JP2012256239A - Destination prediction system and program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To achieve accurate prediction of a destination even in the case of non-regular movement.SOLUTION: A data collection part 50 collects access data and movement data for each user in chronological order. A content extraction part 60 extracts each place name from a text on the website, with respect to each of the collected access data, and extracts a place corresponding to the place name. A staying place extraction part 56 extracts staying place data for each user on the basis of the movement data collected in chronological order. A correlation calculation part 62 calculates a correlation for each user upon associating access data for each user with each staying place data extracted from a movement history. A prior probability learning part 64 learns a prior probability of going to a place shown by each staying place data. A destination prediction part 70 calculates a probability of going to the place shown by each staying place data and predicts a destination on the basis of the calculated probability, with respect to a user subject to the prediction.

Description

  The present invention relates to a destination prediction apparatus and program, and more particularly, to a destination prediction apparatus and program for predicting a user's destination from a user's document data browsing history and position information history.

  2. Description of the Related Art Conventionally, a system that estimates a user's interest from Internet browsing history is known (for example, Patent Document 1). In addition, a system is known that generates and predicts an action pattern from a browsing history / movement history and provides information (Patent Document 2).

JP 2009-232415 A JP 2008-311212 A

  However, the system described in Patent Document 1 is a system that enumerates and recommends portions that the user is likely to be interested in, and has a problem that the destination cannot be predicted. In addition, since the movement history data is not used and only the Internet history is used, there is a problem that the estimation accuracy is low.

  In addition, the technique described in Patent Document 2 described above is a point-to-point movement prediction based on a movement distance, and has a problem that only a destination in a regular movement can be predicted.

  The present invention has been made to solve the above-described problems, and provides a destination prediction apparatus and program capable of accurately predicting a destination even when it is not a periodic movement. Objective.

  In order to achieve the above object, a destination prediction apparatus according to the present invention includes a history collection unit that collects document data browsing history and position information history for a plurality of users, and each document data of the collected browsing history. Categorizing means for classifying the browsed document data according to text in the document data, and extracting means for extracting stay place data for each user based on the history of the collected location information Corresponding to each of the combinations of the document data and the staying place data by associating each document data of the browsing history of the user with each staying place data extracted from the history of the position information for each user. Based on the calculated correlation, the probability calculating means calculates the probability of going to the location indicated by each stay location data for the prediction target user, and the calculated Is configured to include a, a destination predicting means for predicting a destination on the basis of probability.

  The program according to the present invention includes a computer, a history collecting unit that collects browsing history of document data and a history of position information for a plurality of users, and the text in the document data for each document data of the collected browsing history. Accordingly, the classification means for classifying the browsed document data, the extraction means for extracting the stay place data for each user based on the collected location information history, and the browsing history of the user for each user. Correlation calculation means for associating each document data with each stay place data extracted from the history of the position information and calculating a correlation for each combination of the document data and the stay place data, and the calculated Based on the correlation, for the prediction target user, the probability of going to the place indicated by each stay place data is calculated, and the target is calculated based on the calculated probability. Is a program for functioning as a destination predicting means for predicting the earth.

  According to the present invention, the history collection unit collects the browsing history of document data and the history of position information for a plurality of users. The classified means classifies the browsed document data according to the text in the document data for each document data of the collected browsing history. Based on the collected history of position information, the extraction means extracts stay place data for each user.

  Then, for each user, the document data of the browsing history of the user and the staying place data extracted from the history of the location information are associated with each other by the correlation calculating means, and the document data and the staying place data are A correlation for each of the combinations is calculated. Based on the calculated correlation, the destination predicting means calculates a probability of going to the place indicated by each stay data for the prediction target user, and predicts the destination based on the calculated probability.

  In this way, for each user, the correlation is calculated by associating each document data of the browsing history with each place of stay data extracted from the history of the position information, and based on the calculated correlation, By calculating the probability of going to the place indicated by each place of stay data, the destination can be predicted with high accuracy even when it is not a regular movement.

