JP5732441B2 - Information recommendation method, apparatus and program - Google Patents

Description

  The present invention relates to an information recommendation method, an apparatus, and a program for recommending information according to a user's preference and situation.

  In recent years, various services for recommending information in accordance with user preferences and situations have been proposed. For example, a restaurant guide service is provided in which peripheral restaurant information is selected from search conditions such as preference information set by the user and information on the current location of the user and presented to the user. However, manually inputting search conditions such as preference information set by the user himself / herself in order to acquire appropriate information imposes an excessive burden on the user as the amount of information on the Internet becomes more detailed. There is a problem that cannot reflect the potential taste as well.

  In addition, information searches using mobile terminals such as smart phones are increasing due to the spread of mobile Internet, but mobile terminals have a small amount of information that can be displayed on the screen at one time, and it takes time to input search conditions. Therefore, there is a growing need for a technique for recommending appropriate information in consideration of user preferences.

  Thus, various techniques for recommending appropriate information by analyzing a user's Internet browsing history and the like have been studied. However, in the conventional technique, there arises a problem of so-called Cold-start in which the recommendation accuracy is not improved until the user history is accumulated more than the first amount. Due to this problem, since the user cannot make an appropriate recommendation at the beginning of service use and the satisfaction level does not increase, the history is not accumulated because it is not continuously used. As a result, the improvement of recommendation accuracy will be prolonged and the service will not spread.

  On the other hand, information recommendation technology based on browsing history and search history is generally called a memory-based method, and is roughly divided into collaborative filtering and content-based filtering. Of these two methods, when a tag indicating a category or the like can be attached to an item, content-based filtering is said to be strong against the Cold-start problem. For this reason, a method of switching or synthesizing these two techniques according to the history accumulation status or the like has been studied (for example, see Non-Patent Document 1).

R. Burke, "Hybrid recommender systems: surveyand, experiments, user-modeling and user-adapted inter-actions," User Modeling and User-Adapted Interaction, Vol. 12, No. 4, pp.331-370, 2002.

  However, the restaurant recommendation system described in Non-Patent Document 1 has the following problems to be improved. That is, in this system, the user manually inputs desired conditions such as price range and gourmet genre at the first stage, and selects candidate restaurants using content-based filtering. The candidates are ranked by collaborative filtering. For this reason, a fixed recommendation accuracy can be realized at an early stage. However, the desired conditions of the restaurant for narrowing down candidates are dependent on user manual input. For this reason, in order to improve the recommendation accuracy when there are many candidates, it is necessary to accumulate history for ranking. That is, this method cannot be expected to have a sufficient effect when recommending information such as restaurants, where there are many candidates but there is little history accumulation.

  The present invention has been made paying attention to the above circumstances, and its purpose is to make it possible to recommend information with higher accuracy without relying on manual input of conditions by the user even in situations where the user's browsing history is low. It is to provide an information recommendation method, apparatus and program.

In order to achieve the above object, a first aspect of the present invention is to analyze a user's interest in a plurality of geographically dispersed spots by a computer, and to recommend spots recommended to the user based on the analysis result. In an information recommendation method, apparatus and program for creating information,
Information that represents the user's browsing history for the browsing site where the computer posts information on the plurality of spots, stores the information in a first storage medium, and information that represents the geographical movement history of the user Is acquired and stored in the second storage medium. In addition, based on the browsing history information stored in the first storage medium, the computer clusters a plurality of spot browsing information included in the browsing history information on the condition that the browsing time is close, thereby browsing chunks. And a reference area that the user can visit is determined based on the movement history information stored in the second storage medium or information input in advance by the user or connection information with the radio base station. A set of spot information included in the determined reference area is extracted as a reference spot set from spot information stored in advance. And the number a of the spot browsing information included in the generated browsing chunk, the number n of the spot browsing information having a predetermined concept among the spot browsing information group included in the browsing chunk, and included in the reference spot set The number of spot browsing information in which the concept appears in the browsing chunk based on the number S of spot information and the number N of spot information to which the concept is given out of the spot information group included in the reference spot set Calculates a first probability that is equal to or greater than n and a second probability that is less than n, and an inverse function of the cumulative distribution function of the standard normal distribution based on the calculated first and second probabilities. by, it is characterized in Rukoto issuing calculation interest score z representing the degree of interest of the user with respect to the concepts.

The first aspect of the present invention is characterized by comprising the following aspects.
In the first aspect, the computer clusters a plurality of pieces of stay information included in the movement history information on the basis of distance proximity based on the movement history information stored in the second storage medium. The function further includes a function of generating a stay chunk and associating a spot group with the stay chunk on condition that the distance is close. The number of pieces of spot information included in the generated stay chunk is set to a, and the stay is stored. The interest score z is calculated by setting the number of pieces of spot information having a predetermined concept in the spot information group included in the chunk as n.

In the second aspect, the computer clusters a plurality of pieces of stay information included in the movement history information on the basis of the proximity of the distance based on the movement history information stored in the second storage medium. A function of generating a stay chunk and associating a spot group with the stay chunk on condition that the distance is close is further provided, and the number of pieces of spot browse information appearing in the browse chunk is N, and The number of spot groups associated with stay chunks in the co-occurrence chunk as a co-occurrence chunk as a co-occurrence chunk on the condition that the distance is close is defined as a. , among the group spots associated with the stay chunks within the predetermined time, the number of spots that appeared in the viewing chunk is n, the interest scores z It is intended to leave.

According to a third aspect, the computer clusters a plurality of pieces of stay information included in the movement history information based on the proximity of the distance based on the movement history information stored in the second storage medium. A function of generating a stay chunk and associating a spot group with this stay chunk on condition of closeness is provided, and the number S of spot information included in the reference spot set is defined as S. The number of spot information appearing in the stay chunk group existing within a predetermined time set in advance from the browsing time of the browsing chunk is N, and the spot information appearing in the browsing chunk is within a certain time. The spot information group that co-occurs with the spot information that appears in the stay chunk having the stay start time is defined as a co-occurrence chunk, and the spots in the co-occurrence chunk are identified. And it calculates the interest scores z as n.

