JP2010250759A - Device and method for predicting action and program - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To predict a user's action with higher precision by considering both of a place of stay and stay time and considering an appearance probability of each action pattern in the past. <P>SOLUTION: The appearance probability of the place of stay included in each piece of feature action pattern information is calculated as prior probability information about each piece of the feature action pattern, when a matched action pattern is selected, a probability that the present place of stay of the user appears in each action pattern included in an action pattern set is calculated as a posterior probability based on likelihood showing a matching degree between the place of stay to be extracted from a set of the feature action patterns and the present place of stay of the user and the prior probability information calculated beforehand, and the place of stay in which the calculated posterior probability and the likelihood satisfy preset conditions is selected. <P>COPYRIGHT: (C)2011,JPO&INPIT

Description

  The present invention relates to a behavior prediction apparatus, a behavior prediction method, and a program for predicting a future behavior of a user based on a user's current place of stay and past behavior patterns.

  Conventionally, a sequence mining technique is known as a technique for extracting frequent patterns from time-series data. This method extracts typical product purchase patterns by customers based on the purchase data recorded in the purchase record database when there is a purchase record database that records the customer ID, the purchased product, and the purchase date and time. To do.

  Recently, a method of applying the above-described series mining method has been proposed in order to extract a behavior pattern representing a transition of a user's staying place, without being limited to the use of extracting a product purchase pattern. In this method, for example, first, a location where a user stays is extracted by mining processing from GPS (Global Positioning System) received data of a terminal possessed by the user, and the transition behavior between the stay locations is expressed as a Markov model. Is used to predict a place where a user staying at a certain place will stay next (see Non-Patent Document 1, for example).

Daniel Ashbrook, Thad Starner, “Using GPS to Learn Significant Locations and Predict Movement across Multiple Users”, Peasonal and Ubiquitous Computing, 2003

  However, the above-described conventional method performs behavior prediction based on the transition probability between the places of stay, and does not consider the time, so there is a risk of erroneous prediction. For example, when a user is using a commuting station, the destination where the person goes next differs depending on whether the user stays at the station in the morning or the situation where the user stays at the same station at night. Is normal. However, if the next destination is predicted only by the transition probability between places without considering the time, the predicted result of the destination may be opposite.

  The present invention has been made paying attention to the above circumstances, and its purpose is to consider both the user's place of stay and the stay time, and to consider the probability of appearance of each past action pattern. It is an object to provide a behavior prediction apparatus, a behavior prediction method, and a program that can predict the behavior of the above with higher accuracy.

In order to achieve the above object, one aspect of the present invention comprises the following means or processes.
That is, a means or process for acquiring stay place data including information indicating a user's stay place and information indicating the stay time, and storing the acquired set of stay place data in the stay history storage section, and the stay history storage section Based on the set of staying place data stored in, the action pattern information that represents the user's characteristic transition behavior by the staying place series and the staying time information is generated, and this generated set of action pattern information is acted on The behavior pattern generation means or process stored in the pattern storage unit, and the appearance probability of the place of stay included in the behavior pattern for each of the generated behavior pattern information is calculated as prior probability information and stored in the prior probability storage unit When predicting the user's behavior and the means or process to perform, the user's current place of stay data and the behavior pattern stored in the behavior pattern recording unit The stay location corresponding to the stay location and the stay time information included in the current stay location data and the stay time information based on the set of event information and the prior probability information stored in the prior probability storage unit The adaptive behavior pattern selection means or the process for selecting the adaptive behavior pattern information including, from the set of the behavior pattern information.
Then, in the adapted behavior pattern selection means or process, a stay place corresponding to the stay place included in the current stay place data is extracted from the set of action pattern information stored in the action pattern recording unit, and this For each of the extracted stay places, a likelihood representing the degree of coincidence with the stay place included in the current stay place data is calculated. Then, based on the calculated likelihood and the prior probability information stored in the prior probability storage unit, the current place of stay appears in each behavior pattern included in the set of behavior pattern information. Probability is calculated as a posterior probability, and based on the calculated posterior probability and the likelihood, the posterior probability and the likelihood satisfying a preset condition are selected from the extracted stay locations, Information indicating the selected stay place and the action pattern including the stay place is output as the matching action pattern information.

  Therefore, according to one aspect of the present invention, when searching for an action pattern that matches a user's current action form from the user's current place of stay and a set of the user's past action patterns, the user's stay The behavior pattern is searched in consideration of not only the place but also the stay time information of the user in the stay place. For this reason, compared with the case where an action pattern is selected based only on a stay place, it becomes possible to select the action pattern which suits a user's present action form with high precision.

  In addition, for each of the generated behavior pattern information, the appearance probability of the stay place included in the action pattern is calculated as prior probability information, and the stay place extracted from the action pattern set when selecting the suitable action pattern And the user's current place of stay among the action patterns included in the action pattern set based on the likelihood representing the degree of coincidence between the user and the user's current place of stay and the prior probability information calculated in advance. The probability of appearing is calculated as a posterior probability, and a place of stay in which the calculated posterior probability and the likelihood satisfy a preset condition is selected. Here, the prior probability is higher for patterns that frequently appear and are not included in other behavior patterns. Therefore, in addition to the likelihood, by considering the prior probability of such a behavior pattern, selecting a behavior pattern that matches the user's current behavior, for example, ignoring a behavior pattern including detailed behavior such as a detour Therefore, it is possible to select an appropriate pattern and improve the accuracy of action prediction.

Furthermore, the following various embodiments can be considered as one aspect of the present invention.
In the first embodiment, when calculating the prior probability information, for each of the stay place series generated from the set of stay place data stored in the stay history storage unit, an action pattern appearing in the stay place series is displayed. Extracted from a set of behavior pattern information stored in the behavior pattern storage unit, and deletes a behavior pattern included in another behavior pattern extracted in the same manner from the extracted behavior patterns. Then, the total number of staying places included in the action pattern remaining as a result of the deletion is calculated, and the inverse of the total number is reflected in the appearance probability of the staying place included in the action pattern remaining as a result of the deletion. Is normalized based on the number of places of stay, and the appearance probability of the stay after the normalization is output as the prior probability information.
If it does in this way, it will become possible to give the appearance probability which becomes high with respect to the produced | generated action pattern information that the pattern which appears well and is not contained in another action pattern.

In the second embodiment, based on the stay place series and stay time information included in the conforming action pattern information selected by the conforming action pattern selecting means, the stay place where the user will move and its arrival time are predicted. The information representing the prediction result is output.
If it does in this way, the place where a user will move from now on, and its arrival time are outputted, for example, are shown to an information manager. For this reason, when trying to grasp the user's behavior, the information manager does not need to predict the user's destination and arrival time from the behavior pattern information, thereby improving workability and prediction accuracy. be able to.

In the third embodiment, when the behavior pattern generation means or process generates behavior pattern information including a value representing the average and variance of the stay start time or the stay end time at the stay place as the stay time information, The means or process for calculating likelihood adds a preset constant to a value representing the dispersion of the stay start time or the stay end time, and the stay time included in the user's current place data and the action pattern The likelihood is calculated based on the value representing the average of the stay start time or the stay end time at the place of stay corresponding to the current location extracted from the value and the value representing the variance to which the constant is added.
In general, when the same action is performed at the same time in the past, the temporal dispersion of this action becomes extremely small, which may cause overfit. Therefore, by adding a constant in advance to the variance value of the stay start time or the stay end time as described above, the overfit as described above is reduced, and as a result, the accuracy of selecting the suitable behavior pattern can be improved. It becomes possible.

  That is, according to the present invention, it is possible to predict the behavior of the user with higher accuracy by considering both the place of stay and the time of stay of the user and considering the appearance probability of each past behavior pattern. An apparatus, a behavior prediction method, and a program can be provided.