  The destination prediction apparatus according to the present invention browses the document data belonging to the classification for each combination of the classification of the document data and each of the stay data based on the correlation calculated by the correlation calculation means. Probability calculation means for calculating a prior probability to go to the place indicated by the stay location data sometimes, wherein the destination prediction means is calculated by the correlation calculated for the prediction target user and the probability calculation means. Based on the prior probabilities that have been made, for the prediction target user, a probability that each destination is a destination can be calculated and the destination can be predicted.

  The correlation calculating means according to the present invention relates each document data of the browsing history of the user and each stay place data extracted from the history of the position information, and a combination of the document data and the stay place data. For each of the above, a correlation according to a time difference between the viewing time of the document data and the staying time of the staying place data can be calculated.

  The correlation calculation means is configured to calculate, for each combination of the document data and the stay place data, a time difference between the view time of the document data and the stay time of the stay place data, and the length of the view time of the document data. The correlation can be calculated based on the corresponding browsing probability and the stay probability according to the length of the stay time of the stay place data.

  For each document data of the collected browsing history, the classification means extracts a place name or a product name from the text in the document data, and uses the place data indicating the extracted place name or the extracted product name. The location data of the store to sell can be assigned as the classification of the document data.

  The classification means extracts words from the text in the document data for each document data of the collected browsing history, performs clustering on the extracted set of words, and extracts the words extracted from the document data. The cluster to which the set belongs can be assigned as the classification of the document data.

  The program of the present invention can also be provided by being stored in a storage medium.

  As described above, according to the destination prediction apparatus and program of the present invention, for each user, the correlation is calculated by associating each document data of the browsing history with each place of stay data extracted from the history of the position information. Then, based on the calculated correlation, by calculating the probability of going to the location indicated by each place of stay data for the prediction target user, it is possible to accurately predict the destination even if it is not a regular movement The effect of being able to be obtained.

It is a block diagram which shows the destination prediction system which concerns on the 1st Embodiment of this invention. It is a block diagram which shows the portable terminal which concerns on the 1st Embodiment of this invention. It is a block diagram which shows the processing server which concerns on the 1st Embodiment of this invention. It is a figure for demonstrating the process which calculates a correlation. It is a flowchart which shows the content of the correlation learning process routine in the process server which concerns on the 1st Embodiment of this invention. It is a flowchart which shows the content of the destination prediction process routine in the process server which concerns on the 1st Embodiment of this invention.

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

  As shown in FIG. 1, the destination prediction system 10 according to the first embodiment includes a client PC 12 for a user to browse a website, and a mobile phone that measures the location information while the user browses the website. A terminal 14, a mobile terminal 16 that is mounted on a mobile body (for example, a vehicle) and that measures position information, and a processing server 18 that performs a process of predicting a user's destination are provided. Each of the client PC 12, the portable terminal 14, and the mobile terminal 16 and the processing server 18 are connected via the Internet 20.

  As shown in FIG. 2, the mobile terminal 14 controls the access to the position measuring unit 22 that measures the position of the own terminal, the operation unit 24 for the user to operate the own terminal, and the Web site, and browsing A computer 26 that generates data and movement data, a display unit 28 that displays the accessed website to the user, and a communication unit 30 that accesses the website and transmits browsing data and movement data. Yes.

  The position measurement unit 22 is configured using, for example, a GPS sensor or an inertial sensor, and measures the current position of the terminal itself.

  The communication unit 30 accesses the website via the Internet 20 by wireless communication, and transmits browsing data and movement data generated by the computer 26 to the processing server 18. Further, the communication unit 30 transmits a destination prediction request to be described later to the processing server 18 by wireless communication.

  The computer 26 includes a CPU, a RAM, and a ROM, and is functionally configured as follows. The computer 26 accesses the website in accordance with the operation of the operation unit 24 by the user, displays the website screen on the display unit 28, and the browsing data regarding the website displayed by the browsing control unit 32. And a movement data generation unit 36 that generates movement data indicating the position of the terminal measured by the position measurement unit 22 and outputs the movement data to the communication unit 30. I have.

