquires the target log number input to the target log number input unit 43. When acquiring the permission determination, the log extraction unit 51 extracts a target number of logs from the logs accumulated in the history data accumulation unit 21. The log extraction process is arbitrary. For example, in the log extraction process, logs may be extracted at random, and the number of logs may be reduced to the target number of logs. However, in the present embodiment, it is preferable to use a similarity theory value and a similarity threshold value, which will be described later, thereby improving the accuracy of filtering. Details of this characteristic extraction process will be described later.

  FIG. 13 is a flowchart showing the operation of the processing target data extraction unit 29. The total number x of recommended users is input (S21), and a basic data acquisition experiment is performed by the basic data acquisition experiment processing unit 53 (S23), and processing end prediction graph data that is basic data is generated (S25). Then, the minimum log value is input (S27), and the predicted processing time is obtained from the basic data and displayed (S29). Whether or not the permission determination is accepted is determined (S31), and if the permission determination is not accepted (S31, No), the process returns to step S27 and another log minimum value is input. When the permission determination is accepted (S31, Yes), the total number of logs in the history data storage unit 21 is compared with the target log number, and it is determined whether or not the total log number is larger than the target log number (S33). If the total number of logs is equal to or less than the target number of logs (S33, No), the process ends. The processing target data extraction unit 29 causes the collaborative filter unit 23 to process all the logs in the history data storage unit 21. If the total number of logs is larger than the target number of logs in step S33 (S33, Yes), the logs to be processed are extracted from all the logs, and the number of logs is reduced to the target number of logs (S35). The extracted log is transmitted from the processing target data extraction unit 29 to the collaborative filter unit 23, and the collaborative filter unit 23 performs filter processing.

  FIG. 14 shows a screen display corresponding to the operation of the processing target data extraction unit 29. The screen of FIG. 14 is displayed on a computer display. As shown in FIG. 14, first, an input request screen for x (the number of recommended users) for the basic data experiment is displayed. When the number of recommended users is input, a screen indicating that a basic experiment is being performed is displayed. When the basic experiment is completed, an input request screen for the minimum log value (target log number) is displayed. When the log minimum value is input, the predicted processing time is displayed, and a “process execution” button and a “cancel” button are displayed. When the “Cancel” button is displayed, Step S31 in FIG. 13 is No, and the log minimum value input request screen is displayed again. When the “process execution” button is pressed, step S31 is Yes and log extraction is performed. By keeping the relationship between the number of logs obtained in the basic experiment and the processing time, it is not necessary to repeat the basic experiment. In this case, the basic experiment process is omitted in FIGS. 13 and 14. Good.

  Next, the log extracting unit 51 in FIG. 10 will be described in detail. In the present embodiment, the log extraction unit 51 can extract a part of all logs using the theoretical similarity value as described below while keeping the accuracy of the filtering process high.

  Here, first, the similarity based on the Jackard coefficient in the two specific examples of FIG. 15 is considered. In FIG. 15A, the number of purchases of the recommendation target user A is 10, the number of purchases of the comparison target user (similar user) B is 10, and the number of common purchase products is 5. In this case, the similarity is 0.33. On the other hand, in FIG. 15B, the number of purchases of the recommendation target user A is 10, the number of purchases of the comparison target user (similar user) C is 100, and the number of common products is 5. In this case, the similarity is 0.04. Therefore, the purchase product b of the user B, who has a smaller purchase quantity, is the recommended product with higher reliability than the purchase product c of the user C, whose purchase quantity is far.

  In this way, even if the common purchase product is the same, the similarity of users having similar preference information such as the number of purchases as a base increases. This means that the number of logs can be reduced without reducing reliability, that is, accuracy as much as possible, by extracting logs of users whose purchases are close in consideration of similarity.

  In the present embodiment, log extraction processing is performed based on such a viewpoint. In particular, in the present embodiment, as described below, the number of purchases that maximizes the similarity as much as possible is specified by objective data processing by means of processing based on the theoretical similarity value, and this specified purchase is determined. An extraction process that leaves a number of comparison target users' logs is realized.

  FIG. 16 is an example of a similarity theoretical value table used in the present embodiment. In this example, the number of recommended users to be purchased is fixed at 10. Then, a Jackard coefficient is calculated as a theoretical value of similarity from the purchase quantity of each comparison target user and each common purchase quantity. FIG. 16 is a table of similarity degree theoretical values for various combinations of the user purchase quantity to be compared and the common purchase quantity.

  In FIG. 16, a triangular zone A is a zone having a theoretical similarity value of 0.33 or more. On the other hand, zone B is a zone having a theoretical similarity value of less than 0.33. Note that the zone C is an unnecessary zone (the number of purchases ≦ the number of common zones, and when the number of purchases = the common number, there is no recommended product).

  Thus, when the minimum value of the theoretical similarity value is set, the illustrated zone A is formed. This value is called a similarity threshold in the present invention. That is, in the example of the figure, the similarity threshold is 0.33.

  When the similarity threshold is determined, zone A is determined, and the minimum value (5) and maximum value (30) of the number of comparison target users purchased in zone A are determined. Then, purchase history data of the comparison target user corresponding to the number of purchases from the minimum value to the maximum value can be extracted.

  Here, based on the concept described in the specific example of FIG. 15, if the similarity threshold value is increased, comparison target users whose purchase numbers with the recommendation target user are close are narrowed down. In addition, the comparison target users are narrowed down so that the degree of similarity is as high as possible, that is, the recommendation accuracy is increased. However, if the similarity threshold value is increased, the comparison target users are narrowed down (zone A is narrowed), and the number of history data is reduced.

  Therefore, in the present embodiment, the similarity threshold is maximized within a range in which the amount of history data to be extracted can be achieved. As a result, a necessary amount of history data can be extracted while keeping the accuracy as high as possible.

  In the example of FIG. 16, in the zone A when the similarity threshold is 0.33, the maximum value of the purchased number is 30 and the minimum value is 5. In the present embodiment, history data of all comparison target users in the purchased number range is extracted. Accordingly, history data other than zone A (history data of zone B on the left side of zone A) is also mixed. Strictly speaking, the accuracy is improved when history data is extracted from only zone A (however, the processing time becomes longer).

