JP5392679B2 - Decision analysis server, decision analysis method, program, and decision analysis system - Google Patents

Description

  The present invention relates to an intention analysis technique using sensitivity engineering.

  In recent years, diversification and variety of products and digital contents and interests and purposes have diversified, and digital contents such as clothes, miscellaneous goods, and web pages that some people are interested in are also diversified (hereinafter, products) -Refer to the entity to be selected, such as digital content, as the recognition target.) The subject who accessed the recognition target (hereinafter referred to as the access subject) has interest / interest such as “see”, “touch”, “take in hand”, “browse”, “download” with respect to the recognition target. Specific actions to reflect the will based on

  For example, when the access subject is interested only to a certain extent that the recognition target is “there”, the action of the access subject with respect to the recognition target is considered to end with only “see”. On the other hand, if the access subject is interested in the recognition target more than “the recognition target is there”, the action is more “approaching”, “touching”, “holding”, etc. It can be said that it develops into an action that increases closeness. Therefore, it is considered that by analyzing the behavior of the access subject to the recognition target, it is possible to analyze the intention state that cannot be judged from the appearance such as the degree of interest / interest in the recognition subject of the access subject.

  Corresponding to the diversification of the values of access subjects, we have estimated "what kind of products the person is interested in" based on the purchase history of a specific person, Attempts have also been made to use it for information recommendation and marketing. However, in the case of information recommendation or marketing based on the purchase history, products similar to those already purchased are often presented. In such a case, the access subject may already have lost interest because it already has the object to be presented, and providing such information would be rather annoying. There is also.

  In addition, when a new product or the like is developed based on the purchase history, there is a possibility that a product that is not different from the product currently on the market is produced. For such problems, quantitative analysis of intriguing elements in products, etc. is performed by data mining the intention of the access subject to the action to the recognition target in the middle of the decision process of the access subject. In addition, it is expected to be possible to carry out with high accuracy.

  There has been known a technique for performing management by analyzing buyers' interest in products and the like. For example, in Japanese translations of PCT publication No. 2008-511066 (patent document 1), after detecting the position of the customer in the store using RFID, the shopping time consumed in the area in the store is checked, and the position of the customer is determined. There has been proposed a system for analyzing a customer's shopping flow by tracking and analyzing a customer's shopping pattern using this.

  Japanese Patent Laid-Open No. 2001-331628 (Patent Document 2) includes a server device, a mobile terminal device, and an advertising medium, and each advertising medium is given a unique mark. By photographing this mark with the device and sending the image data to the server device, the server device identifies each advertisement medium from the image data sent from the mobile terminal device and collects it in association with the time at that time A system for obtaining marketing data by registering in a data DB has been proposed.

  Further, in Japanese Patent Application Laid-Open No. 2003-114968 (Patent Document 3), the number of visits to a customer's product display shelf is counted once for the same product category such as the same product category, and is not excessively counted. Therefore, even if multiple customers visit the same category of products on multiple product display shelves, this is counted as a single product category, so it is possible to grasp customer trends by product category unit. An analysis system has been proposed.

  Also, Philip Kotler, translated by Maki Tsukiya: Introduction to Kotler Marketing, Pearson Education, pp.179-185, October 2002, ISBN-10: 978-4894716582 (Non-Patent Document 1) The state of the mind of a person in the process of purchasing a product includes (1) a problem recognition stage, (2) an information retrieval stage, (3) an evaluation stage, (4) a decision stage, and (5) an evaluation stage for purchased products. It is analyzed that there is.

Special table 2008-511066 gazette JP 2001-331628 A JP 2003-114968 A

Philip Kotler, translated by Tsukiya Maki: Introduction to Kotler Marketing, Pearson Education, pp.179-185, October 2002, ISBN-10: 978-4894716582

  As described above, Patent Documents 1 to 3 describe a technique for performing marketing based on an action to a recognition target such as a product by an access subject, and Non-Patent Document 1 describes an interest and behavior of an access subject. The relationship between is described. However, Patent Documents 1 to 3 and Non-Patent Document 1 disclose a technique for analyzing the behavior of the access subject reflecting the internal interest / interest in the recognition subject of the access subject from the behavior manner of the access subject to the recognition subject, In addition, it is not possible to conduct marketing that predicts the future, such as product development based on the analysis.