It is a block diagram which shows the structure of the hardware and software of the action prediction apparatus concerning one Embodiment of this invention. It is a block diagram which shows the function structure of the action prediction apparatus shown in FIG. It is a figure which shows an example of the user's stay place data obtained by the action prediction apparatus shown in FIG. It is a figure which shows an example of the user's stay place data obtained by the action prediction apparatus shown in FIG. It is a figure which shows an example of the user's stay pattern obtained by the action prediction apparatus shown in FIG. It is a flowchart which shows the procedure and content of the action pattern generation process by the action prediction apparatus shown in FIG. 1, and the characteristic action pattern prior probability calculation process. It is a block diagram which shows the 1st Example of a function structure of the appearance frequency threshold value determination means shown in FIG. It is a flowchart which shows the procedure and content of the threshold value determination process using the number of stays by the appearance frequency threshold value determination means shown in FIG. It is a figure which shows an example of the correspondence of a stay pattern set and a stay place set. It is a flowchart which shows the procedure and content of the stay count processing of the threshold value determination process using the stay count shown in FIG. It is a flowchart which shows the procedure and content of the threshold value determination process from the stay number of the threshold value determination processes using the stay number shown in FIG. It is a figure which shows an example of the relationship between appearance frequency and the number of stays. It is a block diagram which shows the 2nd Example of the function structure of the appearance frequency threshold value determination means shown in FIG. It is a flowchart which shows the procedure and content of the appearance frequency threshold value determination process using the number of patterns by the appearance frequency threshold value determination means shown in FIG. It is a figure which shows an example of the relationship between an appearance frequency threshold value, a pattern total number, and an approximate expression. It is a figure which shows an example of the relationship between an appearance frequency threshold value, a pattern total number, and an approximation error. It is a block diagram which shows the 3rd Example of a function structure of the appearance frequency threshold value determination means shown in FIG. It is a flowchart which shows the procedure and content of the appearance frequency threshold value determination process using the coverage by the appearance frequency threshold value determination means shown in FIG. It is a figure which shows an example of the action pattern produced | generated by the action pattern production | generation means shown in FIG. It is a flowchart which shows the procedure and content of the characteristic action pattern prior probability calculation process in the flowchart shown in FIG. It is a flowchart which shows the whole procedure and content of the action prediction process by the action prediction process means shown in FIG. It is a flowchart which shows the procedure and content of the adaptation action pattern search process in the flowchart shown in FIG. It is a flowchart which shows the procedure and content of the destination prediction process in the flowchart shown in FIG.

Embodiments according to the present invention will be described below with reference to the drawings.
FIG. 1 is a block diagram showing the hardware and software configurations of an action prediction apparatus according to an embodiment of the present invention.
The behavior prediction apparatus 1 is composed of a server device capable of communicating with a user terminal device such as a portable terminal or a personal computer possessed by a user via a communication network such as an IP (Internet Protocol) network.

  The behavior prediction apparatus 1 includes a central processing unit (CPU) 11 composed of a microprocessor, a program memory 13 and a data memory 14 are connected to the CPU 11 via a bus 12, and a communication interface 15 and an input / output are connected. The interface 16 is connected.

  The communication interface 15 receives the stay location data from the user's portable terminal under the control of the CPU 11 in accordance with a communication protocol defined by the communication network. For example, TCP / IP (Transmission Control Protocol / Internet Protocol) is used as the communication protocol.

  An input device 17 and an output device 18 are connected to the input / output interface 16. The input device 17 includes a keyboard and a mouse. The output device 18 includes a display and a printer. The input / output interface 16 transmits the operation information of the input device 17 to the CPU 11, and the information indicating the prediction result of the user's next action obtained by the action prediction processing program 135 described later, according to the instruction of the CPU 11, the output device 18. To display or print out. The information representing the prediction result of the user's next action can be transmitted from the communication interface 15 to the user's terminal device or the service provider's terminal device permitted by the user.

  The data memory 14 includes a stay history storage unit 141, a characteristic behavior pattern storage unit 142, a prior probability storage unit 143, and a prediction result storage unit 144 as storage units necessary for carrying out the present invention. Yes. The stay history storage unit 141 stores a set of stay place data that is received from the user terminal device and represents the stay place where the user stayed and the stay time. In the behavior pattern storage unit 142, a stay pattern (feature) corresponding to a behavior pattern characteristic (typical) for the user, extracted from a set of stay patterns obtained by executing a series mining execution program described later. Action pattern) is stored. The prior probability storage unit 143 stores the appearance probability for each place of stay calculated by the prior probability calculation program described below for each of the stored characteristic behavior patterns. In the prediction result storage unit 144, a stay place selected from the stored characteristic stay patterns and having a high degree of coincidence with the current stay place and stay time of the user, and an action pattern including the stay place Identification information (behavior pattern ID) is stored.

  In the program memory 13, as the application programs necessary for carrying out the present invention, the stay place input control program 131, the characteristic action pattern generation program 132, the characteristic action pattern prior probability calculation program 133, and the action prediction processing program 134 is stored.

  The stay place input control program 131 causes the communication interface 15 to receive the stay place data from a user terminal device (not shown) via the communication network, and the received stay place data is stored in the stay history storage unit 141 in the data memory 14. The CPU 11 is caused to execute the storing process. The stay place data is composed of the unique name of the stay place and the stay time information. The unique name of the place of stay is obtained by searching the map database based on the latitude and longitude information of the user terminal device obtained using a position sensor such as GPS (Global Positioning System) or RFID (Radio Frequency IDentification). Is possible. As the map database, one stored in advance in the user terminal device may be used, or when not stored, one disclosed on the map data server may be used. The stay time information includes a stay start time and a stay end time. The stay location input control program 131 also includes a processing function for reading the stay location data from an external storage medium such as a memory card and storing it in the stay history storage unit 141.

The characteristic behavior pattern generation program 132 includes a series mining execution program, an appearance frequency threshold determination program, and a behavior pattern determination program.
Of these, the sequence mining execution program reads a set of stay place data stored in the stay history storage unit 141 and creates a set of stay patterns by applying a sequence mining technique to the read set of stay place data. Then, the CPU 11 is caused to execute a process of storing the set of stay patterns in a stay pattern storage unit (not shown) in the data memory 14.

The appearance frequency threshold value determination program causes the CPU 11 to execute the following series of processes.
(1) Each time the appearance frequency threshold candidate θ is increased by a certain amount within the range of 0 to 1, a stay pattern whose appearance frequency exceeds the candidate from the set of stay patterns stored in the stay pattern storage unit The process to select.
(2) A process of extracting all stay places included in the set from the selected set of stay patterns.
(3) A process of counting the number of stay places that match the extracted stay place from the set of stay place data stored in the stay history storage unit 141.
(4) Based on the count value of the place of stay obtained for each candidate for the appearance frequency threshold, a section where the count value does not change with respect to the change of the appearance frequency threshold candidate is obtained. A process of selecting an appearance frequency threshold value used for extracting the characteristic behavior pattern from among appearance frequency threshold candidates included in.

The behavior pattern determination program causes the CPU 11 to execute the following processing.
(1) A stay pattern that exceeds the appearance frequency threshold from a set of stay patterns stored in the stay pattern storage unit based on the appearance frequency threshold determined by execution of the appearance frequency threshold determination program Extraction process.
(2) Processing for obtaining an average stay start time, an average stay end time, a dispersion of stay start times and a dispersion of stay end times for each stay place included in the extracted stay pattern.
(3) A characteristic behavior pattern composed of the extracted stay pattern and the obtained average stay start time, average stay end time, stay start time variance, and stay end time variance is associated with a pattern ID. Processing to be stored in the characteristic behavior pattern storage unit 142 in the data memory 14.

  The characteristic behavior pattern prior probability calculation program 133 calculates the appearance probability of the place of stay included in the characteristic behavior pattern for each of the characteristic behavior pattern information stored in the characteristic behavior pattern storage unit 142, and the calculated value Is stored in the prior probability storage unit 143 as prior probability information.

The behavior prediction processing program 134 causes the CPU 11 to execute the following processing.
(1) When the user's current place of stay data is obtained under the control of the place of stay input control program 131, the stay place whose name matches the place of stay included in the stay place data is selected as the characteristic behavior pattern. Processing to extract from the characteristic behavior pattern stored in the storage unit 142.
(2) A process of calculating a likelihood representing the degree of coincidence with the current stay place data for each of the extracted stay places. This likelihood calculation is performed based on the average stay start time or average stay end time and the variance value of the stay start time or stay end time. At this time, a constant set in advance to reduce overfit is added to the dispersion value of the stay start time or the stay end time.
(3) Based on the calculated likelihood and the prior probability of the characteristic behavior pattern stored in the prior probability storage unit 143 in the data memory 14, the current place of stay of the user is A process of calculating a posterior probability indicating a probability corresponding to a pattern.
(4) A process of selecting, from the extracted stay places, stay places where the calculated likelihood and posterior probability satisfy the preset conditions.
(5) Predicting the user's next destination and the arrival time to the destination based on the selected destination, the stay pattern that includes this place of stay, and the stay time information, A process of outputting the prediction result to the output device 18 via the input / output interface 16.