  Next, with reference to FIG. 17, the log extraction processing of the present embodiment will be described more specifically. FIG. 17 is also a table of theoretical similarity values similar to FIG. However, FIG. 16 is a diagram of general explanation using the purchased number. On the other hand, FIG. 17 is a table expressed using a log that is history data actually processed in the present embodiment.

  That is, in FIG. 17, the number of purchases is replaced with a log. The number of logs for a single user is referred to as the number of individual logs in order to clarify the distinction from the total number of logs. Therefore, in FIG. 17, the number of personal logs of the recommendation target user is fixed to 10. FIG. 17 is a table of similarity theoretical values corresponding to combinations of the number of personal logs of the comparison target user and the number of logs of the common acquired item.

  Also in the example of FIG. 17, the similarity threshold is 0.33. When the similarity threshold is set, the “threshold reference personal log number” is determined. The “threshold reference personal log number” is the number of personal logs of the comparison target user corresponding to the similarity similarity value equal to or higher than the similarity threshold. The “threshold reference personal log number” is the number of personal logs from the maximum personal log number (30) to the minimum personal log number (5) in zone A corresponding to the similarity theoretical value equal to or higher than the similarity threshold. It is.

  When the “threshold reference personal log number” is determined, the number of logs of the comparison target user corresponding to the “threshold reference personal log number” is counted from all the logs of the history data, and the count value is obtained. In the example of FIG. 17, the number of all logs of all the comparison target users whose personal logs are 5 to 30 is obtained. In this specification, this count value means a value obtained from the actual number of logs corresponding to a given condition among logs of history data. The counting process is a process for obtaining the actual number of logs corresponding to such a given condition, and may be an arbitrary process and may include a search or the like. Here, the given condition is the similarity threshold value or the threshold reference personal log number determined therefrom. At the time of counting, the number of people corresponding to the number of logs can be obtained, the product of the number of logs and the number of people can be obtained, and the sum of the products can be calculated. In the present embodiment, based on the count value, the maximum similarity threshold is obtained in a range where logs equal to or more than the target number of logs remain. A log is extracted from the history data by setting the maximum similarity threshold.

  FIG. 18 is a block diagram showing the configuration of the log extraction unit 51 that implements the above processing. The log extraction unit 51 includes a similarity degree theoretical value table storage unit 61, a similarity threshold setting unit 63, a threshold reference personal log number setting unit 65, a count value acquisition unit 67, a count value determination unit 69, and an application threshold. A value determination unit 71 and an extraction execution unit 73 are included.

  The similarity degree theoretical value table storage unit 61 stores a table of similarity degree theoretical values shown in FIG. The similarity threshold setting unit 63 sets a similarity threshold. The threshold standard personal log number setting unit 65 refers to the similarity theoretical value table and sets the “threshold standard personal log number” corresponding to the similarity threshold. The “threshold reference personal log number” is the number of personal logs from the maximum personal log number to the minimum personal log number in the zone A corresponding to the similarity degree theoretical value equal to or higher than the similarity threshold value.

  The count value acquisition unit 67 searches the log of the history data storage unit 21. And the count value acquisition part 67 counts the number of the logs of the comparison object user corresponding to the threshold value reference | standard personal log number, and acquires a count value. The count value determination unit 69 determines the count value of the number of logs obtained by the count acquisition unit 67. The target log number is input to the count value determination unit 69. The count value determination unit 69 compares the target log number with the count value, and determines whether the count value is larger than the target log number.

  The application threshold value determination unit 71 determines an application threshold value based on the determination result of the count value determination unit 69. The application threshold is a maximum similarity threshold within a range in which the number of logs obtained by the count value acquisition unit 67 is equal to or greater than the target number of logs. The application threshold value determination unit 71 determines the application threshold value from the determination result of the count value when the similarity threshold value setting unit 63 changes the similarity threshold value by a process described later. . The extraction execution unit 73 extracts a log from the history data storage unit 21 using the application threshold value determined by the application threshold value determination unit 71. The extraction execution unit 73 extracts the log of the comparison target user corresponding to the number of personal logs corresponding to the application threshold value (threshold reference personal log number). The extraction execution unit 73 passes the extracted log to the collaborative filter unit 23.

  FIG. 19 is a flowchart showing an operation example of the log extracting unit 51 described above. In this example, the target number of logs is input to the log extraction unit 51 (S41), and the initial value of the similarity threshold is set (S43). The initial value of the similarity threshold is set to a predetermined maximum value (for example, 1). Then, the threshold standard personal log number corresponding to the similarity threshold is set (S45), and the number of logs of the comparison target user corresponding to the threshold standard personal log number is counted (S47). In this processing, the number of people corresponding to the number of logs can be obtained by searching or the like, the sum of products of the number of logs and the number of people can be calculated, and the count value can be obtained (the same applies hereinafter). Then, the count value of the log number is compared with the target log number to determine whether the count value is equal to or greater than the target log number (S49). If the count value is less than the target number of logs (S49, No), the similarity threshold value setting unit 65 resets the similarity threshold value (S51). At this time, the similarity threshold is decreased by a predetermined width. After changing the threshold value, the process returns to step S45.

  If the determination in step S49 is Yes (count value ≧ target log number), the application threshold value determination unit 71 determines the similarity threshold value set in step S51 or S43 as the application threshold value, and applies it. The number of threshold reference personal logs corresponding to the threshold is obtained (S53). Then, the extraction execution unit 73 extracts logs from the history data storage unit 21 according to the threshold reference personal log number (S55).

  In this way, in the processing of FIG. 19, the operation of determining the number of logs by reducing (decreasing) the similarity threshold is repeated, and when the count value of the number of logs exceeds the target number of logs, the similarity threshold is reached. The value is confirmed and log extraction is performed.