  In addition, with conventional technology, if the action does not proceed at least to the decision stage, it is not possible to analyze the interests and interests of products, etc., and the actions of the access subject that did not lead to purchase are effectively reflected in marketing etc. It does not make it possible to make it happen.

  That is, the present invention analyzes a behavior of an access subject with respect to a recognition target, analyzes an access subject's intention level from the behavior of the access subject, and enables an access subject's intention prediction with respect to a specific recognition target. An object is to provide a decision analysis method and program, and a decision analysis system.

The present invention has been made in view of the above-mentioned problems of the prior art. In the present invention, the intention analysis server analyzes the intention of the access subject from the behavior of the access subject, and the intention analysis server includes:
When the access subject approaches the recognition target, the access subject identifies the closeness of the access subject's behavior mode performed on the recognition target while the access subject approaches the recognition target, and the behavior An action determination unit that generates an action vector that associates the duration of the action with the action pattern;
An action normalization unit that normalizes element values of the action vector and normalizes actions of the access subject with respect to the recognition target;
An action cluster unit that classifies the action vector normalized by the action normalization unit into a plurality of clusters based on a closest access action of the access subject in the action mode;
The element values of the action vector attributed to the cluster generated by the action cluster unit are ranked based on the numerical range assigned to the assigned cluster, and the ranking value is a condition attribute, An action ranking unit that generates a rough set having a classification attribute as a decision attribute;
And a rule generation unit that generates a maximum likelihood rule that reduces the condition attribute of the rough set and gives the decision attribute.

  The behavior determination unit of the present invention can level the behavior pattern of the access subject with respect to the recognition target in association with closeness using the duration of the behavior pattern and image analysis. The behavior cluster unit of the present invention can execute cluster classification by hierarchical clustering using the Ward method.

  Furthermore, the intention analysis server of the present invention can include a database for managing the rules generated by the rule generation unit and the data table used by the behavior determination unit, the behavior ranking unit. In the present invention, the intention analysis server acquires information on the behavior of the access subject from a intention analysis device connected via a network, and the intention analysis device sends the information to the recognition subject of the access subject. An access detection unit that detects the approach of the access, an image change detection unit that detects that the access subject has acted on the recognition target, and sends a video image to the intention analysis server; and the intention analysis server And an intention analysis unit that issues an inquiry and executes an analysis of the intention of the access subject.

  Furthermore, the present invention can provide an intention analysis method, a program, and an intention analysis system using the intention analysis server having the above configuration.

The figure explaining the data definition 100 of the basic method of intention analysis of this embodiment. The figure which showed schematic structure of the action observation system 200 which observes the action of an access subject in a ubiquitous environment in order to acquire the action vector 140 of the access subject 202 in this embodiment. The functional block diagram of the intention analysis system 300 of this embodiment. The figure which showed embodiment of the data table which the database 324 stores in this embodiment. The flowchart of the process which registers the action of the access subject 202 which the intention analysis apparatus 230 of this embodiment performs into the database 324. FIG. The figure which showed embodiment of the data table of each process part which the database 324 manages. The figure which showed the flowchart of the process which the action cluster part of this embodiment performs. The figure explaining the cluster classification | category process demonstrated in FIG. 7 using the data structure. The figure explaining the process which the action rank attaching part 316 of this embodiment performs using the data structure 900. FIG. The figure which showed embodiment of the rule which the rule production | generation part 318 demonstrated in FIG. 9 produced | generated by the reduction process. The figure which showed the simulation shop and sensor structure which are used by this experiment. The figure which showed the result of the action pattern obtained by this experiment. The figure which showed the result table 1300 which assumed a normal distribution for the appearance frequency of the action style in each cluster classification which shows a test subject's intention stage, and was processed statistically from the data of the category classification shown in FIG. The figure which showed the table which extracted partly the result which contrasted the stage evaluation of decision-making and the questionnaire result to a test subject based on the rule attachment which took the posture by rough set.

  Hereinafter, although this invention is demonstrated with embodiment, this invention is not limited to embodiment mentioned later. FIG. 1 is a diagram for explaining a data definition 100 of a basic method of intention analysis according to this embodiment. The access subject changes the behavior pattern according to the degree of interest / interest for recognition objects such as products and digital contents that can be selected according to preference. The behavior pattern is characterized by the behavior of the access subject and the duration of the behavior.