  When the user operates the operation unit 24 and inputs the user ID and password to the computer 26, the input user ID and password are transmitted to the processing server 18. As a result, the mobile terminal 14 is in a login state with respect to the processing server 18.

  The browsing data generation unit 34 includes the URL of the website displayed by the browsing control unit 32, the HTML text of the website, the time when the website was accessed (viewing time), the site display time, the user ID, and the mobile terminal 14 generates browsing data including items including a terminal ID and a terminal type set in advance.

  The movement data generation unit 36 includes a position (latitude, longitude, and altitude) measured by the position measurement unit 22, a position time when the position is measured, a user ID, a terminal ID preset in the mobile terminal 14, and The movement data including each item including the terminal type is generated.

  The client PC 12 has the same configuration as the operation unit 24, the display unit 28, the communication unit 30, the browsing control unit 32, and the browsing data generation unit 34. On the other hand, a screen of the website is displayed, and browsing data relating to the displayed website is generated and transmitted to the processing server 18.

  The mobile terminal 16 is a car navigation system, for example, and has the same configuration as the position measurement unit 22, the operation unit 24, the communication unit 30, and the movement data generation unit 36. The mobile terminal 16 measures the position of the own terminal, generates movement data indicating the measured position of the own terminal, and transmits the movement data to the processing server 18.

  As shown in FIG. 3, the processing server 18 collects and stores the browsing data, the movement data, and the destination prediction request, the browsing data and the movement data, and stores the destination data of the prediction target user. And a computer 42 for predicting the above.

  The communication unit 40 receives browsing data, movement data, and a destination prediction request from the client PC 12, the mobile terminal 14, and the mobile terminal 16 via the Internet 20.

  The computer 42 includes a CPU, a RAM, and a correlation learning processing routine and a destination prediction processing routine ROM described later, and is functionally configured as follows. The computer 42 includes a data collection unit 50 that collects the received browsing data and movement data, a log database 52 that stores the collected browsing data and movement data as a browsing history and a movement history for each user, and a log database 52 for each user. The travel data acquisition unit 54 for acquiring the travel history, the stay location extraction unit 56 for extracting the stay location based on the acquired travel history, and the browsing data acquisition unit 58 for acquiring the browsing history for each user from the log database 52. And a content extracting unit 60 that extracts a location based on the HTML text of each browsing data of the acquired browsing history, and a correlation calculating unit 62 that calculates a correlation for each combination of browsing data and movement data for each user. Based on the calculated correlation, when browsing a website where a place is extracted, It includes a prior probability learning unit 64 for learning the prior probability indicating the Ku probability, and a learning result database 66 for storing the calculated prior probability of correlation and learned was. The content extraction unit 60 is an example of a classification unit, and the prior probability learning unit 64 is an example of a probability calculation unit.

  The data collection unit 50 collects the browsing data and movement data received by the communication unit 40, stores the browsing data time series in the log database 52 as browsing history for each user ID, and moves the movement data time series. Stored in the log database 52 as a history.

  The stay location extraction unit 56 extracts a stay location for each user ID as described below based on the time series of the movement data acquired by the movement data acquisition unit 54.

  First, a portion where the positions indicated by the movement data are dense is searched from the time series of the movement data. For example, a portion where the velocity data obtained from the position indicated by the movement data is a certain value or less is extracted as a dense portion. Then, for each dense portion, a time window T (start time t1 to end time t2) including the dense portion is set as an initial value.

  Next, for each time window T, the start time t1 is moved to maximize the stay probability P (j) expressed by the following equation (1).

ただし、α１は予め定められたパラメータである。

Here, α1 is a predetermined parameter.

  Further, for each time window T, the end time t2 is moved to maximize the stay probability P (j) shown in (1) above.