According to a fourth aspect, the computer clusters a plurality of pieces of stay information included in the movement history information based on the proximity of the distance based on the movement history information stored in the second storage medium. A function of generating a stay chunk and associating a spot group with this stay chunk on condition of closeness is provided, and the number S of spot information included in the reference spot set is defined as S. In the spot information group appearing in the stay chunk, N is the number of spot information appearing in the browse chunk group whose browsing time is within a predetermined time from the stay start time of the stay chunk. A spot information group that co-occurs with spot information appearing in a browsing chunk within a certain time is defined as a co-occurrence chunk, and the number of spots in the co-occurrence chunk is n. Te and calculates the interest score z.

In a fifth aspect, for the browsing chunk, stay chunk, or co-occurrence chunk, the interest score according to at least one of a user operation history, a time and situation of each chunk, and a degree of association between the spot and the concept z is weighted.

  In the sixth aspect, the first probability and the second probability are calculated by hypergeometric distribution, binomial distribution or normal distribution.

The seventh aspect further includes a function of calculating the user's evaluation score for the spot information using the interest score z of each concept that appears in the spot information for the spot information in which one or more concepts appear. It is a thing.

In the eighth aspect, the interest score z is weighted according to at least one of the degree of association between the spot information and each concept and the importance of each concept.

  That is, according to the present invention, it is possible to provide an information recommendation method, apparatus, and program that enable highly accurate information recommendation without relying on manual input of conditions by a user even in a situation where the user's browsing history record is small. it can.

The figure which shows the whole structure of the system provided with the information recommendation apparatus which concerns on 1st Embodiment of this invention. The block diagram which shows the function structure of the preference learning process part in the information recommendation apparatus shown in FIG. The block diagram which shows the function structure of the restaurant recommendation process part in the information recommendation apparatus shown in FIG. The flowchart which shows the whole process sequence and process content by the information recommendation apparatus shown in FIG. The flowchart which shows the process sequence and process content of a preference learning process among the whole processes shown in FIG. The flowchart which shows the process sequence and process content of a restaurant recommendation process among the whole processes shown in FIG. The figure for demonstrating the process which extracts a browsing chunk and a stay chunk in the preference learning process shown in FIG. The figure for demonstrating the process which calculates the deviation probability of the browsing chunk and stay chunk with respect to a reference restaurant set in the preference learning process shown in FIG. The figure for demonstrating the process which calculates the deviation probability of the co-occurrence chunk with respect to a reference restaurant set in the preference learning process shown in FIG. The figure which shows the evaluation result of the recommendation precision by the information recommendation apparatus shown in FIG.

Embodiments according to the present invention will be described below with reference to the drawings.
(First embodiment)
[Constitution]
FIG. 1 is a schematic configuration diagram of a system according to a first embodiment of the present invention, in which user terminals UT1 to UTn possessed by respective users are connected to an information server SV having a function as an information recommendation device via a communication network NW. It is possible.
The communication network NW includes the Internet and an access network for accessing the Internet. As the access network, a mobile communication network such as a mobile phone network, Hotspot (registered trademark), WiMAX (registered trademark), or wireless LAN is used.

  The user terminals UT1 to UTn are, for example, mobile terminals such as mobile phones, smart phones, PDAs (Personal Digital Assistants), tablet terminals, notebook personal computers, and the like, and use the GPS (Global Positioning System). A function for measuring the position of the user, a browser function for accessing a Web site, and a mailer function for sending and receiving electronic mail.

The information server SV is composed of, for example, a Web server, and includes a communication interface unit (not shown), a control unit 1, and a storage unit 2.
The communication interface unit performs information communication with the user terminals UT1 to UTn via the communication network NW. For example, TCP / IP (Transmission Control Protocol / Internet Protocol) is used as the communication protocol.

  The storage unit 2 uses a non-volatile memory that can be written and read as needed, such as an HDD or a NAND flash memory, as a storage medium. As a storage function necessary to implement this embodiment, a GPS movement history storage unit 21, a detailed browsing history storage unit 22, a preference database 23, a restaurant database 24, and a recommendation list storage unit 25.

  The GPS movement history storage unit 21 is used for accumulating user's GPS movement history data periodically collected from the user terminals UT1 to UTn. This GPS movement history data consists of a set of positioning data representing the user's place of stay, and each positioning data consists of positioning time and positioning coordinates.

  The detailed browsing history storage unit 22 is used for accumulating a set of data representing a browsing history with respect to a restaurant URL (Uniform Resource Locator) transmitted from the user terminals UT1 to UTn. Each browsing history data is represented by browsing time and store ID.

  A preference database (Preference DB) 23 is used to store preference data calculated by a preference learning processing unit 13 described later. Each preference data includes a tag and a preference score.

  In the restaurant database 24, information representing restaurants existing in the service area of the system is stored in advance. The information representing the restaurant includes a “store ID”, a “summary” representing a summary of the store, a “detail” representing a menu, a price range, etc., a “URL” which is link information to the store website, It is composed of “address coordinates” and “tag” of the store represented by movement / longitude. “Tag” indicates attribute information such as the gourmet genre of the restaurant and the presence / absence of a parking lot.

  The recommendation list storage unit 25 is used for storing a restaurant recommendation list calculated by the restaurant recommendation processing unit 14 described later. The recommendation list (Recommend List) is a collection of restaurant data serving as recommendation candidates, sorted in order of preference score, and includes a recommendation rank, a store ID, an outline, details, and a URL.

  The control unit 1 includes a central processing unit (CPU). As a control function necessary to implement this embodiment, a GPS movement history collection unit 11, a detailed browsing history collection unit 12, The preference learning processing unit 13 and the restaurant recommendation processing unit 14 are provided.

  The GPS movement history collection unit 11 periodically collects GPS movement history data measured at the user terminals UT1 to UTn from the user terminals UT1 to UTn via the communication network NW, and collects the collected GPS movement history data to the GPS. It has a function to be stored in the movement history storage unit 21.

  The detailed browsing history collection unit 12 collects the access history of the user terminals UT1 to UTn with respect to the URL of the restaurant, that is, detailed browsing history data, from the user terminals UT1 to UTn via the communication network NW. The collected browsing history data is stored in the GPS movement history storage unit 21.