FIG. 2 is a block diagram showing a functional configuration of the behavior prediction apparatus 1 shown in FIG. The function of the behavior prediction apparatus 1 is roughly divided into an input unit 21, a recording unit 22, a processing unit 23, and an output unit 24.
The input unit 21 includes a stay location input unit 211. This stay place input means 211 corresponds to the stay place input control program 131 and the communication interface 15 shown in FIG.
The output unit 24 includes a prediction result output unit 241. The prediction result output unit 241 corresponds to the input / output interface 16, the output device 18, and the communication interface 15.

  The recording unit 22 includes a stay history storage unit 221, a characteristic behavior pattern storage unit 222, a prior probability storage unit 223, and a prediction result storage unit 224. The stay history storage unit 221 corresponds to the stay history storage unit 141 of FIG. 1, and the characteristic behavior pattern storage unit 222 corresponds to the characteristic behavior pattern storage unit 142. The prior probability storage unit 223 corresponds to the prior probability storage unit 143 in FIG. 1, and the prediction result storage unit 224 corresponds to the prediction result storage unit 144.

  The processing unit 23 includes a characteristic behavior pattern generation unit 231, a characteristic behavior pattern prior probability calculation unit 232, and a behavior prediction processing unit 233. Among these, the characteristic behavior pattern generation unit 231 includes a sequence mining execution unit, an appearance frequency threshold value determination unit, and a behavior pattern determination unit. Further, the feature behavior pattern prior probability calculation means 232 corresponds to the feature behavior pattern prior probability calculation program 133 shown in FIG. 1, and the behavior prediction processing means 234 corresponds to the behavior prediction processing program 134.

  FIG. 7 is a block diagram showing the functional configuration of the appearance frequency threshold value determining means 2311 in more detail. That is, the appearance frequency threshold value determining unit 2311 includes an appearance frequency threshold value exceeding pattern selecting unit 2321, a stay location set calculating unit 2322, a stay number counting unit 2323, and a threshold value determining unit 2324. .

Next, the operation of the behavior prediction apparatus configured as described above will be described. 6 and 21 are flowcharts showing the entire processing procedure and processing contents.
(1) Input of data on the place of stay
CPU11 starts the stay place input control program 131 in step S51. Then, according to this stay place input control program 131, the stay place data is received from the user terminal device via the communication network by the communication interface 15, and the received stay place data is stored in the stay history storage unit 141 in the data memory 14. Remember me.

  The stay place data is data related to the stay place where the user stayed for a certain period of time or more, and is generated by the stay place detection function of the user terminal device, for example. The stay location detection function represents a unique name of the point from the map database based on the function of measuring the position of the terminal using, for example, GPS or RFID, and the position information including the latitude and longitude measured by this function And a function for searching for information. The stay place data includes a stay place name, a stay start time, and a stay end time. FIG. 3 shows an example of the stay place data, and shows data of “house” and “A station”.

のように表される。 In addition, stay series data can be generated by arranging the plurality of stay places in the order of movement of the user. The stay series data includes a place name column and a series number. At this time, if the stay sequence is d and the sequence number is i, the stay sequence di having the sequence number i is

It is expressed as

で表される。 Here, uij is called the j-th place of stay (or stay) of series i. The value of uij represents the name of the place where the user stayed jth in series i. Mi is the total number of stays in sequence i. FIG. 4 shows an example of a stay sequence data set for N sequences stored in the stay sequence storage unit 141.

It is represented by

(2) Series mining processing
Subsequently, the CPU 11 starts up the series mining execution program in step S52. Then, according to the sequence mining execution program, a set of stay sequence data is read from the stay history storage unit 141, and a sequence mining process is executed on the set of stay sequence data to create a set of stay patterns. The set of stay patterns is stored in a stay pattern storage unit (not shown) in the data memory 14.

Sequence mining is a method of receiving a set of stay sequence data and selecting from the stay sequence data set all stay patterns whose appearance frequency is equal to or greater than a certain number or whose appearance frequency is greater than or equal to a certain speed. Note that the appearance frequency refers to the total number of sequences in the stay sequence data set in which the stay pattern appears when a certain stay pattern and stay sequence data set are given, and the total number of sequences included in the stay sequence data set. Is the ratio. The sequence mining process is described in detail in the following documents.
Rakesh Agrawal and Ramakrishnan Srikant, "Mining Sequential Patterns", 11 th International Conference on Data Engineering, pp 3-14,1995.

と表される。ｖijは滞在パターンｐiに出現するｊ番目の滞在地であり、Kiは滞在パターンｐiの滞在地の総数である。 FIG. 5 shows an example of a stay pattern stored in a stay pattern storage unit (not shown) in the data memory 14, and the stay pattern is composed of a pattern number, the number of appearances thereof, and a stay column. The pattern number is a number uniquely assigned to one stay pattern, and the stay column is a list of places of stay. At this time, if the stay pattern is p and the pattern number is i, the pattern pi having a certain pattern number i is

It is expressed. vij is the jth place of stay that appears in the stay pattern pi, and Ki is the total number of places of stay of the stay pattern pi.

  The number of occurrences is defined as follows. That is, the appearance of a pattern is a concept defined between a stay pattern and a stay sequence, and 1 ≦ α (1) <such that vai = ubα (i) for each stay pattern pa and stay sequence db. When α (2) <... α (Ka) ≦ Mb exists, the pattern pa is defined as appearing in the stay sequence db.

  For example, the stay pattern <house, school, school> appears in the stay series <house, station, school>, but does not appear in <house, station, school>. Use this occurrence definition to define the number of appearances. The number of appearances of the stay pattern pj refers to the number of the stay series data included in the stay series data set accumulated in the stay series data storage unit 141 in which the pattern pj appears. The number of appearances of the pattern pi is expressed as O (pi).

のように表される。 A set of stay patterns is defined as follows. That is, the number of rows (L rows) shown in FIG. 5 is the total number of stay patterns created by the series mining process, and the set P of all created stay patterns is

It is expressed as

(3) Appearance frequency threshold determination processing
Next, in step S53, the CPU 11 starts an appearance frequency threshold value determination program, and executes an appearance frequency threshold value determination process according to the appearance frequency threshold value determination program. There are three possible methods for determining the appearance frequency threshold.

(3-1) First method
FIG. 8 is a flowchart showing a processing procedure and processing contents for realizing the method for determining the appearance frequency threshold.
That is, the CPU 11 first sets the appearance frequency threshold value candidate θ to the minimum value θ = 0 in step S71. Subsequently, in step S72, a stay pattern set is read from the stay pattern storage unit in the data memory 14, and the appearance frequency threshold candidate value θ is selected from a plurality of stay patterns included in the read stay pattern set. Select all stay patterns that exceed.

  Next, the CPU 11 proceeds to step S73, and extracts a stay place set from the selected stay pattern set. A stay place set is defined as a set of all stay places included in a stay pattern included in this set when a set of stay patterns stored in the stay pattern storage unit is given. FIG. 9 shows an example of this. When the set of stay patterns given is {<house, cram school>, <house, bus stop>, <house>, <bus stop>}, the set of stay patterns. Will extract {house, cram school, bus stop} as a place of stay.

  Subsequently, the CPU 11 proceeds to step S74 and performs a process of counting the number of staying places that match the extracted staying place from the set of staying series data stored in the staying series storage unit 141 as follows. To run. FIG. 10 is a flowchart showing the processing procedure and processing contents.

  That is, the CPU 11 fetches the stay place set extracted in step S73 in step S91, fetches the stay series data set from the stay series data storage unit in step S92, and initializes the count value count to 0 in step S93. To do.