  Here, if the time windows T overlap, a plurality of overlapping time windows T are integrated. Through the above processing, a time window T (t1 to t2) that is a candidate for extracting a stay place is determined.

  Then, for each time window T, the stay probability P (j) is calculated for the movement data in the time window T according to the above equation (1). In the above equation (1), the smaller the position variance V, the greater the stay probability P (j), and the longer the time window T width (t2-t1), the greater the stay probability P (j).

  Next, the time window T in which the stay probability P (j) is larger than the threshold value β1 is specified as a place of stay extraction, and the average coordinate is calculated from the movement data of the specified time window T to specify the place. Then, the location data g prepared in advance is classified into the nearest location data g. The location data g is common to all users.

  As described above, the location data g is extracted as the place of stay from the time series of the movement data for each user ID.

  The content extraction unit 60 extracts, for each user ID, a place name included in the browsed website, based on the time series of the browse data acquired by the browse data acquisition unit 58, as described below. A location ID corresponding to the place name is assigned as the site content f. The website is an example of document data, and the location ID is an example of document data classification.

  First, a word that matches a place name / product name included in dictionary data prepared in advance is extracted from the time series of browsing data. When the product name is extracted, the nearby stores are further searched by the product name, and the place name of the search result is extracted. Here, if a plurality of place names are extracted and there is ambiguity, all place names are listed.

  Then, based on a place name dictionary which is a set of place name / latitude / longitude data prepared in advance, place data (latitude / longitude, place ID) is acquired from the place name extracted for the view data, and the view data and Connect with location data. Further, the browsing probability P (i) is calculated according to the following equation (2).

ただし、α0は、予め定められたパラメータであり、Ｔは、Ｗｅｂサイトの表示時間である。上記（２）式によれば、サイト表示時間Ｔが短いほど、閲覧確率が小さくなる。

However, α0 is a predetermined parameter, and T is the display time of the website. According to the above equation (2), the shorter the site display time T, the smaller the browsing probability.

  If a place name dictionary is used, the nearest place name can be acquired from the place data (latitude and longitude).

As shown in FIG. 4, the correlation calculation unit 62 associates each of the browsing history browsing data i with each of the stay location data j extracted from the movement history for each user ID, and browses by the user k. The correlation w ^k _i, j between the data i and the stay place data j is calculated according to the following equation (3).

ただし、α２は予め定められたパラメータであり、ｄは、閲覧データｉに対して結び付けられた場所データと、滞在地データ（場所データ）との距離であり、Ｔは、滞在地データの時刻ｔjと閲覧データの閲覧時刻ｔiとの時間間隔（Ｔ＝ｔｊ−ｔｉ）である。

Here, α2 is a predetermined parameter, d is the distance between the location data linked to the browsing data i and the stay location data (location data), and T is the time tj of the stay location data. And the time interval (T = tj−ti) between the browsing time ti of the browsing data.

According to equation (3), as the distance where d is close correlation w ^k _{i, j} is increased, the shorter the time interval T, the correlation w ^k _{i, j} is increased. Moreover, when the time of the stay place data is earlier than the browsing time of the browsing data, the correlation is zero.

  Instead of the above equation (3), the following equation (4) may be used to calculate the correlation considering the browsing probability P (i) and the stay probability P (j).

また、上記（３）式は一例であり、最も簡単な例としては、Ｔ≧０の場合に、ｗ^k _i,j＝１としてもよい。

Further, the above equation (3) is an example. As the simplest example, when T ≧ 0, w ^k _{i, j} = 1 may be set.

  When the prior probability learning unit 64 browses the website of the browsing data from which the location f is extracted, as will be described below, based on the correlation between the browsing data i calculated for each user and the stay location data j. In addition, the prior probability of going to the place of stay at the place g is obtained. This prior probability is assumed to be common to all users.

First, in accordance with the following formulas (4) and (5), based on the correlation calculated for each user, a total result W ^k _f ^obtained by summing up correlations regarding the location f of the browsing data for the user k is calculated, and browsing to calculate the aggregate result W _f obtained by aggregating the correlation about the location f of the data.