  The preference learning processing unit 13 stores the GPS movement history data of the user stored in the GPS movement history storage unit 21, the detailed browsing history data stored in the detailed browsing history storage unit 22, and the restaurant database 24. Based on the stored restaurant data, a user's preference score for each restaurant is calculated, and the calculated preference score is stored in the preference database 23. Specifically, the configuration is as follows. Is done. FIG. 2 is a block diagram showing the functional configuration.

  That is, the preference learning processing unit 13 includes a reference restaurant set creation unit 131, a detailed browsing history analysis unit 132, a GPS movement history analysis unit 133, a restaurant co-occurrence analysis unit 134, and a preference score calculation unit 135. Yes.

  The reference restaurant set creation unit 131 analyzes the user's action range based on the GPS movement history data stored in the GPS movement history storage unit 21 for each user, and the restaurants that exist within this action range. Is extracted from the restaurant database 24 to create a reference restaurant set.

  For each user, the detailed browsing history analysis unit 132 creates the reference restaurant set creation for the total number of tags held by each store for each store that appears in the detailed browsing history data accumulated in the detailed browsing history storage unit 22. It has a function of calculating a deviation probability from the case of random extraction from the reference restaurant set created by the unit 131. In order to calculate a divergence probability in consideration of integration with a GPS movement history described later, a group of detailed browsing is extracted from the detailed browsing history based on the proximity of the browsing time. This group is called a browsing chunk.

  The GPS movement history analysis unit 133 extracts, for each user, a GPS movement history in which the user stays within a certain range for a certain period of time based on the GPS movement history data stored in the GPS movement history storage unit 21. . This extracted GPS movement history is referred to as a stay chunk. Then, the visit possibility restaurant in the extracted stay chunk is considered as a sample from the reference restaurant set, and has a function of calculating a deviation probability from the random extraction in the tag stay chunk.

  The restaurant co-occurrence analysis unit 134 reflects the possibility that the user has actually visited the restaurant after the detailed browsing in order to reflect in the analysis the co-occurrence of the restaurant that appeared in the browsing chunk and the restaurant that can be visited in the stay chunk As in the analysis to each history so far, it has a function of calculating the deviation probability from the case of random extraction.

  The preference score calculation unit 135 integrates the divergence probabilities calculated by the detailed browsing history analysis unit 132, the GPS movement history analysis unit 133, and the restaurant co-occurrence analysis unit 134, thereby calculating the user's preference score for each tag. A function of calculating and storing the calculated preference score in the preference database 23;

  The restaurant recommendation processing unit 14 creates a restaurant set serving as a recommendation candidate in response to a user request, calculates a score for each restaurant included in the candidate restaurant set, creates a recommendation list, The recommendation list is returned to the requesting user terminals UT1 to UTn. Specifically, the recommendation list is configured as follows. FIG. 3 is a block diagram showing the functional configuration.

That is, the restaurant recommendation processing unit 14 includes a candidate restaurant set creation unit 141, a restaurant score calculation unit 142, and a recommendation list creation unit 143.
The candidate restaurant set creation unit 141, when a user request for requesting a restaurant is sent from the user terminals UT1 to UTn, the restaurant corresponding to the store condition represented by the region and the tag included in the user request. It has a function of retrieving a set of shops from the restaurant database 24 and thereby creating a candidate restaurant set.

  The restaurant score calculation unit 142 has a function of calculating the score of each restaurant included in the candidate restaurant set based on the user's preference score for the tag stored in the preference database 23. This restaurant score is information representing the degree of coincidence with the user's preference score.

The recommendation list creation unit 143 sorts the restaurant information in descending order based on the calculated score of each restaurant, thereby rearranging the restaurant information in descending order of the degree of coincidence with the user preference score. Create a recommended list. The prepared recommendation list is stored in the recommendation list storage unit 25 and returned to the requesting user terminals UT1 to UTn via the communication network NW.
Each processing function of the control unit 1 described above is realized by causing the CPU to execute an application program stored in a program memory (not shown).

[Operation]
Next, an operation of information recommendation processing by the information recommendation server SV configured as described above will be described. FIG. 4 is a flowchart showing the overall processing procedure and processing contents.
(1) Collection processing of GPS movement history and browsing history The control unit 1 of the information recommendation server SV monitors the collection timing of the GPS movement history and the collection timing of the detailed browsing history at step S1 and step 3, respectively. In this state, for example, when the GPS movement history collection timing set at a certain time interval comes, the GPS movement history collection unit 11 is activated in step S2, and each user is controlled under the control of the GPS movement history collection unit 11. A GPS movement history collection request is sequentially transmitted to the terminals UT1 to UTn.

  On the other hand, each user terminal UT1 to UTn stores a set of GPS positioning data measured during the period from the previous collection timing to the current collection timing, and when receiving the collection request, each user terminal UT1 to UTn. Reads out the stored GPS positioning data set and transmits it to the information recommendation server SV.

  The control unit 1 of the information recommendation server SV receives a set of GPS positioning data sent from each of the user terminals UT1 to UTn under the control of the GPS movement history collection unit 11, and receives the received GPS positioning data. The set is stored in the GPS movement history storage unit 21 as GPS movement history data.

  On the other hand, when the detailed browsing history collection timing comes, the detailed browsing history collecting unit 12 is activated in step S4, and the GPS movement history is collected under the control of the detailed browsing history collecting unit 12 as described above. In addition, a detailed browsing history collection request is sequentially transmitted to each of the user terminals UT1 to UTn.

  On the other hand, in each user terminal UT1 to UTn, the URL and browsing time of the restaurant that the user browsed to confirm the detailed information of the restaurant during the period from the previous collection timing to the current collection timing are stored. ing. When receiving the collection request, each of the user terminals UT1 to UTn reads the browsing history data including the stored restaurant URL and browsing time, and transmits the reading history data to the information recommendation server SV.

  The control unit 1 of the information recommendation server SV receives the browsing history data transmitted from each of the user terminals UT1 to UTn under the control of the detailed browsing history collection unit 12, and performs detailed browsing of the received browsing history data. Store in the history storage unit 22.