  In this state, the CPU 11 first selects one stay place from one of the stay series data included in the stay series data set in step S94, and an element equal to the selected stay place is included in the stay place set. It is determined in step S95 whether or not it exists. As a result of the determination, if there is an element equal to the selected stay place in the stay place set, the process proceeds to step S96 and the count value count is incremented. On the other hand, if it is determined that it does not exist, the count value is not incremented.

  Next, in step S97, the CPU 11 determines whether or not there is a staying place that has not been checked yet in the stay series data set. If there is, the process returns to step S94 and selects one unselected staying place. . Then, in step S95, it is determined whether or not an element equal to the selected stay place is present in the stay place set, and if it exists, the count value count is incremented. In the same manner, each time one unselected stay place is selected from the stay series data set, it is determined whether or not an element equal to the stay place exists in the stay place set. The process of incrementing the count value count is repeated.

  When the determination process and the increment process of the count value count based on the determination result are finished for all the stay places included in the stay series data set, the CPU 11 proceeds from step S97 to step S98, and the count The value of the value count, that is, the number of times each element included in the stay location set when the appearance frequency threshold candidate value θ = 0 appears in the stay series data set is output. The output count value of the number of stays is stored in a stay count value storage unit (not shown) in the data memory 14 in association with the appearance frequency threshold candidate value θ = 0.

として表される。 The result of the stay count processing described above is expressed as follows. In other words, when the counted number of stays is a function s (θ) of the appearance frequency threshold θ, the function s (θ) is

Represented as:

When the stay count processing when the appearance frequency threshold candidate value θ = 0 is completed, the CPU 11 determines whether or not the appearance frequency threshold candidate value θ has reached the maximum value “1”. The determination is made at S75. As a result of this determination, if θ has not reached “1”, the appearance frequency threshold candidate value θ is increased by a fixed amount (eg, 0.01) in step S76, and the process returns to step S72. Then, this time, the stay frequency counting process in steps S72 to S74 is executed using the increased appearance frequency threshold candidate value θ. Thereafter, the stay count process in steps S72 to S74 is repeatedly executed every time θ is increased by a certain amount until the appearance frequency threshold candidate value θ reaches “1”.
As a result of the above stay count processing, information indicating the change in the stay count with respect to the change in the appearance frequency threshold candidate is obtained in the stay count count storage unit in the data memory 14. FIG. 12 shows an example.

  When the appearance frequency threshold candidate value θ reaches “1” which is the maximum value, the CPU 11 proceeds to step S77, and changes in the appearance frequency threshold candidate θ obtained in steps S72 to S74. Based on the information indicating the change in the number of stays with respect to, a process of selecting one optimum appearance frequency threshold to be used for the characteristic stay pattern extraction process is executed as follows. FIG. 11 is a flowchart showing the processing procedure and processing contents.

That is, in step S101, the CPU 11 first reads the stay number count value obtained for each appearance frequency threshold candidate θ (0 to 1) from the stay number count value storage unit in the data memory 14. Next, in step S102, all sections where the stay count value does not change with respect to the change in the appearance frequency threshold candidate θ are detected based on the read count value of the stay place. Subsequently, in step S103, of all the detected sections, the section including the appearance frequency threshold candidate θ = 1 is first deleted, and the section in which the range of θ is the longest among the remaining sections is determined. Select one. In step S104, the lower limit value of θ in the selected longest interval is determined as an appearance frequency threshold value used for extracting a characteristic action pattern.
For example, in the example of FIG. 12, an interval where θ = 0.1 to 0.5 is selected as the longest interval, and θ = 0.1 which is the lower limit value in the longest interval is used for extracting a characteristic action pattern. Determine as the frequency threshold.

  As described above, in the first method for determining the appearance frequency threshold value, the appearance frequency threshold value candidate θ is stored in the stay pattern storage unit 142 every time it is increased by a certain amount in the range of 0 to 1. A stay pattern whose appearance frequency exceeds the candidate is selected from the set of stay patterns, all stay places included in the set are extracted from the selected set of stay patterns, and the stay series storage unit 141 The number of stay places that coincide with the extracted stay place from the set of stored stay series data is counted. Then, based on the count value of the place of stay obtained for each candidate for the appearance frequency threshold, a section where the count value does not change with respect to the change of the candidate for the appearance frequency threshold is obtained. The lower limit value of the longest section excluding the section including θ = 1 is selected as the appearance frequency threshold value used for extracting the characteristic action pattern.

  This first method has the following advantages. In other words, there is a tendency for the number of stays to change more generally in regions where the appearance frequency threshold is low, but this reflects the large change in the set of stay patterns that have an appearance frequency equal to or greater than the appearance frequency threshold. Yes. The stay pattern included in such a set is likely to be noise that does not reflect the user's daily behavior, and is not suitable for use in evaluating the user's daily behavior. On the other hand, when the appearance frequency threshold is high, the number of frequent patterns obtained by itself becomes small, and it is difficult to appropriately evaluate the user's daily behavior in consideration of the diversity.

  On the other hand, when the first method is used, as described above, the longest interval in which the number of stays does not change with respect to the change in the appearance frequency threshold, that is, the set of appearing stay patterns is most stable and includes noise. An appearance frequency threshold is selected from a section with a low possibility. For this reason, there is no worry that the appearance frequency threshold is set low. In addition, when the appearance frequency threshold is increased from 0, the longest interval in which the number of stays does not change often appears immediately after the region where the number of stays changes rapidly with respect to the change in the appearance frequency. An appearance frequency threshold is set as the lower limit value of the section. For this reason, there is no concern that the appearance frequency threshold is set to a high value.

  That is, according to the first method for determining the appearance frequency threshold, it is possible to set the appearance frequency threshold to an appropriate value that is not too low and not too high, thereby reducing the influence of noise. It is possible to sufficiently extract the stay pattern necessary for evaluating the daily behavior of the user while keeping the minimum.

(3-2) Second method
The second method for determining the appearance frequency threshold value is to increase the appearance frequency threshold candidate value θ every time the candidate value θ of the appearance frequency threshold value is increased by a certain amount in the range of 0 to 1. Means or a method for counting the number of stay patterns exceeding and determining an appearance frequency threshold value using the count value of the number of stay patterns is used.

  FIG. 13 is a block diagram showing a functional configuration of the appearance frequency threshold value determining means for carrying out the second method for determining the appearance frequency threshold value. In the figure, the same parts as those in FIG. Further, FIG. 1 will be described with respect to the software and hardware configuration of the behavior prediction apparatus, and FIG. 2 will be described with respect to the functional configuration of the entire apparatus.

As shown in FIG. 13, the appearance frequency threshold value determination means 2312 includes an appearance frequency threshold value excess pattern selection means 2321, a pattern number count means 2325, an approximate error calculation means 2326, and a minimum error threshold detection means 2327. It consists of.
The appearance frequency threshold value excess pattern selection unit 2321 increases the appearance frequency threshold value candidate θ from the set of stay patterns stored in the stay pattern storage unit 142 every time the candidate frequency threshold value θ is increased by a certain amount within a range of 0 to 1. A stay pattern in which the appearance frequency exceeds the candidate is selected.
The pattern number counting unit 2325 counts the number of stay patterns selected by the appearance frequency threshold value exceeding pattern selection unit 2321 for each candidate value θ of each appearance frequency threshold value.

The approximate error calculation means 2326 defines an approximate expression representing a change in the count value of the stay pattern with respect to a change in the appearance frequency threshold candidate value θ.
The error minimum threshold value detecting means 2327 obtains an error between the approximate expression defined above and the count value of the stay pattern, and characterizes the candidate value θ of the appearance frequency threshold value when the approximation error is minimized. Is selected as an appearance frequency threshold value used for extraction of a simple action pattern.

Next, the threshold value determination processing operation by the appearance frequency threshold value determination unit 2312 configured as described above will be described. FIG. 14 is a flowchart showing the processing procedure and processing contents. In the figure, the same parts as those in FIG.
That is, the CPU 11 first sets the appearance frequency threshold value candidate θ to the minimum value θ = 0 in step S71. Subsequently, in step S72, a stay pattern set is read from the stay pattern storage unit 142 in the data memory 14, and the appearance frequency threshold candidate value θ is selected from a plurality of stay patterns included in the read stay pattern set. Select all stay patterns that exceed. Next, the process proceeds to step S131, the number of stay patterns selected in step S72 is counted, and the count value is stored in the data memory 14.