  In many cases, in a store or the like, there are more access entities that view or pick up products than access entities that purchase products. For this reason, even if the access entity does not lead to the purchase of products etc., it is possible to guide the next purchase of the same access entity by effectively using the behavior of the access entity. It is possible to connect interests and preferences for the recognition target to subsequent marketing. For this purpose, in the present embodiment, the behavior pattern of the access subject with respect to the recognition target is analyzed, and the intention of the access subject with respect to the recognition target is analyzed from the behavior pattern.

  In this embodiment, in order to separate the behavior of the access subject from the activity of the behavior itself, the behavior of the access subject is defined as a normalized behavior vector 140 as shown in FIG. It shall be reflected as a projection onto the coordinate axis corresponding to the intention level. In the embodiment where the intention analysis in the marketing is performed in the present embodiment, the action vector 140 does not reach the determination stage of (4), (1) problem recognition, (2) information search, and (3) evaluation. It can be formulated as a three-dimensional vector having three elements. Note that the dimension of the action vector may be two-dimensional or may be formulated as three or more dimensions depending on the specific embodiment.

  The action vector 140 shown in FIG. 1 has a value projected on each coordinate axis in the coordinate axis system such as the problem recognition axis 110, the information search axis 120, and the evaluation axis 130 that reflects the intention level. In this embodiment, the value of each coordinate axis is quantified by using a proximity wireless sensor such as an image analysis by a camera and an RFID tag. In this embodiment, each element value of the action vector 140 is used to classify the action of the access subject into clusters corresponding to the intention level of the access subject. Then, by assigning an intensity level for the same action in each cluster, a rough set condition attribute is generated from the observed action pattern, and the cluster classification is used as a decision attribute, which is used for decision rule generation. . The behavior of the access subject is ruled based on the behavior vector shown in FIG. 1, and the degree of the degree of interest such as the interest or preference of the access subject based on the behavior of the specific access subject based on the rule Can be estimated.

  FIG. 2 shows a behavior observation system 200 for observing the behavior of an access subject in a ubiquitous environment in order to acquire the behavior vector 140 of the access subject 202 in this embodiment (Naoki Imamura, Sadano Kanno, Shunichi Kato, Ubitaki Inter It is the figure which showed the outline composition of consumer behavior pattern using the face, Kansei Engineering Society paper, 8,3,23,741-747,2009). The access subject 202 moves in the store 210 along the route 204. Then, the access subject 202 stops when there is a recognition target 212 such as a product that catches eyes, and responds to the recognition target 212 such as a T-shirt placed on the display shelf 208 according to the intention based on the interest of the access subject 202. It approaches a recognition target 212 such as a product.

  At this time, the access subject 202 increases the closeness of the action according to the intention based on the interest and interest, and takes an action pattern such as “see”, “touch”, and “take in hand”. In the present embodiment, the access subject 202 has the RFID 216 for the purpose of eliminating, as noise, an action with a low degree of interest when walking around the route 204. The RFID 216 can be embedded in a medium for providing a purchase benefit such as a member's card or a service card, and can identify a specific individual as an access subject ID or the like.

  On the other hand, a camera 218 is installed in the store 210 in association with a specific product, and monitors the state of the recognition target 212. The camera 218 is a digital camera capable of acquiring a moving image such as a video camera or MOVING MPEG. Information of the camera 218 and the proximity wireless sensor 214 is sent to the intention analysis device 230 via the network 220. The intention analysis device 230 stores an image acquired at the end of the monitoring process immediately before the camera 218, and when a difference image is detected from the image, the video image after the detection of the difference image is displayed together with time data in the present embodiment. In the form, data to be registered is sent to an externally connected intention analysis server.

  The intention analysis device 230 issues an SQL query to the intention analysis server or the like via the display device 234, the keyboard 236, and the mouse 238, and receives the response. The intention analysis device 230 stores data and a program necessary for the processing in a hard disk device (HDD) 232 and reads the data into an execution space such as a RAM in response to a call from the MPU to enable the processing. Note that the video image from the camera 218 may be continuously sent to the intention analysis server if the capacity of the intention analysis server (not shown) allows. In this case, the detection of the approach of the access subject 202 holding the RFID to the recognition target 212 is triggered to start the processing of this embodiment.