  FIG. 20 is a flowchart illustrating another example of the log extraction process. In FIG. 19, the similarity threshold is reduced from the maximum value, but in FIG. 20, the similarity threshold is increased conversely. That is, in FIG. 20, in step S43a, the initial value of the similarity threshold is set to a predetermined minimum value. In step 49a, the count value determination unit 69 determines whether or not the log count value is smaller than the target log count. If the determination in step S49a is No, the similarity threshold is increased in step S51a. If the determination in step S49a is Yes, the application threshold value determination unit 71 determines the similarity threshold value set in the previous loop in step S51a or S43a as the application threshold value (S53a). Other processes are the same as those in FIG.

  In this way, in the process of FIG. 20, the operation of determining the number of logs by increasing (increasing) the similarity threshold is repeated. When the count value of the number of logs falls below the target number of logs, the similarity threshold is determined (the similarity threshold becomes the value of the previous loop), and log extraction is executed.

  The loop processing related to the similarity threshold is not limited to FIGS. The width of the threshold change may not be constant. Then, convergence calculation may be performed. For example, initially the threshold is increased, and this time the threshold is decreased if the number of logs decreases too much. Such an operation is repeated, and the change width of the threshold value is gradually reduced.

  In this embodiment, since the similarity threshold is used as a reference, the number of logs to be extracted is slightly larger than the target number of logs. Therefore, the number of logs may be further adjusted, and the final number of logs may be aligned with the target number of logs. In this case, the log may be reduced at random.

  The preferred log extraction processing of the present embodiment using the similarity theoretical value and the similarity threshold has been described above. According to the present embodiment, by extracting the log from the history data based on the theoretical similarity value, it is possible to leave a log of the user whose acquisition target number is close to the recommendation target user, and to keep the filtering accuracy higher. The number of logs can be reduced.

  More specifically, by setting a similarity threshold of the similarity theoretical value, and extracting the log from the history data using this similarity threshold, the log of the user whose acquisition target number is close to the recommended target user And the number of logs can be reduced while maintaining higher filtering accuracy.

"Processing data extractor"
(Pattern for entering the setting processing time)
FIG. 21 shows another configuration example of the processing target data extraction unit 29. In the configuration of FIG. 10 described above, the target log number is input, and the predicted processing time is calculated from the target log number. On the other hand, in the configuration of FIG. 21, the setting processing time is input, and the target number of logs is calculated from the setting processing time.

  The processing target data extraction unit 29 in FIG. 21 includes a basic data storage unit 81 that stores basic data on the relationship between the number of logs and the collaborative filtering processing time, a setting processing time input unit 83 that inputs collaborative filtering setting processing time, and the like. Then, the basic data is read from the basic data storage unit 81, and the target log number calculating unit 85 for obtaining the target log number corresponding to the set processing time input by the processing time input unit 83 from the basic data, and the target log number calculating unit 85 A target log number output unit 87 that outputs the target log number calculated by the above, a permission determination reception unit 89 that receives a permission determination for the target log number output by the target log number output unit 87, and a target log number calculation unit 85. Based on the calculated target log number, a log extracting unit 91 that extracts a log of the target log number from the logs accumulated in the history data accumulating unit 21; And a basic data acquisition experiment processor 93. The basic data storage unit 81 and the basic data acquisition experiment processing unit 93 may have the same configurations as those in FIG.

  The setting processing time input unit 83, the target log number output unit 87, and the permission determination reception unit 89 are configured by input / output devices such as a keyboard and a display. The input device is operated by the operator, and the maximum value of the desired processing time is input. The maximum processing time is input to the setting processing time input unit 83 as the setting processing time, and is acquired by the target log number calculation unit 85.

  The target log number calculation unit 85 calculates the target log number from the setting processing time. The basic data is read from the basic data storage unit 41, the set processing time is applied to the basic data, and the target number of logs is calculated. The basic data is in the form of a relationship between the number of logs y and the processing time t. The setting processing time may be substituted into this equation. The target log number output unit 87 outputs the calculated target log number. The target number of logs is displayed on the display. If the target number of logs is appropriate, the operator performs a permission operation on the input device, and the permission determination receiving unit 89 receives a permission determination (satisfaction determination). The permission determination is transmitted from the permission determination reception unit 89 to the log extraction unit 91. The log extraction unit 91 also acquires the target log number calculated by the target log number calculation unit 85. When acquiring the permission determination, the log extraction unit 91 extracts a target number of logs from the logs accumulated in the history data accumulation unit 21.

  FIG. 22 is a flowchart showing the operation of the processing target data extraction unit 29. The total number x of recommended users is input (S61), the basic data acquisition experiment is performed by the basic data acquisition experiment processing unit (S63), and the process end prediction graph data as basic data is generated (S25). Then, the maximum processing time value is input (S67), and the target number of logs is obtained from the basic data and displayed (S69). Whether or not the permission determination is accepted is determined (S71), and if the permission determination is not accepted (S71, No), the process returns to step S67 and another set time maximum value is input. When the permission determination is accepted (S71, Yes), the total number of logs in the history data storage unit 21 is compared with the target log number, and it is determined whether the total log number is larger than the target log number (S73). If the total number of logs is less than or equal to the target number of logs (S73, No), the process ends. The processing target data extraction unit 29 causes the collaborative filter unit 23 to process all the logs in the history data storage unit 21. If the total number of logs is larger than the target number of logs in step S73 (S73, Yes), processing target logs are extracted from all the logs, and the number of logs is reduced to the target number of logs (S75). The extracted log is transmitted from the processing target data extraction unit 29 to the collaborative filter unit 23, and the collaborative filter unit 23 performs filter processing.

  The configuration and operation of the log extraction unit 91 are generally the same as those of the log extraction unit 51 of FIG. 10 (see FIGS. 18, 19, and 20). However, in the log extraction unit 51, the number of extracted logs is greater than or equal to the target number of logs. On the other hand, in the log extraction unit 91 of the present embodiment, the number of extracted logs is equal to or less than the target number of logs. In relation to such a difference, the configuration and operation of the log extraction unit 91 are also slightly different. Specifically, the application threshold value determination process is different, and the count value of the number of logs is equal to or less than the target number of logs. The application threshold is determined to be the maximum similarity threshold in the range.