  When the selection of the number of stay patterns and the process of counting the number of stay patterns when the appearance frequency threshold candidate value θ = 0 is completed, the CPU 11 determines that the appearance frequency threshold candidate value θ is “1”, which is the maximum value. In step S75, it is determined whether or not the value has been reached. As a result of this determination, if θ has not reached “1”, the appearance frequency threshold candidate value θ is increased by a fixed amount (eg, 0.01) in step S76, and the process returns to step S72. Then, using this increased appearance frequency threshold candidate value θ, the process of selecting the stay pattern and counting the number of stay patterns in steps S72 and S131 is executed. Thereafter, similarly, every time the appearance frequency threshold candidate value θ is increased by a certain amount until the appearance frequency threshold candidate value θ reaches “1”, which is the maximum value, the stay pattern in step S72 and step S131. The process of selecting and counting the number is repeatedly executed.

として定義される。 Next, the CPU 11 initializes a variable φ set separately in step S132 to 0, and then defines a function fφ (x) in step S133. This function fφ (x) has three points (0, R () when the count value of the stay pattern corresponding to each appearance frequency threshold candidate value θ obtained up to step S132 is R (θ). 0)), (θ, R (θ)), (1, R (1)) are connected in order by line segments

Is defined as

として定義される。ここで、０＝θ1 ＜θ2 ＜…＜θT ＝１である。 Subsequently, in step S134, the CPU 11 obtains an approximate error between the function fφ (x) defined above and the count value R (θ) of the stay pattern for each θ. The approximate error G (φ) corresponding to the parameter φ is

Is defined as Here, 0 = θ1 <θ2 <... <θT = 1.

The meaning of this approximate error G (φ) will be described with reference to FIG. FIG. 15 is a plot of the function R (θ) of the appearance frequency threshold candidate value θ and the count value of the stay pattern, and the function fφ (x) when φ = 0.2. The shaded area in the figure is an error when the function R (θ) of the count value of the stay pattern is approximated by the function fφ (x). The stay pattern count value tends to be large when the threshold value is small and small when the threshold value is large, with respect to the change in the appearance frequency threshold candidate value θ. . Therefore, when the count value of the stay pattern is approximated by the function fφ (x), the error is minimized when θ is present at the boundary between the portion where the change in the count value of the stay pattern is large and the small portion. In other words, the smaller error means that the value of the parameter φ is approaching the boundary of the slope change in the function R (θ) of the count value of the stay pattern.
When the approximate error is calculated in step S134, the CPU 11 temporarily stores the calculated approximate error in the data memory 14 in association with the value of φ in step S135.

  When the stay number counting process when the parameter φ = 0 is completed, the CPU 11 determines in step S136 whether or not the parameter φ has reached 1, which is the maximum value. If φ does not reach 1 as a result of this determination, the parameter φ is increased by a fixed amount (eg, 0.01) in step S137, and the process returns to step S133. This time, the increased φ is used to execute the approximation error calculation process in steps S133 to S135. Similarly, until the parameter φ reaches 1, every time the parameter φ is increased by a certain amount, the approximation error calculation process in steps S133 to S135 is repeatedly executed.

  When the approximation error for all φs from 0 to 1 is calculated by the approximation error calculation process, the CPU 11 calculates the smallest approximation error from among the calculated approximation errors for all φ in step S138. Then, φ at which this error is minimized is output as an appearance frequency threshold value used for feature pattern extraction.

  The reason for selecting φ as the appearance frequency threshold when the approximation error is minimum is as follows. FIG. 16 is a diagram in which the appearance frequency parameter φ is changed and the change in the count value R (φ) of the stay pattern and the change in the approximation error G (σ) are plotted simultaneously. From the figure, it can be seen that the appearance frequency parameter φ that minimizes the approximation error G (σ) is close to the boundary of the change in the stay pattern count value R (φ). Since it can be considered that such a region is the lower limit value of the appearance frequency threshold value where the stay pattern is stable, this φ is output as the appearance frequency threshold value used for extracting the feature pattern.

  As described above, in the second method for determining the appearance frequency threshold value, the appearance frequency threshold value is increased each time the candidate value θ of the appearance frequency threshold value is increased by a certain amount in the range of 0 to 1. The number of stay patterns exceeding the candidate value θ is counted. Each time the separately set appearance frequency parameter φ is increased by a certain amount within the range of 0 to 1, an approximation error between the count value of the number of stay patterns and the approximation function fφ (x) is obtained, and this approximation error is minimized. Is selected as an appearance frequency threshold value used for feature pattern extraction.

  Therefore, also in the second method, as in the first method described above, the appearance frequency threshold is set to the lower limit value of the appearance frequency threshold where the stay pattern is stable. For this reason, it is possible to set the appearance frequency threshold value to an appropriate value that is not too low and not too high, thereby minimizing the influence of noise and evaluating the user's daily behavior. Necessary stay patterns can be extracted sufficiently.

(Third method)
A third method for determining the appearance frequency threshold value is to increase the appearance frequency threshold candidate value θ every time the candidate value θ of the appearance frequency threshold value is increased by a certain amount in the range of 0 to 1. The number of stay patterns exceeding the selected number of stay patterns is extracted from all the stay places included in the set, and the count of the stay places relative to the total number of stay places included in the stay series data set is selected. The ratio of the values (coverage) is obtained, and the appearance frequency threshold is determined based on this coverage.

  FIG. 17 is a block diagram showing a functional configuration of the appearance frequency threshold value determining means 2313 for carrying out the third method. In the figure, the same parts as those in FIG. Also in this embodiment, FIG. 1 is used for the software and hardware configuration of the behavior prediction apparatus, and FIG. 2 is used for the functional configuration of the entire apparatus.

As shown in FIG. 17, the appearance frequency threshold value determination means 2313 includes an appearance frequency threshold value exceeding pattern selection means 2321, a stay location set calculation means 2322, a stay count counting means 2323, a coverage rate calculation means 2328, And a threshold value determining means 2324.
Of these, the coverage calculation means 2328 is the stay location obtained by the stay count counting means 2323 with respect to the total number of stay locations included in the set of stay sequence data stored in the stay sequence data storage unit (stay sequence storage means). The ratio of the count value is calculated as the coverage.

Next, the threshold value determination processing operation by the appearance frequency threshold value determination unit 2313 configured as described above will be described. FIG. 18 is a flowchart showing the processing procedure and processing contents. In the figure, the same parts as those in FIG.
That is, the CPU 11 first sets the appearance frequency threshold value candidate θ to the minimum value θ = 0 in step S71. Subsequently, in step S72, a stay pattern set is read from the stay pattern storage unit in the data memory 14, and the appearance frequency threshold candidate value θ is selected from a plurality of stay patterns included in the read stay pattern set. Select all stay patterns that exceed.

  Next, the CPU 11 proceeds to step S73, and extracts a stay place set from the selected stay pattern set. A stay place set is defined as a set of all stay places included in a stay pattern included in this set when a set of stay patterns stored in the stay pattern storage unit is given. Subsequently, the CPU 11 proceeds to step S74, and counts the number of stay places that match the extracted stay place from the set of stay series data stored in the stay series storage unit 141.

のように表される。 Next, in step S161, the CPU 11 covers the ratio of the count value of the stay places obtained in steps S72 to S74 to the total number of stay places included in the set of stay series data stored in the stay series data storage unit. Calculate as a rate. Now, if the coverage is σ (θ), the coverage is

It is expressed as

  Next, in step S162, the CPU 11 determines an appearance frequency threshold value used for feature pattern extraction based on the calculated coverage σ (θ). The coverage is monotonously non-increasing with respect to the appearance frequency threshold similarly to the number of stays described in the first method. For this reason, if the candidate value of the appearance frequency threshold value is increased and the coverage is once lower than a certain value, the coverage does not exceed that value no matter how much the candidate value is increased thereafter. Therefore, if a threshold value is set for the coverage ratio, when the candidate frequency threshold value is increased, the point at which the coverage ratio falls below the threshold value for the first time is detected. Can do.