(2) Preference learning process (Learning Process)
The control unit 1 of the information recommendation server SV monitors whether or not the preference learning timing has come in step S5 while periodically collecting the GPS movement history data and the detailed browsing history data. When the predetermined processing timing is reached in this state, the preference learning processing unit 13 is activated in step S6, and the user's preference for the restaurant is learned as described below under the control of the preference learning processing unit 13. FIG. 5 is a flowchart showing the processing procedure and processing contents.

(2-1) Creation of reference restaurant set The control unit 1 of the information recommendation server SV first activates the reference restaurant set creation unit 131 in step S61. Then, under the control of the reference restaurant set creation unit 131, the GPS movement history data related to the learning target user is read from the GPS movement history storage unit 21 in step S <b> 611. Next, in step S612, a reference area representing the user's action range is obtained based on the read GPS movement history data, and restaurants existing in the reference area are read from the restaurant database 24 in step S613. The reference restaurant set. Note that the reference area can be determined based on information collected by manual input based on the user's range of action, residence, work location questionnaire, profile information, and the like. Further, other than the GPS movement history, it can be determined based on the connection history to the radio base station whose position information is known in advance, the position positioning information based on the radio wave intensity from the radio base station, and the like.

  The following method can be considered as a specific processing method for creating the reference restaurant set. That is, first, the center of gravity of all positioning points is calculated for the user's GPS movement history, and the distance including 95% of each positioning point is calculated from this center of gravity. The reason for setting it to 95% is to consider outliers due to travel due to travel or the like. Next, a circle area centered on the calculated center of gravity is drawn with a constant multiple of this distance as a radius, and restaurants existing in this reference area are read from the restaurant database 24 using this circle area as a reference area.

(2-2) Analysis of Detailed Browsing History Next, the control unit 1 activates the detailed browsing history analyzing unit 132 in step S62. Then, under the control of the detailed browsing history analysis unit 132, first, the detailed browsing history data of the learning target user is read from the detailed browsing history storage unit 22 in step S621, and based on the closeness of the browsing time from the detailed browsing history. Extract a group of detailed browsing. This group is called a browsing chunk. Specifically, when there are many detailed browsing histories of a certain user and there are other histories within a certain time (for example, 30 minutes) before and after each history, these histories are collected as the same browsing chunk. By sequentially performing this process in time series, a browsing chunk group is extracted from the detailed browsing history data. FIG. 7A shows an example of the extraction result. In this example, browsing chunk 1, browsing chunk 2, and browsing chunk 3 are extracted.

  In general, user information browsing is concentrated in a certain time zone. At this time, when selecting each detailed browsing store in a series of information browsing, it is considered that there is also a relationship with other detailed browsing stores. For this reason, it is effective to improve analysis by recommending browsing chunks for analysis.

とする。 Next, in step S622, for each store appearing in the detailed browsing history data, the total number of tags held by each store is randomly extracted from the reference restaurant set created by the reference restaurant set creating unit 131. Is calculated.
For example, each of the extracted browsing chunks is regarded as a sample from the reference restaurant set, and the deviation probability pij from the random sampling in the browsing chunk j of the tag i is expressed as follows:

And

  In addition, r is the number of restaurants in the reference restaurant set, and di is the number of restaurants to which the tag i in the reference restaurant set is assigned. Further, vj is the number of restaurants that appear in the browsing chunk j, and nij is the number of restaurants that are assigned the tag i in the browsing chunk j. These r, di and nij are read from the restaurant database 24 and calculated.

(2-3) Analysis of GPS Movement History Next, the control unit 1 activates the GPS movement history analysis unit 133 in step S63, and under the control of the GPS movement history analysis unit 133, the GPS movement history analysis is performed as follows. Perform analysis processing.
That is, first, in step S631, for each user, a GPS movement history in which the user stays within a certain range for a certain period of time is extracted based on the GPS movement history data stored in the GPS movement history storage unit 21. Is a stay chunk. FIG. 7B shows an example of the extraction result. In this example, the stay chunk 1 and the stay chunk 2 are extracted.

  For the extraction of the stay chunk, there is a method of using the DBSCAN method in order to exclude outliers often appearing at GPS positioning points and fluctuations in positioning results to a certain extent. In addition, there is a method of extracting a case where there are a certain number of GPS positioning points for a certain period, a certain range, as a stay chunk. The DBSCAN method is a method for obtaining a group connected by a connection relationship (Directly Density Reachable), which is determined by two parameters of a threshold value Eps of the distance to be analyzed and a threshold value MinPts of the number of adjacent stores. However, here, the connection relationship search is limited to the time series, and when the connection can be extracted, the time interval between the start point and the end point is calculated, and the group exceeding the predetermined time is set as the stay chunk. At that time, the first time is the stay start time and the last time is the stay end time.

  Subsequently, in step S632, restaurants that exist within a certain distance are extracted from the restaurant database 24 with the center of gravity of the GPS positioning points belonging to the extracted stay chunk as a central point. This extracted restaurant is referred to as “visitable restaurant” in the stay chunk. The constant distance parameter is determined by the positioning accuracy of the GPS device to be used, the history characteristics, and the like.

とする。 Next, in step S633, the visit possibility restaurant in the extracted stay chunk is considered as a sample from the reference restaurant set, and the deviation probability pik from the random extraction in the tag stay chunk is calculated as follows.
That is, assuming that the stay chunk of tag i is k, the divergence probability pik with random sampling in this stay chunk k is

And

Note that Sk is the number of restaurants that can be visited by stay chunk k, and nik is the number of restaurants that can be visited by tag i in stay chunk k.
Assuming that there is one store in principle for each stay chunk, in equation (5), if tag i is assigned to any one of the visitable restaurants The divergence probability pik of is calculated.

  Next, in step S634, the possibility of eating and drinking behavior according to time is calculated. In other words, in order to recommend a store with higher accuracy, it is appropriate to increase the influence of stay chunks that are highly likely to be visiting a restaurant to be recommended. Since each stay chunk has information on a stay start time and a stay end time, the importance considering the time is calculated and used for analysis.