  That is, when the process of FIG. 19 is modified and the count value exceeds the target log number in step S49, the determination is Yes. In step S53, the application threshold value determination unit 71 determines the similarity threshold value set in step S51 or S43 of the “previous loop” as the application threshold value. In other words, the operation for determining the number of logs is repeated by decreasing (decreasing) the similarity threshold, and when the count value of the number of logs exceeds the target number of logs, the similarity threshold is determined (similarity level). The threshold value is the value of the previous loop), and log extraction is performed.

  Also, the process of FIG. 20 is modified. In step S49, if the count value of the number of logs is equal to or less than the target number of logs, the determination is yes. In step S53a, the application threshold value determination unit 71 determines the similarity threshold value set in step S51a or S43a as the application threshold value. In other words, the operation of judging the number of logs by increasing (increasing) the similarity threshold is repeated, and when the count value of the number of logs falls below the target number of logs, the similarity threshold is fixed and log extraction is executed. Is done.

  FIG. 23 shows a screen display corresponding to the operation of the processing target data extraction unit 29. The screen of FIG. 23 is displayed on the computer display of the processing target data extraction unit 29. As shown in FIG. 23, first, an input request screen for x (the number of recommended users) for the basic data experiment is displayed. When the number of recommended users is input, a screen indicating that a basic experiment is being performed is displayed. When the basic experiment is completed, an input request screen for the maximum processing time (setting processing time) is displayed. When the maximum value of the processing time is input, the target number of logs is displayed, and a “process execution” button and a “cancel” button are displayed. When the “cancel” button is displayed, step S71 in FIG. 22 becomes No, and the input request screen for the maximum set time value is displayed again. If the “execute processing” button is pressed, step S71 is Yes and log extraction is performed.

  The configuration of the pattern for inputting the setting processing time has been described above with reference to FIGS. Of course, both the configuration of FIG. 10 and the configuration of FIG. 21 (the configuration of the pattern for inputting the target number of logs and the configuration of the pattern for inputting the set processing time) may be provided.

  FIG. 24 illustrates another configuration example of the processing target data extraction unit 29. In the configuration of FIG. 10 described above, the target log number is input, and the predicted processing time is calculated from the target log number. On the other hand, in the configuration of FIG. 21, a similarity threshold is input, and the target number of logs and the prediction processing time are calculated from the similarity threshold.

  24, the basic data storage unit 101 and the basic data acquisition experiment processing unit 115 are the same as the corresponding configurations in FIG. The similarity threshold value input unit 103 inputs the similarity threshold value described with reference to FIG. Here, the input similarity threshold is referred to as a set similarity threshold. The target log number acquisition unit 105 obtains the target log number corresponding to the set similarity threshold. The processing time calculation unit 107 calculates a predicted processing time corresponding to the target number of logs. The calculation result output unit 109 outputs the target log number obtained by the target log number acquisition unit 105 and the predicted processing time calculated by the processing time calculation unit 107. The permission determination reception unit 111 receives a permission determination for the output target log number and the predicted processing time. The log extraction unit 113 extracts a target number of logs from the logs accumulated in the history data accumulation unit 21 using the set similarity threshold.

  The similarity threshold value input unit 103, the calculation result output unit 109, and the permission determination reception unit 111 are configured by input / output devices such as a keyboard and a display. The input device is operated by the operator, and a desired similarity threshold value is input. This similarity threshold is input to the similarity threshold input unit 103 as a set similarity threshold and is acquired by the target log number acquisition unit 105.

  The target log number acquisition unit 105 obtains the target log number from the similarity threshold according to the principle of the process using the similarity threshold described with reference to FIG. That is, the target log number acquisition unit 105 reads the similarity theoretical value table from the similarity theoretical value table storage unit 117, refers to the similarity theoretical value table, and corresponds to the “threshold value corresponding to the similarity threshold value”. "Reference personal log number" is calculated. Further, the target log number acquisition unit 105 has a function corresponding to the above-described count value acquisition unit, and the log data of the history data storage unit 21 indicates that the comparison target user corresponding to the “threshold reference personal log number” Count the log to get the count value. The number of logs counted becomes the target number of logs.

  The processing of the processing time calculation unit 107 is the same as that in FIG. The basic data is read from the basic data storage unit 101, the target log number is applied to the basic data, and the predicted processing time is calculated.

  The calculation result output unit 109 displays the target number of logs and the predicted processing time on a display or the like. If the target log number and the predicted processing time are appropriate, the operator performs a permission operation on the input device, and the permission determination receiving unit 111 receives a permission determination (consent determination).

  The permission determination is transmitted from the permission determination reception unit 111 to the log extraction unit 113. In addition, the log extraction unit 113 acquires the set similarity threshold value input from the similarity threshold value input unit 103. Then, the log extraction unit 113 extracts a log using the set similarity threshold. That is, the log extraction unit 113 extracts the log of the comparison target user corresponding to the “threshold reference personal log number” from the history data storage unit 21.

  The internal configuration of the log extraction unit 113 may be different from that in FIG. In the configuration of FIG. 18, the similarity threshold is sequentially changed to determine the application threshold. However, here, the threshold value to be applied is already determined. Therefore, the log extraction unit 113 extracts the log using the acquired setting similarity threshold as it is.

  FIG. 25 is a flowchart showing the operation of the processing target data extraction unit 29. The total recommended number of users x is input (S81), the basic data acquisition experiment processing unit performs a basic data acquisition experiment (S83), and processing end prediction graph data that is basic data is generated (S85). Then, the set similarity threshold is input (S87), the target number of logs is obtained (S89), the predicted processing time is calculated (S91), and these are displayed (S93). Whether or not the permission determination is accepted is determined (S95), and if the permission determination is not accepted (S95, No), the process returns to step S87 and another set similarity threshold is input. When the permission determination is accepted (S95, Yes), a log is extracted using the set similarity threshold (S97).