For this reason, a coverage threshold value of 0.3, for example, is set in advance for the coverage. Then, when the appearance frequency threshold candidate value is increased and the coverage ratio exceeds the set threshold value, the maximum value of the appearance frequency threshold candidate value at this time is determined as a feature pattern. Is output as an appearance frequency threshold value used for extraction.
By determining the appearance frequency threshold value used for feature pattern extraction in this way, a set of characteristic stay patterns can be created so that the behavior coverage is constant regardless of the nature of the user's stay series data. It becomes possible to extract.

(4) Characteristic behavior pattern generation processing
When the appearance frequency threshold is determined by any one of the first to third methods, the CPU 11 proceeds to step S54 and activates the behavior pattern determination program. Then, according to the action pattern determination program, first, a set of stay patterns is read from the stay pattern storage unit, and the determined appearance frequency threshold is exceeded from each stay pattern included in the read set of stay patterns. A stay pattern is extracted. Further, a time feature amount is obtained for each stay location included in the extracted stay pattern. This time feature amount includes an average stay start time, an average stay end time, a dispersion of stay start times, and a dispersion of stay end times. In step S55, the extracted stay pattern and the obtained time feature amount are stored in the feature behavior pattern storage unit 142 in the data memory 14 in association with the pattern ID.

  FIG. 19 shows an example of characteristic behavior pattern information stored in the characteristic behavior pattern storage unit 142 as described above. In the figure, for the sake of simplicity, only two stay patterns are illustrated, and only time feature amounts corresponding to the first stay place 1 appearing in the stay pattern are shown.

(5) Prior probability calculation processing
When the characteristic action pattern of the user is generated as described above, the CPU 11 next activates the characteristic action pattern prior probability calculation program 133 in step S60, and the generated characteristic action pattern according to the action prediction processing program 133. The process for calculating the appearance probability of the stay place included in the pattern, that is, the prior probability is executed as follows. FIG. 20 is a flowchart showing the processing procedure and processing contents.

  That is, the CPU 11 first reads a set of stay place series data from the stay place series data storage unit in the data memory 14 in step S181, and further sets a set of feature action patterns from the feature action pattern storage unit 142 in the data memory 14 in step S182. Is read. The read set of stay place series data and the set of characteristic behavior pattern information are stored in a work memory in the data memory 14. Subsequently, in step S183, the CPU 11 initializes the appearance probability of the stay place included in the pattern for each of the read characteristic behavior patterns to 0.

  Next, in step S184, the CPU 11 selects one place of stay series data from the read set of stay place series data, and in step S185, the behavior pattern appearing in the selected place of stay series data is selected. All the extracted feature behavior patterns are extracted. Subsequently, in step S186, all the action patterns that are partial patterns of other action patterns included in the set are deleted from the action patterns included in the extracted set of action patterns. In step S187, the total number of staying places included in the action pattern remaining after the deletion process is calculated, and this calculated value is set as NUM. Finally, in step S188, the reciprocal of the total number of NUM stays, that is, 1 / NUM is added to each stay place included in the action pattern remaining after the deletion process. CPU11 performs the process by step S184-S188 mentioned above with respect to all the stay place series data contained in the said set of stay place series data.

  Then, when it is confirmed in step S189 that the processing for all stay place series data has been completed, the CPU 11 proceeds to step S190. Then, the appearance probability of the stay place included in each of the read feature action patterns is normalized by dividing by the number of stay place series included in the set of stay place series data. The appearance probability of the place of stay included in the pattern is stored in the prior probability storage unit 143 in the data memory 14 as prior probability information.

(6) Behavior prediction processing
As described above, a set of user's characteristic behavior patterns is generated in (4), and the appearance probability of the stay place included in the generated characteristic behavior pattern in (5), that is, the prior probability is calculated. Then, the CPU 11 activates the behavior prediction processing program 134 and executes processing for predicting the user's next behavior according to the behavior prediction processing program 134 as follows. FIG. 21 is a flowchart showing the overall processing procedure and processing contents.

  That is, the CPU 11 first in step S201, when the latest stay place data received from the user terminal device changes with respect to the last received place place data stored in the stay history storage unit 141 in step S201. The latest stay place data, that is, data representing the user's current stay place is read from the stay history storage unit 141. The latest stay place data includes at least the stay start time. At the same time, the CPU 11 reads a set of characteristic behavior pattern information from the characteristic behavior pattern storage unit 142 in step S202. The set of the stay location data and the characteristic behavior pattern information indicating the read current position of the user is temporarily stored in the work memory in the data memory 14.

  Next, in step S203, the CPU 11 executes a process for searching for a characteristic behavior pattern that matches the stay location data representing the read user's current position as follows. FIG. 22 is a flowchart showing the processing procedure and processing contents.

(6-1) Matching action pattern search process
First, in step S301, the CPU 11 reads the name of the latest place of stay data from the work memory, and in step S302, sequentially reads a set of action pattern information from the work memory. In step S303, all feature action patterns including the stay place name of the latest stay place data are extracted from the stay pattern, and a set of the extracted feature action patterns is created on the work memory. For example, if the name of the place of stay in the latest place of stay data (the current place of stay of the user) is “A station”, all feature behavior patterns having a stay pattern including this “A station” are extracted. As a result of this extraction processing, if it is determined in step S304 that the set of feature action patterns is “empty”, that is, if the feature action pattern including the stay place name of the latest stay place data cannot be extracted, the CPU 11 The matching action pattern search process is terminated.

のように定義される。また、このとき分散の値σ_ij ² には、オーバーフィットを低減するために、予め設定した定数εを付加する。 On the other hand, if the action pattern information including the name of the place of stay of the latest place of stay data can be extracted, the CPU 11 proceeds to step S305, extracts one feature action pattern from the set of feature action patterns, and then continues to step S306. From the stay place patterns of the extracted characteristic behavior pattern, one stay place having the same name as the stay place corresponding to the current position of the user is selected. In step S306, the likelihood of the stay start time included in the latest stay place data is calculated using the average and variance values of the stay start times corresponding to the selected stay places. The likelihood p (l | v _ij ) is the stay start time of the current stay place t _l , and the average and variance values of the stay start times of the stay places selected from the stay patterns of the characteristic behavior pattern are μ, respectively. _ij and σ _ij ²

Is defined as follows. At this time, a preset constant ε is added to the dispersion value σ _ij ² in order to reduce overfit.

In addition, in the calculation of the likelihood P (l | v _ij ), the stay end time may be used in addition to the stay start time of the current place of stay. In this case, the same calculation is performed for the stay end time using the above formula, the product of the calculation result and the calculation result for the stay start time is obtained, and the calculation result is defined as likelihood p (l | v _ij ). .

ここで、ｐ(ｌ)は滞在地ｌについての事前確率であり、ある単一のlに対しては定数値となる。以降の処理においてはｐ(ｌ)を未知の正の定数としたうえで、事後確率ｐ(ｖ_ij｜ｌ)の定数倍を事後確率ｐ(ｖ_ij｜ｌ)の代わりに用いる。 Next, in step S307, the CPU 11 calculates the likelihood p (l | v _ij ) calculated above and the prior probability p of the characteristic behavior pattern previously calculated in (5) and stored in the prior probability storage unit 143. Based on (v _ij ), the posterior probability p (v _ij | l) is calculated. The following formula is used to calculate the posterior probability p (v _ij | l).

Here, p (l) is a prior probability for the place of stay l, and is a constant value for a single l. In terms of the p and (l) an unknown positive constant in the subsequent processes, the posterior probability p is used in place of | | (l v _ij) a constant multiple of (v _ij l) posterior probability p.

Next, in step S308, the CPU 11 compares the calculation result of the posterior probability p (v _ij | l) with a preset threshold value, and the posterior probability p (v _ij | l) is the maximum so far. It is determined whether or not. If it is determined that the posterior probability p (v _ij | l) is the maximum, then the calculation result of the likelihood p (l | v _ij ) is compared with a threshold value set in advance. It is determined whether it is above. As a result of this determination, if the value of the likelihood p (l | v _ij ) is greater than or equal to a threshold value, the CPU 11 performs the likelihood p (l | v _ij ) and the posterior probability p (v _ij |) in step S309. The working memory of the data memory 14 is used as a candidate of the behavior pattern information corresponding to the user's current place of stay, using the number of the place of stay calculated as l) and the pattern ID of the action pattern information including the stay place in the stay pattern. Save to. Then, the process proceeds to step S310.