ただし、ｗ^k _i=f,jは、ユーザｋについて算出された、場所ｆが抽出された閲覧データｉと滞在地データｊとの相関である。

However, w ^k _{i = f, j} is the correlation between the browsing data i from which the place f is extracted and the place of stay data j calculated for the user k.

Further, according to the following formulas (6) and (7), based on the correlation calculated for each user, a total result W ^k _g ^obtained by totaling correlations regarding the location g of the stay location data for the user k is calculated, A total result W _g obtained by totaling the correlations regarding the place g of the stay place data is calculated.

ただし、ｗ^k _i,j=gは、ユーザｋについて算出された、閲覧データｉと、場所ｇが抽出された滞在地データｊとの相関である。

However, w ^k _{i, j = g} is a correlation between the browsing data i calculated for the user k and the stay place data j from which the place g is extracted.

Moreover, the correlation regarding the combination of the browsing data location f and the stay location data location g for the user k is tabulated based on the correlation calculated for each user in accordance with the following formulas (8) and (9). The tabulation result W ^k _{f, g} is calculated, and the tabulation result W _{f, g} obtained by tabulating the correlation regarding the combination of the browsing data location f and the stay location data location g is calculated.

ただし、ｗ^k _i=f,j=gは、ユーザｋについて算出された、場所ｆが抽出された閲覧データｉと場所ｇが抽出された滞在地データｊとの相関である。

Here, w ^k _{i = f, j = g} is the correlation between the browsing data i from which the location f is extracted and the stay location data j from which the location g is calculated, which is calculated for the user k.

  Moreover, the total result W which totaled all the correlations is computed according to the following (10) Formula.

そして、上記のように算出された集計結果に基づいて、以下の式に従って、滞在地データの場所ｇの事前確率Ｐ（ｇ）、閲覧データの場所ｆの事前確率Ｐ（ｆ）、滞在地データの場所ｇであり、かつ、閲覧データの場所ｆである事前確率Ｐ（ｆ，ｇ）を算出する。

Then, based on the calculation results calculated as described above, the prior probability P (g) of the location g of the stay location data, the prior probability P (f) of the location f of the browsing data, and the stay location data according to the following formulas: And a prior probability P (f, g) that is the location f of the browsing data.

そして、閲覧データの場所ｆと、滞在地データの場所ｇとの全ての組み合わせの各々について、場所ｆが抽出される閲覧データのＷｅｂサイトを閲覧したときに、場所ｇの滞在地へ行く事前確率Ｐ（ｆ，ｇ）を以下の（１１）式に従って算出する。

Then, for each of all combinations of the location f of the browsing data and the location g of the stay location data, the prior probability of going to the stay location of the location g when browsing the browsing data website from which the location f is extracted P (f, g) is calculated according to the following equation (11).

上記の相関の集計結果と、事前確率Ｐ（ｆ、ｇ）の算出結果とが、学習結果データベース６６に格納される。

The result of the above correlation and the calculation result of the prior probability P (f, g) are stored in the learning result database 66.

  The computer 42 also receives a prediction request receiving unit 68 that receives a destination prediction request from a user received by the communication unit 40, and a correlation total result and a prior probability P (f, g) stored in the learning result database 66. A destination prediction unit 70 that predicts the destination of the prediction target user based on the calculation result is further provided.

  The communication unit 40 receives a destination prediction request including a user ID from the client PC 12, the mobile terminal 14, or the mobile terminal 16.

  The prediction request receiving unit 68 receives the destination prediction request received by the communication unit 40.

The destination prediction unit 70 uses the total result W ^k _f ^obtained by totaling the correlations regarding the browsing data location f with respect to the user k for the prediction target user k, for each location g of the stay destination prepared in advance. The probability of going to the place g is calculated according to the equation (12).