  FIG. 3 is a functional block diagram of the intention analysis system 300 of the present embodiment. The intention analysis system 300 shown in FIG. 3 can be configured as a server that includes a database function as a whole, or the right-hand side functional blocks in FIG. 3 are separated as a intention analysis server and implemented as a distributed processing system. You can also. The specific mounting format of the intention analysis system 300 can be appropriately set according to the capability and purpose of the computer device to be used.

  The intention analysis system 300 illustrated in FIG. 3 includes a intention analysis device 230 and a intention analysis server 340 that acquires various types of information from the intention analysis device and executes intention analysis.

  The intention analysis device 230 can be implemented using a personal computer or a workstation, and the microprocessor (MPU) may include any single-core processor or multi-core processor known so far. The OS may be controlled by any known OS such as WINDOWS (registered trademark), UNIX (registered trademark), LINUX (registered trademark), or MAC OS.

  Further, the intention analysis device 230 accesses Internet Explorer (registered trademark), Mozilla (registered trademark), Opera (registered trademark), and FireFox (registered trademark) in order to access a intention analysis server connected via a network such as the Internet. ) And other browser software are implemented to enable file transmission based on file transfer protocols such as HTTP and FTP. In addition, the intention analysis device 230 has a database access component such as JDBC (Java Database Connectivity) in order to access an externally connected intention analysis server using SQL. It is preferable to send various sentences to the intention analysis server to access various databases.

  In addition, the intention analysis server 340 includes a CISC architecture microprocessor such as PENTIUM (registered trademark), XEON (registered trademark), and a PENTIUM (registered trademark) compatible chip, or a RISC architecture microprocessor such as POWERPC (registered trademark). Can be implemented in single core or multi-core form. In addition, the intention analysis device 230 is controlled by an operating system (OS) such as WINDOWS (registered trademark) 200X, UNIX (registered trademark), LINUX (registered trademark), C, C ++, JAVA (registered trademark), A server program such as CGI, servlet, APACHE, or IIS (Internet Information Server) implemented using a programming language such as PERL or RUBY is executed, and data from the RFID 216 or the camera 218 is processed.

  Further, the intention analysis server 340 implements a database application in order to register, search, and process data using SQL (Structured Query Language) or the like. Although the database is not particularly limited, it is preferable to use a relational database in terms of versatility. For example, DB2 (registered trademark), Oracle (registered trademark), MySQL, ACCESS (registered trademark), PostgreSQL, etc. are used. can do. In addition, if the application environment permits, an object-oriented database (OODB), XMLDB, or the like can be used as appropriate.

  In the embodiment shown in FIG. 3, the intention analysis system 300 is configured such that a intention analysis device 230 including a intention analysis unit 320 and a intention analysis server 340 including a DB access unit 322 are connected via a second network 330. ing. However, in other embodiments, as long as the processing overhead is not a problem and the capacity of the information processing apparatus is sufficient, the intention analysis apparatus 230 and the intention analysis server 340 may be functionally integrated and implemented. it can.

  More specifically, the intention analysis apparatus 230 of FIG. 3 includes a network interface 304, and performs data transmission via the network 302 using a transmission protocol such as Ethernet (registered trademark) or FTTH. . In the embodiment shown in FIG. 3, the network 302 is configured as a local area network (LAN) that receives data from the store 210 or the like. Furthermore, the intention analysis device 230 includes an access detection unit 306 and an image change detection unit 308.

  The access detection unit 306 is an access subject for uniquely identifying the access subject 202 to be sent by detecting that the access subject 202 having the RFID 216 has approached by the proximity wireless sensor 214 installed on the display shelf 208 of the store 210. Receive an identification value (ID). Further, the access detection unit 306 receives position information where the proximity wireless sensor 214 is installed from the proximity wireless sensor 214, sends each received data to the intention analysis server 340 and registers it in the database 324.

  The image change detection unit 308 stores the last video image of the previous camera operation acquired by the camera 218, and compares the video image sent from the network 302 with the stored image image. It is determined whether or not the video image acquired by the camera 218 has changed by determining that the recognition target 212 is touched or the action of picking up the recognition target 212. When the image change detection unit 308 determines that the video image has changed, the image change detection unit 308 sends the video image together with time information to the intention analysis server unit 340 and registers it in the database 324.