On the other hand, when the posterior probability p (v _ij | l) is not the maximum and when the value of the likelihood p (l | v _ij ) is less than the threshold, the CPU 11 stores the behavior pattern information. The user's current place of stay is not saved as a corresponding candidate, and the process proceeds to step S310. In step S <b> 310, it is determined whether there is another stay place having the same name as the current stay place of the user in the stay pattern of the behavior pattern information. As a result of this determination, if there are others, the process returns to step S305 to select one other staying place having the same name as the current staying place. Thereafter, the likelihood p (l | v) is obtained through steps S306 to S309. and calculation processing _ij), the posterior probability p (v _ij | executes the calculation processing l), the determination process behavior pattern candidate based on these calculation results.

In addition, in the search process of the suitable behavior pattern candidate based on the likelihood p (l | v _ij ) and the posterior probability p (v _ij | l), the value of the likelihood p (l | v _ij ) When a plurality of stay places are found, the name of the stay place and the pattern IDs of all characteristic action patterns including the stay place in the stay pattern are stored as action pattern information corresponding to the user's current stay place. It may be. Further, without setting a threshold value, a stay location where the value of the posterior probability p (v _ij | l) is maximized and a pattern ID of one action pattern information including the stay location in the stay pattern are displayed as the user's current You may make it preserve | save as action pattern information applicable to the place of stay.

  When the selection of all the stay places having the same name as the current stay place included in the stay pattern of the behavior pattern information is completed, the CPU 11 proceeds from step S310 to step S311. Then, it is determined whether or not there is a feature behavior pattern not yet selected in step S305 in the set of feature behavior patterns including the current place of stay created in step S303. If there is a feature action pattern that has not been selected as a result of this determination, the process returns to step S303 to select one unselected feature action pattern. And the process by said step S305-step S310 is similarly performed about this selected action pattern information.

  Then, when the processes in steps S305 to S311 are completed for all the characteristic behavior patterns included in the set of characteristic behavior patterns created in step S303, the CPU 11 proceeds from step S311 to step S312 and the data The number of the stay place and the pattern ID of the action pattern information including the stay place in the stay pattern, which are finally stored as candidates in the work memory in the memory 14, are output as the search result of the matching action pattern.

(6-2) Destination prediction process
When the process for searching for a suitable behavior pattern is completed, the CPU 11 next determines in step S204 in FIG. 21 whether or not behavior pattern information suitable for the user's current place of stay has been found. If the result of this determination is that it is not found, the process is terminated as it is, but if it is found, the process proceeds to step S205 and the destination prediction process is executed as follows. FIG. 23 is a flowchart showing the processing procedure and processing contents.

  That is, first in step S401, the CPU 11 uses the number of the stay place finally obtained by the adaptive action pattern search process and the pattern ID of the action pattern information including the stay place in the stay pattern for the work in the data memory 14. Read from memory. In step S402, the corresponding behavior pattern information is read from the behavior pattern storage unit 143 based on the read pattern ID, and the stay location series in the stay pattern included in the behavior pattern information and the user's current Based on the number of the above-mentioned place of stay determined to correspond to this place of stay, the place of stay that will be the next destination of the user is predicted.

  As a prediction method, there are multiple methods depending on whether the prediction is based only on the stay pattern, the prediction is based on the stay pattern and the stay time information, or the prediction of multiple stay places is permitted as the destination. Can be considered. If prediction of multiple destinations is allowed, simply enter the next corresponding destination of the user's current place of stay on each of the stay patterns for all of the entered matching behavior pattern information. Each staying place is selected, and the set of the selected staying places is set as the destination set. On the other hand, when only one destination is allowed to be predicted, for example, from the set of stay destinations selected as the destination, the stay destination having the earliest average value of the stay disclosure times is set as the destination prediction result. In addition, by referring to the time information, for example, among the stay places arranged after the stay place corresponding to the current stay place of the user in each stay pattern of the plurality of conforming behavior pattern information, the average of the stay start times All values whose values are within a certain time (for example, within 30 minutes) from the current time are used as the predicted results of the destination. If it is considered that the user is acting according to the pattern, this place of stay is likely to stay by the user.

  Finally, in step S206 (step S403), the CPU 11 stores the destination predicted as described above and the average value of the stay start times at the destination in the prediction result storage unit 144 in the data memory 14, The data is output from the output device 18 via the input / output interface 16.

  As described above, in this embodiment, based on the current location data of the user and a set of information representing a characteristic behavior pattern created in advance for the user, from this set of behavior pattern information, Action pattern information including a stay place having the same name as the current place of stay of the user and a stay start time or an end time of the stay place having a likelihood equal to or greater than a threshold and the stay place. select. Then, the user's next destination is predicted based on the selected stay place and action pattern information.

  Therefore, not only the user's place of stay but also the user's stay time at the place of stay is considered, and an action pattern that matches the user's current place of stay is searched. For this reason, compared with the case where an action pattern is selected based only on a stay place, it becomes possible to select the action pattern which suits a user's present action form with high precision.

  In addition, for each piece of characteristic behavior pattern information, the appearance probability of the place of stay included in the characteristic behavior pattern is calculated as prior probability information, and the stay extracted from the set of characteristic behavior patterns when selecting a suitable behavior pattern Based on the likelihood indicating the degree of coincidence between the place and the current place of stay of the user and the prior probability information calculated in advance, the current place of stay of the user in each action pattern included in the action pattern set Is calculated as a posterior probability, and a place of stay in which the calculated posterior probability and the above likelihood satisfy a preset condition is selected. Here, the prior probability is higher for patterns that frequently appear and are not included in other behavior patterns. Therefore, in addition to the likelihood, by considering the prior probability of such a behavior pattern, selecting a behavior pattern that matches the user's current behavior, for example, ignoring a behavior pattern including detailed behavior such as a detour Therefore, it is possible to select a suitable pattern, and it is possible to improve the accuracy of behavior prediction.

Further, when the likelihood p (l | v _ij ) is calculated, a preset constant ε is added to a value σ _ij ² representing the dispersion of the stay start time or the stay end time of the characteristic behavior pattern information. . In general, when the same action is performed at the same time in the past, the temporal dispersion of this action becomes extremely small, which may cause overfit. However, by adding a constant ε in advance to the dispersion value σ _ij ² of the stay start time or stay end time as described above, the overfit as described above is reduced, and as a result, the accuracy of selection of the conforming action pattern is reduced. Can be increased.

  In addition, the user's next destination and its arrival time are output and presented to the information manager. For this reason, when trying to grasp the user's behavior, the information manager does not need to perform the work of predicting the user's destination and arrival time from the compatible behavior pattern information, thereby improving workability and predicting. The accuracy can be improved.

として定義することができる。非日常的な滞在地を発見することはユーザの行動を理解することにつながる。 Moreover, according to this embodiment, it is also possible to discover an unusual place of stay in the process of obtaining the number of stays. An extraordinary place of stay is a place of stay that is not included in any set of characteristic stay patterns. Assuming that a set of patterns having an appearance frequency equal to or higher than an appearance frequency threshold φ is Qφ 非 P, a set E _i of extraordinary staying places included in the stay series d _i on a certain day is:

Can be defined as Discovering an extraordinary place of stay leads to understanding the user's behavior.

  The present invention is not limited to the above embodiment. For example, the case where the function as the behavior prediction device is provided in the server device has been described as an example, but the function as the behavior prediction device may be provided in the user terminal device such as the user's portable terminal or personal computer. Further, a plurality of functions constituting the behavior prediction device, that is, the feature behavior pattern generation unit 231 and the feature behavior pattern prior probability calculation unit 232, and the behavior prediction processing unit 233 are divided, and a part of them is assigned to the server device. May be provided in a user terminal device such as a portable terminal or a personal computer. That is, a plurality of means for realizing the functions constituting the behavior prediction device may be provided in a distributed manner in the server device and the user terminal device.