目的地予測部７０は、上記（１２）式で算出される確率が閾値以上となる場所ｇを特定し、特定された場所ｇを、予測される目的地として出力する。目的地の予測結果は、通信部４０により、目的地予測要求を送信したクライアントＰＣ１２、携帯端末１４、又は移動体端末１６へ送信される。

The destination prediction unit 70 identifies a place g where the probability calculated by the above equation (12) is equal to or greater than a threshold value, and outputs the identified place g as a predicted destination. The destination prediction result is transmitted by the communication unit 40 to the client PC 12, the mobile terminal 14, or the mobile terminal 16 that has transmitted the destination prediction request.

  Next, the operation of the destination prediction system 10 according to the first embodiment will be described. First, in the client PC 12, when a user ID and a password are input by a user operation, they are transmitted to the processing server 18 to enter a login state. When the Web site is browsed by the user operation in the client PC 12, the browsing data is transmitted to the processing server 18, and the browsing data is collected in the processing server 18.

  In addition, when a user ID and a password are input by a user operation on the mobile terminal 14, it is transmitted to the processing server 18 and enters a login state. When a Web site is browsed by a user operation on the mobile terminal 14, browsing data is transmitted to the processing server 18, and browsing data is collected in the processing server 18. Moreover, in the portable terminal 14, the position of the own terminal is measured at any time, and movement data indicating the measured position is transmitted to the processing server 18 every time the position is measured. In the processing server 18, movement data is collected.

  In addition, when a user ID and a password are input by the user's operation on the mobile terminal 16, it is transmitted to the processing server 18 and enters a login state. In the mobile terminal 16, the position of the own terminal is measured as needed, and each time the position is measured, movement data indicating the measured position is transmitted to the processing server 18, and the movement data is collected in the processing server 18. .

  Then, in the processing server 18, the correlation learning processing routine shown in FIG.

  In step 100, a user k to be processed is set among a plurality of users registered in advance. In the next step 102, the movement history (time series of movement data) of the user k is read from the log database 52 based on the user ID of the user k.

  In step 104, the stay location data (location data g) of the user k is extracted based on the time series of the movement data read in step 102. In the next step 106, the browsing history (time series of browsing data) of the user k is read from the log database 52 based on the user ID of the user k.

  In step 108, for each time series of the browsing data read in step 106, a place name or a product name is extracted as content from the HTML text of the browsing data website, and the place data f is used using the place name dictionary. To extract.

  In the next step 110, the correlation is calculated according to the above equation (3) for all combinations of each browsing data of the user k and each stay place data extracted in the above step 104. In step 112, it is determined whether or not the processes in steps 100 to 110 have been executed for all the users registered in advance, and there are users who have not executed the processes in steps 100 to 110. If so, the process returns to step 100 and the user is set as a user to be processed.

  On the other hand, when the processing of Step 100 to Step 110 is executed for all users, in Step 114, based on the correlation calculated for all users in Step 110, Equations (4) to (10) above. In accordance with the formula, a correlation total result is calculated and stored in the learning result database 66.

  In the next step 116, the browsing data from which the location data g is extracted according to the above equation (11) for all the combinations of the location data g and the location data f based on the correlation result calculated in the above step 114. , The prior probability P (f, g) of going to the place of stay where the place data f is extracted is calculated, stored in the learning result database 66, and the correlation learning processing routine is terminated.

  Since the stored contents of the log database 52 are updated as needed, the correlation learning processing routine is periodically executed repeatedly, and the correlation total results and the prior probabilities stored in the learning result database 66 are updated.

  Next, when a destination prediction request including the user ID is transmitted from the client PC 12, the portable terminal 14, or the mobile terminal 16 that is in the login state to the processing server 18, the computer 42 in the processing server 18 The destination prediction processing routine shown in FIG. 6 is executed.

  First, in step 130, the received destination prediction request is received, and a user to be predicted is specified. In step 132, the correlation result and the prior probability are read from the learning result database 66.