In the first embodiment, in particular, for the purpose of recommending restaurants, meal activity rate data for each time zone is acquired and stored in advance, and the stay end time from the stay start time of each stay chunk. The average value of the meal action rate in the period until is set as the importance wk of the stay chunk k. For the calculation of wk, for example, the value of the social life time survey of the Statistics Bureau of the Ministry of Internal Affairs and Communications can be used.
FIG. 8 is a diagram showing the concept of the divergence probability pik between the random extraction in the browsing chunk j and the random extraction in the stay chunk k, calculated as described above.

(2-4) Restaurant Co-occurrence Analysis Next, the control unit 1 activates the restaurant co-occurrence analysis unit 134 in step S64, and under the control of the restaurant co-occurrence analysis unit 134, a detailed visit is made as follows. Perform a later visit analysis. That is, first, in step S641, a restaurant that appears in each browsing chunk based on the browsing chunk extracted by the detailed history analysis unit 132 and the stay chunk extracted by the GPS movement history analysis unit 133 has its browsing time. The restaurant is extracted as a co-occurrence restaurant when it appears as a restaurant that may be visited in a stay chunk whose stay start time is within a certain time (for example, 24 hours). It is speculated that this co-occurring restaurant may have a visit based on information about the restaurant where the user browsed the website in detail.

  However, it is necessary to eliminate the possibility that it stays frequently in the downtown area during the time and closes to many restaurants and co-occurs. For this reason, the rarity that each co-occurrence restaurant has become a visitable restaurant in a stay chunk within a certain period is calculated as follows from the deviation from random sampling.

として算出する。 That is, in step S642, all the visiting possibility restaurants of the stay chunk within a certain time from the browsing time of the browsing chunk are regarded as samples from the reference restaurant set created by the reference restaurant set creation unit 131. And the deviation probability picj with the random sampling for the co-occurrence restaurant in the browsing chunk j of the tag i,

Calculate as

Note that vij is the number of detailed browsing stores to which the tag i is assigned in the browsing chunk j, oj is the total number of restaurants that can be visited by all the staying chunks within a certain time from the browsing time of the browsing chunk j, and cij is the browsing chunk j It is the number that co-occurs as a restaurant that can visit a stay chunk within a certain period of time from the browsing time at the detailed browsing store to which the tag i is assigned.
FIG. 9 shows the concept of the co-occurrence restaurant extraction process and the divergence probability calculation process.

(2-5) Calculation of Preference Score Finally, the control unit 1 activates the preference score calculation unit 135 in step S65. Under the control of the preference score calculation unit 135, the user preference score for each tag is as follows. To calculate.

により統合し、各タグｉに対するユーザの嗜好スコアｚi を算出する。この嗜好スコアｚi が大きいほど店舗に対するユーザの嗜好が強いと言える。 That is, first, in step S651, the divergence probabilities pij, pic, picj calculated by the detailed browsing history analysis unit 132, the GPS movement history analysis unit 133, and the restaurant co-occurrence analysis unit 134, respectively,

And the user's preference score zi for each tag i is calculated. It can be said that the greater the preference score zi, the stronger the user's preference for the store.

  Φ1 (p) is the reciprocal of the cumulative density function of the standard normal distribution. J is the total number of browsing chunks, K is the total number of staying chunks, and wk is the importance of the staying chunk k. Expression (7) is based on the Stouffer's Z-score method, and the probability can be combined by introducing a weight in a natural manner with respect to the importance of each combination target. That is, the preference score z is weighted according to at least one of the degree of association between the restaurant information and each concept and the importance of each concept.

(3) Restaurant recommendation process (Recommend process)
The control unit 1 of the information recommendation server SV monitors the arrival of an information recommendation request (user request) from the user in step S7 while periodically performing the collection processing of the GPS movement history data and the detailed browsing history data. Yes. In this state, when the user request transmitted from the user terminals UT1 to UTn is received, the restaurant recommendation processing unit 14 is activated in step S8, and under the control of the restaurant recommendation processing unit 14, the restaurant's recommendation is as follows. Execute the recommendation process. FIG. 6 is a flowchart showing the processing procedure and processing contents.

(3-1) Creation of candidate restaurant set First, in step S81, under the control of the candidate restaurant set creation unit 141, based on the store conditions represented by the region and the tag included in the received user request, A restaurant corresponding to the condition is extracted from the restaurant database 24, thereby creating a candidate restaurant set. The candidate restaurant set is created by searching the restaurant database 24 based on the conditions specified by the user request, for example, “within 1 km from the X station” as the regional condition.

により算出する。 (3-2) Calculation of Restaurant Score Next, in step S82, the restaurant score calculation unit 142 is activated, and the user corresponding to the tag i stored in the preference database 23 under the control of the restaurant score calculation unit 142. Based on the preference score zi, the restaurant score (r) of each restaurant included in the candidate restaurant set is

Calculated by

  Note that r is an identifier of a restaurant that is an element of a candidate restaurant set, Tr is a set of tags assigned to the restaurant r2, and tr indicates the number of tags included in Tr. It should be noted that the sorting of restaurants by the Restaurant Score according to this formula considers the norm of the preference score zi as a constant at each calculation point when the preference score zi and the tag of each restaurant are regarded as a matrix for the tag i. The result is equivalent to the cosine similarity.

(3-3) Creation and Distribution of Recommendation List Next, in step S83, the recommendation list creation unit 143 is activated, and each food and drink calculated by the restaurant score calculation unit 142 under the control of the recommendation list creation unit 143. Based on the store scores zi, the restaurant information is sorted in descending order of the scores zi (step S831). That is, the restaurant information is sorted in descending order of the degree of coincidence with the user's preference score. Thereby, a recommendation list of restaurants suitable for the user is created. Then, the created recommendation list is stored in the recommendation list storage unit 25 in step S832, and is returned to the recommendation request source user terminals UT1 to UTn via the communication network NW in step S833.

(4) Evaluation of Recommendation Method According to Embodiment The recommendation accuracy according to the recommendation method according to this embodiment was compared with the recommendation accuracy according to Random Sort and Cosine Similarity.
Random Sort is a method of randomly sorting each store included in the candidate restaurant set. For example, if there is no preference, condition input, or other recommendation method by the user in a search that is “within 1 km from the current location” or the like, this method is the same. The expected value is used for evaluation of this Random Sort.