  In addition, the type and configuration of the behavior prediction device, the configuration of the characteristic behavior pattern generation unit 231 and the characteristic behavior pattern prior probability calculation unit 232 and the processing procedure and processing content thereof, the behavior prediction processing unit 233, particularly the procedure of the adaptive behavior pattern search processing The processing contents and the like can be variously modified and implemented without departing from the gist of the present invention.

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

  DESCRIPTION OF SYMBOLS 1 ... Action prediction apparatus, 11 ... CPU, 12 ... Bus, 13 ... Program memory, 14 ... Data memory, 15 ... Communication interface, 16 ... Input / output interface, 17 ... Input device, 18 ... Output device, 131 ... Stay place input Control program, 132 ... feature action pattern generation program, 133 ... feature action pattern prior probability calculation program, 134 ... action prediction processing program, 141 ... stay history recording section, 142 ... feature action pattern recording section, 143 ... prior probability storage section, 144 ... Prediction result storage unit, 21 ... Input unit, 211 ... Stay place input unit, 22 ... Recording unit, 221 ... Stay history storage unit, 222 ... Characteristic behavior pattern recording unit, 223 ... Prior probability storage unit, 224 ... Prediction result Storage means, 23... Processing unit, 231... Feature action pattern generation means, 232. Pattern generation means, 233... Action prediction processing means, 2311, 212, 2313 ... appearance frequency threshold value determination means, 2321 ... frequency threshold value exceeding pattern selection means, 2322 ... stay place set calculation means, 2323 ... stay count counting means 2324... Threshold value determining means, 2325... Pattern number counting means, 2326... Approximate error calculating means, 2327... Minimum error threshold detecting means, 2328.

Claims (9)

Means for acquiring stay place data including information indicating the stay place of the user and the stay time, and storing a set of the obtained stay place data in the stay history storage unit;
Based on a set of stay place data stored in the stay history storage unit, action pattern information that represents the user's characteristic transition behavior by a stay place series and stay time information is generated. Action pattern generation means for storing a set of action pattern information in an action pattern storage unit;
For each of the generated behavior pattern information, the appearance probability of the place of stay included in the behavior pattern is calculated as prior probability information, and the calculated prior probability information is stored in the prior probability storage unit;
When predicting the user's behavior, the current location data of the user, a set of behavior pattern information stored in the behavior pattern recording unit, and prior probability information stored in the prior probability storage unit, Based on the above, the staying place included in the current staying place data and the staying action corresponding to the staying time information and the matching action pattern information including the staying time information are selected from the set of the action pattern information. Action pattern selection means,
The conforming behavior pattern selection means includes:
Means for extracting a place of stay corresponding to a place of stay included in the current place of stay data from a set of action pattern information stored in the action pattern recording unit;
Means for calculating a likelihood representing a degree of coincidence with a stay place included in the current stay place data for each of the extracted stay places;
Based on the calculated likelihood and the prior probability information stored in the prior probability storage unit, the probability that the current place of stay appears in each behavior pattern included in the set of behavior pattern information Is calculated as a posterior probability, and based on the calculated posterior probability and the likelihood, a means of selecting a staying place satisfying a preset condition of the posterior probability and likelihood from the extracted staying places When,
An action prediction apparatus comprising: means for outputting the selected stay place and information representing an action pattern including the stay place as matching action pattern information. The means for calculating the prior probability information includes:
For each stay place series generated from a set of stay place data stored in the stay history storage unit, an action pattern appearing in the stay place series is obtained from a set of action pattern information stored in the action pattern storage unit. Means for extracting;
Means for deleting a behavior pattern included in another behavior pattern extracted in the same manner from the extracted behavior patterns;
Means for calculating the total number of staying places included in the behavior pattern remaining as a result of the deletion;
Means for reflecting the reciprocal of the calculated total number of staying places in the appearance probability of staying places included in the behavior pattern remaining as a result of the deletion;
Means for normalizing the appearance probability of the stay place included in the extracted behavior pattern based on the number of the stay place series, and outputting the appearance probability of the stay place after the normalization as the prior probability information; The behavior prediction apparatus according to claim 1, further comprising:   Based on the stay place sequence and stay time information included in the fit action pattern information selected by the fit action pattern selection means, the stay place where the user will move and its arrival time are predicted, and this prediction result is expressed. The behavior prediction apparatus according to claim 1, further comprising behavior prediction processing means for outputting information. When the behavior pattern generation means generates behavior pattern information including a value representing the average and variance of the stay start time or the stay end time in the place of stay as the stay time information,
The means for calculating the likelihood is:
Means for adding a preset constant to a value representing dispersion of the stay start time or stay end time;
For each of the extracted stay places, the likelihood indicating the degree of coincidence with the stay place included in the current stay place data is set as the stay time included in the current stay place data, and the extracted stay. 4. The behavior predicting apparatus according to claim 3, further comprising means for calculating based on a value representing an average of the stay start time or stay end time on the ground and a value representing the variance to which the constant is added. A process of acquiring stay place data including information indicating the stay place of the user and the stay time, and storing a set of the acquired stay place data in the stay history storage unit;
Based on a set of stay place data stored in the stay history storage unit, action pattern information that represents the user's characteristic transition behavior by a stay place series and stay time information is generated. A behavior pattern generation process for storing a set of behavior pattern information in a behavior pattern storage unit;
For each of the generated behavior pattern information, calculating the appearance probability of the place of stay included in the behavior pattern as prior probability information, and storing the calculated prior probability information in the prior probability storage unit;
When predicting the user's behavior, current place of stay data for the user, a set of behavior pattern information stored in the behavior pattern recording unit, and prior probability information stored in the prior probability storage unit, Based on the above, the staying place included in the current staying place data and the staying action corresponding to the staying time information and the matching action pattern information including the staying time information are selected from the set of the action pattern information. An action pattern selection process,
The adaptive behavior pattern selection process includes:
A process of extracting a place of stay corresponding to a place of stay included in the current place of stay data from a set of action pattern information stored in the action pattern recording unit;
For each of the extracted stay destinations, calculating a likelihood representing the degree of coincidence with the stay place included in the current stay place data;
Based on the calculated likelihood and the prior probability information stored in the prior probability storage unit, the probability that the current place of stay appears in each behavior pattern included in the set of behavior pattern information Is calculated as a posterior probability, and based on the calculated posterior probability and the likelihood, a place of stay where the posterior probability and likelihood satisfy a preset condition is selected from the extracted stay places When,
A behavior prediction method comprising: a step of outputting information indicating the selected stay place and an action pattern including the stay place as matching action pattern information. The process of calculating the prior probability information includes:
For each stay place series generated from a set of stay place data stored in the stay history storage unit, an action pattern appearing in the stay place series is obtained from a set of action pattern information stored in the action pattern storage unit. Extraction process,
A process of deleting a behavior pattern included in another behavior pattern extracted in the same manner from the extracted behavior patterns;
Calculating the total number of places to stay included in the behavior pattern remaining as a result of the deletion;
A process of reflecting the reciprocal of the calculated total number of staying places in the appearance probability of the staying place included in the behavior pattern remaining as a result of the deletion;
Normalizing the appearance probability of the stay place included in the extracted behavior pattern based on the number of stay place series, and outputting the appearance probability of the stay place after the normalization as the prior probability information; The behavior prediction method according to claim 5, further comprising:   Based on the staying place sequence and staying time information included in the matching action pattern information selected by the conforming action pattern selection process, the staying place where the user will move and its arrival time are predicted, and this prediction result is expressed. The behavior prediction method according to claim 5 or 6, further comprising a behavior prediction process for outputting information. When the behavior pattern generation process is to generate behavior pattern information including a value representing the average and variance of the stay start time or stay end time in the stay place as the stay time information,
The process of calculating the likelihood includes:
Adding a preset constant to a value representing dispersion of the stay start time or stay end time;
For each of the extracted stay places, the likelihood indicating the degree of coincidence with the stay place included in the current stay place data is set as the stay time included in the current stay place data, and the extracted stay. The behavior prediction method according to claim 7, further comprising a step of calculating based on a value representing an average of the stay start time or stay end time on the ground and a value representing the variance to which the constant is added.   A program that causes a computer included in the behavior prediction apparatus to execute processing corresponding to each unit included in the behavior prediction apparatus according to any one of claims 1 to 4.

Images