  Then, in step 134, for each location g prepared in advance as stay location data, the probability that the prediction target user will go to location g is calculated according to the above equation (12). In the next step 136, the location g where the probability calculated in the above step 134 is equal to or greater than the threshold is specified, and the specified location g is used as the prediction result of the destination of the prediction target user, and the client that has transmitted the target prediction request It transmits to PC12, the portable terminal 14, or the mobile terminal 16, and complete | finishes a destination prediction process routine.

  As described above, according to the destination prediction system according to the first embodiment, for each user, each browsing data of the browsing history and each staying destination data extracted from the movement history are associated with each other. By calculating the correlation for the combination, calculating the prior probability based on the calculated correlation, and calculating the probability of going to the location indicated by each place of stay data for the prediction target user. Even if it exists, the destination can be accurately predicted.

  Further, by predicting a place where the user wants to go from the browsing history of the Internet, it is possible to predict the future destination and position of a vehicle or a pedestrian on the road.

  Next, a second embodiment will be described. In addition, since the destination prediction system according to the second embodiment has the same configuration as that of the first embodiment, the same reference numerals are given and description thereof is omitted.

  In the second embodiment, the first embodiment is that the browsing data is clustered according to the similarity of the word sets extracted from the browsed website, and the cluster is given as the contents of the browsing data. Is different.

  The content extraction unit 60 of the processing server 18 according to the second embodiment first extracts a word set of websites from each browsing data of all users stored in the log database 52, and the similarity of word sets Based on the above, the browsing data is clustered. Note that a clustering method based on the similarity of word sets may be a conventionally known method, and thus description thereof is omitted. Further, the result of clustering is stored in a memory and used in common for all users.

  The content extraction unit 60 obtains, for each user ID, to which cluster the word set extracted from the browsing data is classified for each time series of the browsing data acquired by the browsing data acquisition unit 58, and the browsing data A cluster ID is assigned as the content f.

The correlation calculating unit 62 associates each of the browsing data i with each of the staying place data j extracted from the movement history for each user ID, and correlates the browsing data i and the staying place data j of the user k. w ^k _{i, j} is calculated according to the following equation (13).

ただし、α２は予め定められたパラメータであり、Ｔは、滞在地データの時刻ｔjと閲覧データの閲覧時刻ｔiとの時間間隔（Ｔ＝ｔｊ−ｔｉ）である。

However, α2 is a predetermined parameter, and T is a time interval (T = tj−ti) between the time tj of the stay place data and the browsing time ti of the browsing data.

According to the above equation (13), the shorter the time interval T _{, the} greater the correlation correlation w ^k _{i, j} .

  Prior probability learning unit 64 calculates a correlation total result according to the same expression as Expressions (4) to (10) above, based on the correlation between browsing data i and stay place data j calculated for each user. At the same time, when browsing the Web site of the browsing data to which the content f (cluster ID) is allocated, the prior probability P (f, g) of going to the place of stay at the location g is obtained.

The destination prediction unit 70 uses the total result W ^k _f ^obtained by totaling the correlations regarding the cluster f of the browsing data with respect to the user k for the prediction target user k, for each of the places g of the stay destinations prepared in advance. The probability of going to the place g is calculated according to an equation similar to the equation (12). The destination predicting unit 70 specifies a place g where the calculated probability is equal to or higher than a threshold, and outputs the specified place g as a predicted destination.

  In addition, about the other structure and effect | action of the destination prediction system which concern on 2nd Embodiment, since it is the same as that of 1st Embodiment, description is abbreviate | omitted.

  In this way, clustering is performed based on the word set of each browsing data in the browsing history, the browsing data is classified, and, for the prediction target user, each staying place is calculated using the correlation regarding the cluster to which each browsing data in the browsing history belongs. The probability of going to the location indicated by the data can be calculated.

  In the second embodiment described above, a case has been described in which a word set is extracted from a website and browsing data is classified based on the similarity of the word set. However, the present invention is not limited to this. Absent. It is also possible to extract feature amounts representing the content / meaning of the browsed website and classify the browse data based on the similarity or distance of the feature amounts. For example, URLs may be extracted from a website, and browsing data may be classified based on the similarity between URLs or the distance between URLs.