  FIG. 26 shows a screen display corresponding to the operation of the processing target data extraction unit 29. The screen of FIG. 26 is displayed on the computer display of the processing target data extraction unit 29. As shown in FIG. 26, first, an input request screen for x (number of recommended users) for the basic data experiment is displayed. When the number of recommended users is input, a screen indicating that a basic experiment is being performed is displayed. When the basic experiment is completed, the input request screen for the similarity threshold is displayed. When the similarity threshold value is input, the target log number and the predicted processing time are displayed, and a “process execution” button and a “cancel” button are displayed. When the “Cancel” button is displayed, Step S95 in FIG. 25 becomes No, and the similarity threshold value input request screen is displayed again. When the “process execution” button is pressed, step S95 is Yes and log extraction is performed.

  The pattern configuration for inputting the similarity threshold value has been described above with reference to FIGS. Of course, the configuration of FIG. 24 may be provided together with at least one of the configurations of FIGS.

"Split"
(Recommendation target users / logs)
Next, the division process, which is another feature of the present embodiment, will be described. Since resources such as the HDD device and the memory of the filter device are limited, if the number of recommended users is enormous, the processing takes time. Moreover, even if the log is enormous, processing takes time. This point is as described with reference to FIG.

  Therefore, in the present embodiment, the processing time is further reduced by dividing the recommendation target user. Further, the processing time is shortened by dividing the log. The present embodiment makes it possible to perform such division and shorten the processing time by using the above-described simulation of the target log number and processing time.

  For the recommendation target user, a plurality of sets (groups) are created by the division, and a filtering process is performed for each set. If the total processing time of the set is shorter than the processing time when no division is performed, the division is effective.

  As for the log, the comparison target user (similar user) is divided into a plurality of sets, thereby dividing the log. That is, the log of history data is divided so as to be divided in units of similar users.

  FIG. 27 shows a change in processing time when the recommendation target user is divided into a plurality of sets (groups). Here, it is assumed that the recommendation target users are equally divided and the number of persons in each set is the same. In FIG. 27, the horizontal axis represents the recommended number of divided persons r (one set of persons r = total number of users / number of divided areas). The vertical axis represents the processing time t. The log number y is fixed. As shown in the figure, the processing time t becomes shorter as the recommended number of divided persons r becomes smaller. However, if the recommended division number r becomes too small, the processing time t becomes conversely longer. In the optimum recommended division number R, the processing time t is minimized.

  The data in FIG. 27 is obtained by executing the basic experiment already described. In the basic experiment, as described above, the filter device actually operates using the sample log. The number of logs is fixed as shown in FIG. And processing time is measured using the division | segmentation set of a recommendation object user.

  FIG. 28 shows some examples of basic experiment processing when division is applied. FIG. 28A shows a pattern in which no division is performed.

  In FIG. 28B, first, as described with reference to FIG. 27, a change in the processing time t according to the recommended number of divided users r of the user to be recommended is obtained by a basic experiment, and the minimum value of the processing time t is obtained. A corresponding recommended division number a is obtained. Then, the relationship between the number of logs y and the processing time t when the recommendation target user is divided into a plurality of sets of recommended division number a is measured. This measurement result is acquired as basic data.

  In FIG. 28 (c), first, a basic experiment is performed without division. The processing time t when the recommendation target user number x is fixed and the log number y is changed is measured, and the same result as in FIG. 28A is obtained. From the measurement result, the number b of logs corresponding to the inflection point is obtained. Logs are divided so that the number of logs in one set is b. The number of logs is fixed to b, the change in the processing time t according to the recommended division number r of the recommendation target user is measured, and the recommended division number c corresponding to the minimum value of the processing time t is obtained. Then, the relationship between the number of logs y and the processing time t when the recommendation target user is divided into a set of recommended division number c is measured. This measurement result is acquired as basic data. In FIG. 28 (c), basic data when both the log and the recommendation target user are divided is obtained.

Now, the division processing described above is applied to the collaborative filter device as follows. That is, in the processing object data extraction unit 29 shown in FIG. 10, the basic data acquisition experiment processing unit 53 acquires a plurality of types of basic data having different division patterns (division forms) as described below, and the plurality of types of basic data Is stored in the basic data storage unit 41.
(1) When neither the recommendation target user nor the log is divided (2) When the recommendation target user is divided (3) When the log is divided (4) When the recommendation target user and the log are divided (2) to (4) For each, basic data may be obtained from each of the plurality of division patterns and stored. That is, basic data when the number of divisions of recommendation target users is changed or the number of log divisions is changed may be stored.

  In FIG. 10, the processing time calculation unit 45 calculates a predicted processing time from each of a plurality of types of basic data in the basic data storage unit 41. That is, the target number of logs is applied to each basic data. Then, the processing time calculation unit 45 obtains the shortest predicted processing time. The processing time output unit 47 displays the division pattern of the basic data obtained with the predicted processing time together with the shortest predicted processing time. For example, a message “Recommendation target user x is divided by R people” is displayed on the display.

  The log extraction unit 51 acquires information on the division pattern corresponding to the prediction processing time (more specifically, information on the division pattern when the prediction processing time is the shortest) from the processing time calculation unit 45. And the log extraction part 51 performs a log extraction process based on the acquired division | segmentation pattern. That is, when the logs of the history data storage unit 21 are divided, the target number of logs is extracted from the divided logs. Again, the process using the above-described similarity threshold is performed in the same manner.

  The data of the division pattern is further transmitted from the processing target data extraction unit 29 to the collaborative filter unit 23. The collaborative filter unit 23 performs a collaborative filter operation in accordance with the acquired division pattern data. If the recommendation target user is divided by the division pattern, the recommendation target is divided according to the division pattern and the filtering process is performed. Further, if the log is divided according to the division pattern, the filtering process is performed using the log extraction data from the division log provided by the log extraction unit 51.

  The above division pattern is data on how the recommendation target user is divided and how the log is divided. Specifically, for example, the number of sets and the number of sets in the division of the recommendation target user, and the number of logs in one set in the log division.