  Furthermore, the intention analysis device 230 includes an intention analysis unit 320. The intention analysis unit 320 accesses the database 324 using SQL or the like, refers to the rules registered in the database 324, analyzes the behavior of the access subject, and reflects the internal interest of the access subject 202 The intention state is determined, and the result is displayed on the display device 234 of the decision making device 230 using an appropriate graphical user interface (GUI) or output to a page printer.

  Further, the intention analysis server 340 shown in FIG. 3 executes a database application such as PostgreSQL and processes an inquiry such as SQL. More specifically, the intention analysis server 340 acquires the access subject ID, location information, time data and corresponding video image from the intention analysis device 230, calls the DB access unit 322, and accesses the access subject ID, location information, And video images are described in the respective fields to be registered.

  The intention analysis server 340 includes an action determination unit 310 in order to execute the decision making process of the present embodiment. The behavior determination unit 310 includes an image analysis program. The behavior determination unit 310 analyzes the video image sent from the intention analysis device 230 and registered in the database 324, and characterizes the behavior pattern. More specifically, the behavior determination unit 310 identifies the behavior of the access subject 202 approaching the specific recognition target 212 with respect to the recognition target 212, and “sees”, “touches” the result of the behavior determination in the database 324, It is classified and registered into behavior patterns associated with closeness based on the distance to the recognition target that reflects the intention to the recognition target, such as “take in hand”. Furthermore, the behavior determination unit 310 registers the time length for which the access subject 202 has performed the behavior in association with the behavior.

  In the present embodiment, for a specific purpose of selecting a product or the like, the determination criteria associated with the closeness to the recognition target such as “see”, “touch”, and “take in hand” are given in Table 1 below. Table 1 below can be registered as a data table of the database 324.

  Note that, in the embodiment of product purchase, “see” is the access behavior having the lowest closeness, and “take in hand” is the access behavior having the highest closeness. Further, for example, in an embodiment of a search result from a search engine on an Internet site, “Search” is the access action with the lowest closeness, and the action of matching the mouse cursor with the search result is the next closest. Access behaviors such as “browsing search results” and “downloading” can be set as access behaviors having the highest closeness. In other embodiments, as long as the action is based on the intention of the access subject, it is possible to appropriately set a judgment criterion related to closeness. The action vector of this embodiment is defined in association with the time length action measured for these access actions.

  Further, the decision making database 340 includes a behavior normalization unit 312. The behavior normalization unit 312 extracts a behavior determination result for a specific access subject ID from the database 324, and normalizes the behavior vector. To do. Various normalization methods can be assumed, but in the present embodiment, since there is a weight of the behavior style on the coordinate axis of the behavior vector, the ratio of the time length of the specific behavior to the total access time is defined as the following formula (1 ) To normalize the action vector. The action vector of the access subject is registered in the database 324 as a table for each recognition target with the access subject ID as an index.

  Further, the intention analysis server unit 340 includes an action cluster unit 314, an action rank attaching unit 316, and a rule generation unit 318. The behavior cluster unit 314 executes processing for classifying the behavior level of the access subject 202 into clusters so as to reflect the intention. The reason for classification into clusters is, for example, the frequency at which specific actions are performed between each cluster when the intention that triggers the action of the access subject is classified as low, medium, or high with respect to the access target. Is different. For this reason, the same strength level is assigned to each of the actions from the most likely action to the least likely action in each cluster so as to be uniformly processed as a rough set described later.

  For example, in the cluster having the least interest / interest with respect to the access target, “see” is the most likely behavior mode and is ranked A by the behavior ranking unit 316 described later. The possible performance action with a high cot is “pick up” and is assigned rank A. In this embodiment, clustering is performed using the Ward method, but clustering will be described in more detail later.

  The behavior ranking unit 316 ranks the elements of the behavior vector of the clustered access subject 202 to, for example, the levels of A, B, and C using the numerical criteria set for the cluster. Rank A is assigned to the most expected behavior of the access subject 202 in a particular cluster, rank C is assigned to the least expected behavior, and rank B is moderately in between. Assigned to the expected behavior. The action ranking unit 316 can generate a rough set having a behavior attribute as a condition attribute from the normalized action vector. Note that the attribute of cluster classification is a decision attribute that reflects the intention.