Further, the correlation may be calculated in consideration of the age of the history. For example, the correlation w ^k _i , _j of the prediction target user may be updated according to the following equation before calculating P (g | k) by the destination prediction unit.

ただし、α３は予め定められたパラメータであり、Ｔは、閲覧の古さ（Ｔ＝現在時刻−閲覧時刻）である。上記の式によれば、閲覧時刻が古いほど、相関が小さくなるように更新される。

Here, α3 is a predetermined parameter, and T is the age of browsing (T = current time−viewing time). According to the above formula, the correlation is updated so that the correlation is smaller as the browsing time is older.

10 Destination prediction system 12 Client PC
14 mobile terminal 16 mobile terminal 18 processing server 22 position measurement unit 26, 42 computer 28 display unit 30, 40 communication unit 32 browsing control unit 34 browsing data generation unit 36 movement data generation unit 50 data collection unit 52 log database 56 place of stay Extraction unit 60 Content extraction unit 62 Correlation calculation unit 64 Prior probability learning unit 66 Learning result database 70 Destination prediction unit

Claims (7)

For a plurality of users, history collection means for collecting document data browsing history and position information history,
Classification means for classifying the browsed document data according to the text in the document data for each document data of the collected browsing history,
Based on the collected history of location information, for each user, extraction means for extracting stay place data;
For each user, correlate each document data of the browsing history of the user with each place of stay data extracted from the history of the location information, and correlate each of the combinations of the document data and the place of stay data. Correlation calculating means for calculating;
Based on the calculated correlation, for the prediction target user, a probability of going to the place indicated by each stay place data is calculated, and a destination prediction means for predicting the destination based on the calculated probability;
Destination prediction device including Based on the correlation calculated by the correlation calculating means, for each combination of document data classification and each stay place data, to the location indicated by the stay place data when browsing the document data belonging to the classification A probability calculating means for calculating a prior probability to go;
The destination predicting means determines each destination for the prediction target user based on the correlation calculated for the prediction target user and the prior probability calculated by the probability calculation means. The destination prediction apparatus according to claim 1, wherein the destination is predicted by calculating a probability of.   The correlation calculation unit associates each document data of the browsing history of the user with each stay place data extracted from the history of the position information, and relates to each combination of the document data and the stay place data. The destination prediction apparatus according to claim 1, wherein a correlation according to a time difference between a viewing time of the document data and a stay time of the stay place data is calculated.   The correlation calculating means, for each combination of the document data and the stay place data, according to a time difference between the view time of the document data and the stay time of the stay place data, and the length of the view time of the document data The destination prediction apparatus according to claim 3, wherein the correlation is calculated based on a viewing probability and a stay probability corresponding to a length of stay time of the stay place data.   For each document data of the collected browsing history, the classification means extracts a place name or a product name from the text in the document data, and sells the location data indicating the extracted place name or the extracted product name The destination prediction apparatus according to any one of claims 1 to 4, wherein location data of a store to be assigned is assigned as a classification of the document data.   For each document data of the collected browsing history, the classification means extracts words from the text in the document data, performs clustering on the extracted set of words, and sets the extracted word set on the document data The destination prediction apparatus according to any one of claims 1 to 4, wherein a cluster to which the data belongs is assigned as a classification of the document data. Computer
History collection means for collecting document data browsing history and position information history for a plurality of users,
Classifying means for classifying the browsed document data according to the text in the document data for each document data of the collected browsing history,
Based on the collected location information history, for each user, extraction means for extracting the stay location data,
For each user, correlate each document data of the browsing history of the user with each place of stay data extracted from the history of the location information, and correlate each of the combinations of the document data and the place of stay data. Correlation calculating means for calculating, and for the prediction target user, calculating the probability of going to the location indicated by each stay place data for the prediction target user, and predicting the destination based on the calculated probability Program to function as a means.

Images