  The above division processing may be similarly applied to the processing target data extraction unit 29 shown in FIG. The configuration of FIG. 21 is a configuration in which the target log number is calculated by inputting the set processing time. Also in this case, a plurality of types of basic data corresponding to a plurality of division patterns are acquired by the basic data acquisition experiment processing unit 93 and stored in the basic data storage unit 81.

  The target log number calculation unit 85 calculates the target log number from each of a plurality of types of basic data in the basic data storage unit 81. Then, the target log number calculation unit 85 obtains the largest target log number. The target log number output unit 87 displays the division pattern of the basic data from which the target log number is obtained together with the target log number. The log extraction unit 81 acquires information on the division pattern of the basic data that maximizes the target log number from the target log number calculation unit 85. Then, the log extraction unit 81 performs log extraction processing based on the acquired division pattern. The collaborative filter unit 23 also acquires the division pattern and performs collaborative filtering with the pattern.

  Furthermore, the above-described division processing may be applied to the processing target data extraction unit 29 shown in FIG. The configuration of FIG. 24 is a configuration in which the target log number and the prediction processing time are calculated by inputting a set similarity threshold. Also in this case, the plurality of types of basic data are acquired by the basic data acquisition experiment processing unit 115 and stored in the basic data storage unit 101.

  The target log number acquisition unit 105 obtains the number of logs corresponding to the set similarity threshold as described with reference to FIG. Thereafter, the processing time calculation unit 107 calculates a predicted processing time from each of a plurality of types of basic data in the basic data storage unit 101. Then, the processing time calculation unit 107 obtains the shortest predicted processing time. The calculation result output unit 109 displays the division pattern of the basic data obtained with the prediction processing time together with the prediction processing time. The log extraction unit 113 acquires information on the division pattern when the predicted processing time is the shortest from the processing time calculation unit 107. Then, the log extraction unit 113 performs log extraction processing based on the acquired division pattern. The collaborative filter unit 23 also acquires the division pattern and performs collaborative filtering with the pattern.

  The division process has been described above. In the above description, the basic data storage unit includes (1) basic data at the time of non-division, (2) basic data when the recommendation target user is divided, and (3) basic data when the log is divided, (4) The basic data when both are divided is stored. However, not all these basic data need be stored. Only one may be stored. For example, only basic data obtained by dividing recommendation target data may be stored (however, a plurality of basic data in which the number of divided persons is changed may be stored).

  Moreover, in said example, the recommendation object user was divided | segmented at random and the number of people of each set was the same. The division pattern of the user to be recommended is not limited to this. For example, the number of logs (number of personal logs) of each recommendation target user is checked. Based on the above-described theoretical value of the Jackard coefficient, the theoretical value is fixed and the users are divided. Specifically, the recommendation target users are divided so that users with a close personal log count belong to the same set. A user may be divided for each number of personal logs.

  The collaborative filter device 1 according to the embodiment of the present invention has been described above. According to the present embodiment, the basic data storage unit 41 that stores the basic data of the relationship between the number of logs and the collaborative filtering processing time is provided, the predicted processing time corresponding to the input target log number is calculated, and , Logs of the input target number of logs are extracted from the history data storage unit 21. Since the prediction processing time is known in advance, the number of logs can be set so that the prediction processing time is appropriate. In addition, by extracting logs of the set number of logs from the history data, the amount of calculation can be reduced so as to achieve the calculated predicted processing time. In this way, the amount of calculation can be reduced and the processing time can be shortened.

  In addition, according to the present embodiment, based on the similarity theoretical value, the number of logs of the comparison target user is limited according to the number of logs of the recommendation target user, and the log is extracted from the history data, so that the recommendation target It is possible to leave a log of a user who has a similar number of items to the user and to reduce the number of logs while maintaining higher filtering accuracy.

  Further, according to the present embodiment, the similarity threshold value of the similarity theory value is set, and the log is extracted from the history data using the similarity threshold value, so that the recommended target user and the number of items to be acquired Logs of users close to can be left, and the number of logs can be reduced while maintaining higher filtering accuracy.

  Further, according to the present embodiment, by dividing the recommendation target user into a plurality of pieces, the calculation amount can be reduced and the processing time can be shortened. Also at this time, the processing can be executed after appropriately setting the processing time and the number of logs.

  Further, according to the present embodiment, by dividing the log into a plurality of logs, the calculation amount can be reduced and the processing time can be shortened. Also at this time, the processing can be executed after appropriately setting the processing time and the number of logs.

  Further, according to the present embodiment, by dividing the recommendation target user into a plurality of parts and dividing the log into a plurality of parts, the calculation amount can be reduced and the processing time can be shortened. Also at this time, the processing can be executed after appropriately setting the processing time and the number of logs.

  In addition, according to the present embodiment, the target number of logs corresponding to the input setting processing time is calculated in response to the input of the setting processing time, and logs of the calculated target log number are extracted from the history data storage unit. Is done. Therefore, also in this respect, it is possible to suitably reduce the amount of calculation and shorten the processing time.

  Further, according to the present embodiment, the target log number is calculated from the input set similarity threshold, and the predicted value of the processing time is calculated from the target log number. Therefore, the log can be reduced so that the processing time is appropriate by adjusting the set similarity threshold. By using the similarity threshold value of the similarity theory value, it is possible to reduce the amount of calculation while keeping the filtering accuracy higher, and to shorten the processing time.

In the present embodiment, the description has been made on the assumption that the product is content. However, the content database may be configured with information describing the product as content. Further, the content may not be electronic data.
Further, although it has been described that the content purchase is downloaded, the history data may be obtained from any means as long as it associates the purchase of the product with the user. For example, the order may be received by facsimile or telephone without using the Internet.

  The preferred embodiments of the present invention have been described above. However, the present invention is not limited to the above-described embodiments, and it goes without saying that those skilled in the art can modify the above-described embodiments within the scope of the present invention.

  As described above, the collaborative filter device according to the present invention has an effect of reducing the amount of calculation and shortening the processing time, and is useful as a content distribution recommendation system.