  Cosine Similarity is a conventional method of memory-based content-based filtering. When a tag is attached to each restaurant that appears in the detailed browsing history, the result of accumulating “1” for each tag type is evaluated as a vector. This is a method of setting a restaurant score as a cosine similarity with a vector by a tag given to a target restaurant.

  As an accuracy evaluation index, DCG (Discounted Cumulative Gain), which is a conformity index that takes into consideration the rank that is frequently used in the information search field, was calculated with respect to the recommendation order of 110 restaurants to be evaluated by the above methods. The reason for adopting DCG is that, for example, in a gourmet recommendation service for portable terminals (smartphones, etc.) that is assumed as an application target of the present invention technique, the amount of information that can be displayed on the terminal at a time is small. It is important to be able to present in order.

  The input value g (i) required for the calculation of DCG is “I want to go” = 7, “I want to go” = 6, "I want" = 5, "Neither" = 2, "I don't want to go" = 1. Among these, the opening of the DCG input value g (i) between “Well I want to go” and “Neither” is “I don't want to go to the restaurant recommendation satisfaction due to subject hearing in advance” This is to increase the attenuation of DCG when a restaurant with a low evaluation is ranked higher because it has been found that how to reduce restaurant recommendations is important. In addition, the DCG calculation range of this evaluation was ranked in the top 32 reflecting the average confirmation range for the recommendation result obtained by subject hearing similarly.

  Under the conditions as described above, when the relationship between the DCG value of the subject average and the history collection period was determined by the recommendation method according to the present embodiment and the Random Sort and Cosine Similarity, the result shown in FIG. 10 was obtained. . As can be seen from FIG. 10, the recommendation accuracy is improved by the integrated analysis of the detailed browsing history and the GPS movement history, and it is understood that the recommendation can be made with higher accuracy even in a situation where the detailed browsing history is small.

  For Cosine Similarity, in order to take into account the difference in the likelihood of appearance in the detailed browsing history of each tag i, the value added at the time of tag appearance is weighted to 1 / log (Ei), and 1 / Ei However, the recommendation accuracy was not improved. Ei is the total number of restaurants to which the tag i in the restaurant database 24 is assigned.

[Effect of the first embodiment]
As described in detail above, in the first embodiment, based on detailed browsing history data for a restaurant website, a collection of detailed browsing according to the proximity of the browsing time is extracted as a browsing chunk, and GPS Based on the movement history data, when the user stays within a certain range at a certain time, a set of the stay location information is extracted as a stay chunk that represents a group of the user's action range. Then, these chunks are regarded as samples from the reference restaurant set, and the deviation probability pij from the random extraction in the browsing chunk and the deviation probability pik from the random extraction in the stay chunk are calculated, respectively, and these deviation probabilities are calculated. The user's preference score for the restaurant tag is calculated in a unified manner, and the restaurant recommendation list for the user is created and distributed using this preference score.

  That is, the restaurants included in the detailed browsing history and the nearby stores extracted from the GPS movement history are regarded as sample extraction by the user from the set of restaurants within the user's behavior range, and are given to each store. The divergence between the number of appearances of tags and the number of appearances by random extraction is calculated, and a tag with a large divergence in the number of appearances is considered to be related to the user's preference and is used for recommendation.

  Therefore, highly accurate information recommendation processing can be performed using the detailed browsing history by the user and the GPS movement history. For this reason, it is possible to perform highly accurate information recommendation without depending on the user's manual input of conditions even in a situation where the user's browsing history record is small. In addition, in the process of obtaining the recommendation list, the detailed browsing history and the GPS movement history are converted into common data, and are analyzed and integrated using a common index. An accuracy recommendation list can be created.

[Second Embodiment]
According to the second embodiment of the present invention, in the co-occurrence analysis of restaurants, a restaurant that appears in each browsing chunk may be visited in a stay chunk whose stay start time is within a certain time from the browsing time. When it appears as a restaurant, this restaurant is extracted as a co-occurrence restaurant, and a restaurant that appears as a restaurant that can be visited in each stay chunk has a browsing time within a certain time from the stay start time When it appears in the browsing chunk, it is extracted as a co-occurrence restaurant in the stay chunk targeting this restaurant.
Also in the second embodiment, the basic configuration and processing operation of the information recommendation device are the same as those in the first embodiment, and therefore description will be made with reference to FIGS.

(2-4) ′ Restaurant co-occurrence analysis The control unit 1 activates the restaurant co-occurrence analysis unit 134 in step S64, and performs the visit analysis as described below under the control of the restaurant co-occurrence analysis unit 134. Do. That is, first, in step S641, a restaurant that appears in each browsing chunk based on the browsing chunk extracted by the detailed history analysis unit 132 and the stay chunk extracted by the GPS movement history analysis unit 133 has its browsing time. When the stay start time within a certain time (for example, one week) appears as a restaurant that may be visited, the restaurant is extracted as a co-occurring restaurant in the target viewing chunk.

At the same time, with regard to a restaurant that appeared as a possibility of visiting in each stay chunk, the restaurant appeared in a browse chunk whose browsing time was within a certain time (for example, one week) from the stay start time in the stay chunk. In this case, the restaurant is extracted as a co-occurrence restaurant in the stay chunk targeting this restaurant.
And this co-occurrence restaurant is guessed to be a restaurant which the user visited with the information of the restaurant which browsed the website in detail as a trigger, or the website with the possibility of browsing the website in detail with the visit as a trigger. It can be assumed that such restaurants are strongly related to user preferences.

However, the above-mentioned co-occurrence restaurants are also extracted when the user frequently stays in a downtown area within a certain time and approaches a large number of restaurants. This type of restaurant may have happened by chance and needs to be excluded.
Therefore, the rarity of each co-occurring restaurant becoming a visitable restaurant in a stay chunk within a certain period of time is calculated as follows from the deviation from random sampling.

として算出する。 That is, in step S642, all the visiting possibility restaurants of the stay chunk within a certain time from the browsing time of the browsing chunk are regarded as samples from the reference restaurant set created by the reference restaurant set creation unit 131. And the deviation probability pijco from the random sampling by the co-occurrence restaurant in the browsing chunk j of the tag i,

Calculate as

  In addition, bj is the number of detailed browsing stores that have appeared in the browsing chunk j, r is the number of restaurants in the reference restaurant set, and Sj is a set of identifiers indicating stay chunks within a certain period from the time of the browsing chunk j. cijco is a detailed browsing store to which the tag i is assigned in the browsing chunk j, and is the number that co-occurs with the visiting possibility restaurants of the stay chunks (stay chunks included in Sj) within a certain time from the browsing time.

として算出する。 In the same manner, the same processing is performed for the co-occurrence of the staying possibility visitor restaurant with the restaurant appearing in the browsing chunk within a certain time from the stay time. In other words, the deviation probability pikco from the random sampling by the co-occurrence restaurant in the stay chunk k of tag i,

Calculate as

  Here, sk is the number of restaurants that can be visited in the stay chunk k, and Bk is a set of identifiers of browsing chunk groups within a certain period from the time of the stay chunk k. In addition, cikco is a visit possibility restaurant where tag i is given in stay chunk k, and is the number that co-occurs with the detailed browsing restaurants within a certain time from the browsing time (viewing chunks included in Bk). is there.

(2-5) ′ Preference Score Calculation The control unit 1 activates the preference score calculation unit 135 in step S65, and under the control of the preference score calculation unit 135, the user preference score for each tag is as follows. calculate.

とする。 That is, first, in step S651, for each divergence probability calculated by the detailed browsing history analysis unit 132, the GPS movement history analysis unit 133, and the restaurant co-occurrence analysis unit 134, the initial state before the history is collected is t = 1. And an identifier obtained by incrementing t in the order in which history chunks are extracted is given. At this time, the divergence probabilities of the stay chunk and the browsing history chunk extracted at time t are pit and pitco, and the preference score for the tag i is

And

  It can be considered that the user's preference for the tag i is stronger as the preference score Zit is larger. Φ-1 (p) is the reciprocal of the cumulative density function of the standard normal distribution, and w t is the importance of the chunk t. Also, equation (13) is based on the Stouffer's Z-score method, and the probability can be combined by introducing a weight in a natural manner with respect to the importance of each combination target. That is, the preference score z is weighted according to at least one of the degree of association between the restaurant information and each concept and the importance of each concept.

  Furthermore, this processing can be updated sequentially with the addition of stay chunks or browsing chunks, thereby increasing the computational resource efficiency. Specifically, the data that needs to be retained in the sequential processing is only the xit and yit of each tag and the store ID group that appears in the browsing history chunk and the stay chunk from the present time to the past certain time.

[Effects of Second Embodiment]
As described above, according to the second embodiment, “proximity store visited after browsing store information” and “store information browsing after store visit” are analyzed as separate models, so proximity before and after store information browsing Compared to the first embodiment in which only a visit to a store to be analyzed is used as a model, a more detailed analysis can be performed. Further, in the first embodiment, in order to analyze a visit to a neighboring store before browsing store information, sequential update for each addition of chunks could not be performed, but sequential update for each addition of chunks of analysis results can be performed. Thus, computer resources such as calculation cost and memory can be reduced.

(Other embodiments)
The present invention is not limited to the above embodiments. For example, in the first embodiment, in the restaurant co-occurrence analysis unit 134 of the preference learning processing unit 13, the user arbitrarily browses the restaurant website before providing the restaurant recommendation list. Based on the acquired detailed browsing information, the possibility that the user actually visited the restaurant was reflected in the analysis. However, the recommendation list created by the restaurant recommendation processing unit 14 is not limited to this, and the detailed browsing information acquired by accessing the website of the restaurant based on this recommendation list is distributed to the user. It may be reflected in the store co-occurrence analysis.

  In the above embodiment, a restaurant has been described as an example of a spot to be recommended. However, other spots may be used as long as the spot is visited by browsing information. For example, a spot visited during leisure time such as a weekend. As an example, a park, an amusement park, a zoo, various shipping facilities, a game center, and a karaoke house can be considered. About these spots, the evaluation, attributes, facilities, etc. are accumulated in the database as tags, the user's preference is analyzed by analyzing the information browsing history and GPS movement history by the user, and the recommendation list is created. It is possible to provide a spot that suits the user's preference. In addition, it can be applied to search for accommodation facilities and travel information, sightseeing spot guide service, real estate information search, and ceremonial hall search.

  Further, each divergence probability may be calculated by a hypergeometric distribution, binomial distribution or normal distribution. In addition, the configuration of the recommendation server SV, the collection process of the GPS movement history and the browsing history, the preference learning process, the processing procedure and processing contents of the restaurant recommendation process, and the like are variously modified without departing from the gist of the present invention. It can be implemented.

  In short, the present invention is not limited to the above-described embodiments as they are, and can be embodied by modifying the components without departing from the scope of the invention in the implementation stage. Moreover, various inventions can be formed by appropriately combining a plurality of constituent elements disclosed in the above embodiments. For example, some components may be deleted from all the components shown in each embodiment. Furthermore, you may combine suitably the component covering different embodiment.

  SV: recommendation server, UT1 to UTn ... user terminal, NW ... communication network, 1 ... control unit, 2 ... storage unit, 11 ... GPS movement history collection unit, 12 ... detailed browsing history collection unit, 13 ... preference learning processing unit, DESCRIPTION OF SYMBOLS 14 ... Restaurant recommendation processing part, 21 ... GPS movement history storage part, 22 ... Detailed browsing history storage part, 23 ... Preference database, 24 ... Restaurant database, 25 ... Recommendation list storage part, 131 ... Reference restaurant group creation part 132 ... Detailed browsing history analysis unit, 133 ... GPS movement history analysis unit, 134 ... Restaurant co-occurrence analysis unit, 135 ... Preference score calculation unit, 141 ... Candidate restaurant set creation unit, 142 ... Restaurant score calculation unit, 143: A recommendation list creation unit.