The block diagram which shows the structure of the collaborative filter apparatus which concerns on embodiment of this invention Flow chart of overall operation of content purchase and history data accumulation The figure which shows the example of log which is history data The figure which shows the collaborative filtering processing which utilizes the Jackard coefficient The figure which shows the collaborative filtering processing which utilizes the Jackard coefficient Flow chart showing collaborative filtering processing The figure which shows the collaborative filtering process using similarity of a cosine The figure which shows the collaborative filtering process using similarity of a cosine The figure which shows the collaborative filtering process using similarity of a cosine The figure which shows the structure of the processing object data extraction part (A) Diagram showing ideal relationship between number of logs and processing time (b) Diagram showing actual relationship between number of logs and processing time The figure which shows the processing of basic data acquisition experiment processing section A flowchart showing the operation of the processing target data extraction unit The figure which shows the screen display corresponding to operation | movement of a process target data extraction part. (A) Diagram showing an example of purchase history (b) Diagram showing another example of purchase history Figure showing a table of similarity theory values Figure showing a table of similarity theory values Block diagram showing the configuration of the log extractor Flow chart showing operation of log extractor Flow chart showing operation of log extractor The figure which shows another structural example of a process target data extraction part A flowchart showing the operation of the processing target data extraction unit The figure which shows the screen display corresponding to operation | movement of a process target data extraction part. The figure which shows another structural example of a process target data extraction part A flowchart showing the operation of the processing target data extraction unit The figure which shows the screen display corresponding to operation | movement of a process target data extraction part. The figure which shows the change of processing time when a user for recommendation is divided into a plurality of sets (A) The figure which shows the basic data of the pattern which does not divide | segment (b) The figure which shows the basic data of the pattern which divides | segments (c) The figure which shows the basic data of the pattern which divides | segments

Explanation of symbols

DESCRIPTION OF SYMBOLS 1 Collaborative filter apparatus 21 History data storage part 23 Collaborative filter part 25 Recommendation information delivery part 27 Merchandise information memory | storage part 29 Process target data extraction part 41 Basic data memory | storage part 43 Target log number input part 45 Processing time calculation part 47 Processing time output part 49 Permission judgment acceptance part 51 Log extraction part 53 Basic data acquisition experiment processing part

Claims (14)

A history data accumulation unit that accumulates, as history data, a log including user identification data and item identification data generated each time an item is acquired by a large number of users;
A collaborative filter unit for obtaining a recommended item for a recommendation target user by collaborative filtering processing using the history data accumulated in the history data accumulation unit;
An output unit for outputting a processing result by the collaborative filter unit;
A processing target data extraction unit that extracts logs to be processed by the collaborative filter unit from a large number of logs stored in the history data storage unit;
The processing object data extraction unit
A basic data storage unit storing basic data on the relationship between the number of logs and the processing time of collaborative filtering;
A target log number input unit for inputting a target log number to be processed by collaborative filtering;
A processing time calculation unit that reads out the basic data from the basic data storage unit and obtains a predicted processing time corresponding to the target log number input by the log number input unit from the basic data;
A processing time output unit that outputs the predicted processing time calculated by the processing time calculation unit;
Based on the target log number input by the log number input unit, a log extraction unit that extracts logs of the target log number from logs accumulated in the history data accumulation unit;
A collaborative filter device comprising:   The log extraction unit calculates the number of personal logs of the recommendation target user based on the similarity similarity value determined from the number of personal logs of the recommendation target user, the number of personal logs of the comparison target user, and the number of logs of the common acquisition items of both users. The collaborative filter according to claim 1, wherein the number of personal logs of the comparison target user is limited in response, and the logs of the comparison target user corresponding to the limited number of personal logs are extracted from the history database. apparatus. The log extraction unit
Means for setting a similarity threshold of the similarity theoretical value;
Means for setting a threshold reference personal log number that is the number of personal logs of a comparison target user corresponding to a similarity theoretical value equal to or greater than the similarity threshold;
Count acquisition means for acquiring a count value of the number of logs of comparison target users corresponding to the threshold standard personal log number from the log of the history data storage unit;
And the log extraction source to the threshold reference personal log number when the similarity threshold is set to the maximum in a range in which the count value obtained by the count acquisition means achieves the target log number The collaborative filter device according to claim 2, wherein the number of personal logs of the comparison target users is limited. The basic data storage unit stores basic data on the relationship between the number of logs and the collaborative filtering processing time when the recommended user is divided into a plurality,
The collaborative filter device according to claim 1, wherein the processing time calculation unit calculates a prediction processing time using the basic data obtained when the recommendation target user is divided. The basic data storage unit stores basic data on the relationship between the number of logs and collaborative filtering processing time when the log is divided into a plurality of logs,
The collaborative filter device according to claim 1, wherein the processing time calculation unit calculates a predicted processing time using the basic data when the log is divided. The basic data storage unit stores basic data on the relationship between the number of logs and the collaborative filtering processing time when the recommended user and the log are each divided into a plurality of pieces,
The collaborative filter device according to claim 1, wherein the processing time calculation unit calculates a predicted processing time using the basic data when the pre-recommendation target user and the log are divided. The processing target data extraction unit further includes a setting processing time input unit that inputs a setting processing time of collaborative filtering, and the setting that is input by the processing time input unit by reading the basic data from the basic data storage unit A target log number calculation unit for obtaining a target log number corresponding to the processing time from the basic data,
The log extraction unit, when the set processing time is input, extracts the log of the target log number obtained by the target log number calculation unit from the logs accumulated in the history data accumulation unit. The collaborative filter device according to claim 1. The processing target data extraction unit inputs a similarity threshold value for setting a similarity degree theoretical value determined from the number of personal logs of a recommendation target user, the number of personal logs of a comparison target user, and the number of logs of a common acquisition item of both users A threshold value input unit, a target log number acquisition unit for obtaining a target log number corresponding to the set similarity threshold, and a processing time calculation unit for calculating a predicted processing time corresponding to the target log number. And
When the set similarity threshold is input, the target log number acquisition unit compares the number of individual logs corresponding to the similarity theoretical value equal to or higher than the similarity threshold from the log of the history data storage unit. The count value of the number of logs of the target user is obtained as the target log number, and the processing time calculation unit reads the basic data from the basic data storage unit, and the target log acquired by the target log number acquisition unit The collaborative filter device according to claim 1, wherein a prediction processing time corresponding to a number is obtained from the basic data. Using a history data storage unit that stores, as history data, a log including user identification data and item identification data that occurs each time an item is acquired by a large number of users, and the history data stored in the history data storage unit For a collaborative filter device comprising a collaborative filter unit that seeks a recommended item for a recommendation target user by collaborative filtering processing, the collaborative filter unit should be processed from a large number of logs accumulated in the history data accumulating unit A collaborative filter support device for extracting logs,
A basic data storage unit storing basic data on the relationship between the number of logs and the processing time of collaborative filtering;
A target log number input unit for inputting a target log number to be processed by collaborative filtering;
A processing time calculation unit that reads out the basic data from the basic data storage unit and obtains a predicted processing time corresponding to the target log number input by the log number input unit from the basic data;
A processing time output unit that outputs the predicted processing time calculated by the processing time calculation unit;
Based on the target log number input by the log number input unit, a log extraction unit that extracts logs of the target log number from logs accumulated in the history data accumulation unit;
A collaborative filter support device characterized by comprising: Using a history data storage unit that stores, as history data, a log including user identification data and item identification data that occurs each time an item is acquired by a large number of users, and the history data stored in the history data storage unit For a collaborative filter device comprising a collaborative filter unit that seeks a recommended item for a recommendation target user by collaborative filtering processing, the collaborative filter unit should be processed from a large number of logs accumulated in the history data accumulating unit A collaborative filter support processing method for extracting logs by computer processing,
The target log number to be processed by collaborative filtering is input from the target log number input unit, the basic data is read from the basic data storage unit storing the basic data of the relationship between the log number and the processing time of collaborative filtering, and the basic data A prediction processing time corresponding to the target log number input by the log number input unit is obtained from the basic data read from the storage unit, and the calculated prediction processing time is output from the processing time output unit, and the number of logs A collaborative filter support processing method, wherein a log having the target number of logs is extracted from logs accumulated in the history data accumulation unit based on the target number of logs input by an input unit. Using a history data storage unit that stores, as history data, a log including user identification data and item identification data that occurs each time an item is acquired by a large number of users, and the history data stored in the history data storage unit For a collaborative filter device comprising a collaborative filter unit that seeks a recommended item for a recommendation target user by collaborative filtering processing, the collaborative filter unit should be processed from a large number of logs accumulated in the history data accumulating unit A collaborative filter support processing program for causing a computer to execute processing for extracting a log,
The target log number to be processed by collaborative filtering is input from the target log number input unit, the basic data is read from the basic data storage unit storing the basic data of the relationship between the log number and the processing time of collaborative filtering, and the basic data A prediction processing time corresponding to the target log number input by the log number input unit is obtained from the basic data read from the storage unit, and the calculated prediction processing time is output from the processing time output unit, and the number of logs A collaborative filter support processing program for causing a computer to execute a process of extracting a log of the target log number from logs accumulated in the history data accumulation unit based on the target log number input by an input unit. Using a history data storage unit that stores, as history data, a log including user identification data and item identification data that occurs each time an item is acquired by a large number of users, and the history data stored in the history data storage unit For a collaborative filter device comprising a collaborative filter unit that seeks a recommended item for a recommendation target user by collaborative filtering processing, the collaborative filter unit should be processed from a large number of logs accumulated in the history data accumulating unit A collaborative filter support device for extracting logs,
A basic data storage unit storing basic data on the relationship between the number of logs and the processing time of collaborative filtering;
A setting processing time input unit for inputting a setting processing time of collaborative filtering;
A target log number calculation unit that reads the basic data from the basic data storage unit and obtains a target log number corresponding to the setting processing time input by the setting processing time input unit from the basic data;
A target log number output unit for outputting the target log number calculated by the target log number calculation unit;
Based on the target log number calculated by the target log number calculation unit, a log extraction unit that extracts logs of the target log number from logs accumulated in the history data accumulation unit;
A collaborative filter support device characterized by comprising: Using a history data storage unit that stores, as history data, a log including user identification data and item identification data that occurs each time an item is acquired by a large number of users, and the history data stored in the history data storage unit For a collaborative filter device comprising a collaborative filter unit that seeks a recommended item for a recommendation target user by collaborative filtering processing, the collaborative filter unit should be processed from a large number of logs accumulated in the history data accumulating unit A collaborative filter support processing program for causing a computer to execute processing for extracting a log,
Input the collaborative filtering setting processing time from the setting processing time input unit, read the basic data from the basic data storage unit storing the basic data of the relationship between the number of logs and the processing time of collaborative filtering, and read from the basic data storage unit The target log number corresponding to the set processing time input by the set processing time input unit is obtained from the basic data, the calculated target log number is output from the target log number output unit, and the target log number calculation is performed. A collaborative filter support processing program for causing a computer to execute a process of extracting a log of the target log number from the logs accumulated in the history data accumulation unit based on the target log number calculated by a unit. Using a history data storage unit that stores, as history data, a log including user identification data and item identification data that occurs each time an item is acquired by a large number of users, and the history data stored in the history data storage unit For a collaborative filter device comprising a collaborative filter unit that seeks a recommended item for a recommendation target user by collaborative filtering processing, the collaborative filter unit should be processed from a large number of logs accumulated in the history data accumulating unit A collaborative filter support device for extracting logs,
Means for inputting a similarity threshold of a theoretical value of similarity determined from the number of personal logs of recommended users, the number of personal logs of users to be compared, and the number of logs of common acquisition items of both users;
Means for setting a threshold reference personal log number that is the number of personal logs of a comparison target user corresponding to a similarity theoretical value equal to or greater than the similarity threshold;
Means for extracting the log of the comparison target user of the threshold reference personal log number from the history data storage unit;
A collaborative filter support device characterized by comprising:

Images