Claims (11)

An information recommendation method for analyzing a user's interest in a plurality of geographically dispersed spots by a computer and creating information on a spot recommended for the user based on the analysis result,
A first step in which the computer acquires information representing the browsing history of the user for a browsing site that posts information on the plurality of spots, and stores the information in a first storage medium;
A second process in which the computer acquires information representing a geographical movement history of the user and stores the information in a second storage medium;
Based on the browsing history information stored in the first storage medium, the computer generates a browsing chunk by clustering a plurality of spot browsing information included in the browsing history information on condition that the browsing time is close A third process to
The computer determines a reference area where the user can visit based on movement history information stored in the second storage medium, information input in advance by the user or connection information with a wireless base station, A fourth step of extracting, from the stored spot information, a set of spot information included in the determined reference area as a reference spot set;
The computer includes a number a of spot browsing information included in the generated browsing chunk, a number n of spot browsing information having a predetermined concept among the spot browsing information groups included in the browsing chunk, and the reference spot set. The spot browsing information in which the concept appears in the browsing chunk based on the number S of spot information included in the reference spot and the number N of spot information to which the concept is assigned in the spot information group included in the reference spot set The first probability that the number of is equal to or greater than n and the second probability that is less than n are calculated, respectively, and the cumulative distribution function of the standard normal distribution is calculated based on the calculated first and second probabilities. the inverse function, and a fifth step of calculating an interest score z representing the degree of interest of the user with respect to the concept
Information recommendation method characterized by comprising the. Based on the movement history information stored in the second storage medium, the computer generates a stay chunk by clustering a plurality of stay information included in the movement history information on the condition of distance proximity, Further comprising a step of associating spots with the stay chunks on condition of proximity.
In the fifth step, the number of spot information included in the generated stay chunk is set to the a, and the number of spot information having a predetermined concept among the spot information groups included in the stay chunk is set to the n. The information recommendation method according to claim 1, wherein the interest score z is calculated. Based on the movement history information stored in the second storage medium, the computer generates a stay chunk by clustering a plurality of stay information included in the movement history information on the condition of distance proximity, Further comprising a step of associating spots with the stay chunks on condition of proximity.
In the fifth step, the number of spot browsing information appearing in the browsing chunk is set as the N, and a stay chunk group existing within a predetermined time from the browsing time of the browsing chunk is set as a co-occurrence chunk. The number of spot groups associated with stay chunks in a co-occurrence chunk on condition that the distance is close is set as a, and among the spot groups associated with stay chunks within the predetermined time, the browsing chunks The information recommendation method according to claim 1, wherein the interest score z is calculated by setting n as the number of spots that appear on the screen. Based on the movement history information stored in the second storage medium, the computer generates a stay chunk by clustering a plurality of stay information included in the movement history information on the condition of distance proximity, Further comprising the step of associating a spot group with this stay chunk on condition of close distance;
In the fifth process, the number S of spot information included in the reference spot set is set to a, and the number of spot information in the browsing chunk is set to a, and exists within a predetermined time from the browsing time of the browsing chunk. The number of spot information appearing in the stay chunk group is N, and in the spot information appearing in the browsing chunk, the spot information group co-occurring with the spot information appearing in the stay chunk having the stay start time within a certain time is shared. 2. The information recommendation method according to claim 1, wherein the interest score z is calculated as a starting chunk, where n is the number of spots in the co-occurring chunk. Based on the movement history information stored in the second storage medium, the computer generates a stay chunk by clustering a plurality of stay information included in the movement history information on the condition of distance proximity, Further comprising the step of associating a spot group with this stay chunk on condition of close distance;
In the fifth step, the number S of spot information included in the reference spot set is set as described above, the number of spot information in the browsing chunk is a, and a predetermined time set in advance from the stay start time of the stay chunk. The number of spot information appearing in the browsing chunk group whose inside is the browsing time is N, and in the spot information group appearing in the stay chunk, the spot information group co-occurring with the spot information appearing in the browsing chunk within a certain time The information recommendation method according to claim 1, wherein the interest score z is calculated as a co-occurrence chunk, where n is the number of spots in the co-occurrence chunk. In the fifth step, for the browsing chunk, stay chunk, or co-occurrence chunk, the interest according to at least one of a user operation history, a time and situation of each chunk, and a degree of association between the spot and the concept. 6. The information recommendation method according to claim 1, wherein the score z is weighted.   The information recommendation method according to claim 1, wherein the fifth step calculates the first probability and the second probability by a hypergeometric distribution, a binomial distribution, or a normal distribution. For spot information in which one or more concepts appear, the method further includes a sixth step of calculating a user evaluation score for the spot information using the interest score z of each concept that appears in the spot information. The information recommendation method according to claim 1. 9. The information recommendation method according to claim 8, wherein in the sixth step, the interest score z is weighted according to at least one of the degree of association between the spot information and each concept and the importance of each concept. . An information recommendation device that analyzes a user's interest in a plurality of geographically dispersed spots and creates information on spots recommended to the user based on the analysis result,
Means for acquiring and storing in a first storage medium information representing the user's browsing history for a browsing site that posts information relating to the plurality of spots, respectively.
Means for acquiring information representing the geographical movement history of the user and storing it in a second storage medium;
Based on the browsing history information stored in the first storage medium, a means for clustering a plurality of spot browsing information included in the browsing history information on the condition of near browsing time and generating a browsing chunk;
A reference area that the user can visit is determined based on the movement history information stored in the second storage medium, and the spot information included in the determined reference area is determined from the spot information stored in advance. Means for extracting the set as a reference spot set;
The number a of the spot browsing information included in the generated browsing chunk, the number n of the spot browsing information having a predetermined concept among the spot browsing information group included in the browsing chunk, and the spot included in the reference spot set Based on the number S of information and the number N of spot information to which the concept is given among the spot information groups included in the reference spot set, the number of spot browsing information in which the concept appears in the browsing chunk is: A first probability that is greater than or equal to n and a second probability that is less than n are calculated, respectively, and based on the calculated first and second probabilities, an inverse function of a cumulative distribution function of a standard normal distribution is obtained . Means for calculating an interest score z representing the degree of interest of the user with respect to the concept ;
Information recommendation apparatus characterized by comprising a.   The program which makes the said computer perform the process of each process with which the information recommendation method in any one of Claims 1 thru | or 9 is provided.

Images