  Further, the rule generation unit 318 uses the expectation of the behavior pattern of the rough set as a condition attribute, uses the cluster classification attribute as a decision attribute, reduces between the condition attributes, and reduces the rule from the reduction of the condition attribute that gives the decision attribute. Alternatively, a rule set is determined and registered in the database 324 in association with the intention state. The registered rule is inquired by the intention analysis unit 320 using SQL or the like, and is used to estimate the intention state of a specific access subject 202 with respect to a specific recognition target.

  Through the processing described above, the behavior pattern of the access subject 202 regarding a specific recognition target can be data mined with the intention of the access subject 202. Therefore, in the present embodiment, it is possible to perform marketing such as obtaining products that appeal to the access subject 202 more strongly from the information of the access subject 202 stored in the database 324.

  The database access unit 322 (hereinafter referred to as the DB access unit 322) is implemented including an SQL parser and the like, receives an SQL query from each unit, creates a target data table, and performs subsequent processing. The data can be used by departments.

  FIG. 4 is an embodiment of a data table stored in the database 324 in this embodiment. The data table 400 associates the recognition target 212 with the position information of the recognition target 212, and enables the behavior determination unit 310 to associate the behavior pattern of the access subject 202 with the recognition subject. The data table 410 associates the intensity level of the behavior style with the criterion for determining the intensity level in a specific cluster, and is used by the action ranking unit 316 to generate a rough set. . In the data table 410, type I to type III indicate classification attributes of clusters, type I is a cluster corresponding to a low degree of interest, and type II is a cluster corresponding to a medium degree of interest. Type III is a cluster with a high degree of interest. In addition, the degree of interest corresponds to the will to perform a specific behavioral pattern.

  In the present embodiment, intensity levels of level A to level C are set for the three behavior styles, and a total of nine determination criteria are registered, and thresholds for “see”, “touch”, and “take in hand” are respectively set. Value is set. Further, as the intention of the access subject 202 increases, the access subject 202 performs a plurality of actions on the recognition target 212. Reflecting this, in the data table 410, with regard to “see”, “touch”, and “take in hand”, the intensity level is set in a format depending on the classification attribute of each cluster in consideration of the closeness of access. Assigned to C.

  The data table 410 will be described in more detail. In the “view” behavior mode that is most easily accessible, in order to determine level A, the element value of the normalized behavior vector is set to be 0.6 or more, The threshold for determining level A decreases as the classification attribute of the cluster increases in will. In addition, the intention state is also reflected in the level C determination, and in the case of a cluster classification attribute with a low intention state, it is assumed that there will be more level C behavior patterns, so the threshold is determined as level C. The value is type I, 0.40 or less, and in type III, the group with high interest rarely takes level C behavior, so it is set to perform level C judgment with a threshold of 0.05 or less. doing.

  In the present embodiment, an intermediate element value between level A and level B is classified into level B. The determination threshold for the intensity level can be set statistically from an ergonomic experiment, and is a value set as appropriate depending on the type of access subject and the recognition target.

  FIG. 5 shows a flowchart of processing for registering the action of the access subject 202 in the database 324, which is executed by the intention analysis apparatus 230 of this embodiment. The process starts in step S500, and changes in the video image sent from the camera are monitored in step S501. In step S502, it is determined whether or not there is an image change in the video image from the camera. If there is no image change (no), the process branches to step S501 and the inspection of the video image from the camera is continued. On the other hand, if an image change is detected in step S502 (yes), the video image and the transmission start time are sent to the database 324 in step S503.

  Thereafter, in step S504, it is determined whether there is no further image change over the set time. If it is determined that the image change has ended (yes), the video image transmission is stopped in step S505, and the stop time is set in the database 324. Send to. On the other hand, if an image change is detected again within the set time (no), the process branches to step S503 and the video image transmission is continued. An appropriate format such as video image wave, MPEG-2, MPEG-4, or H.264 can be used.

  On the other hand, in parallel with the detection of the image change, the intention analysis device 230 monitors the response from the proximity wireless sensor 214 in step S506, and in step S507, determines whether or not the proximity wireless sensor has detected the approach of the access subject. to decide. If it is determined in step S507 that the access subject 202 is not approaching (no), the process branches to step S506, and a response from the proximity wireless sensor 214 is awaited.

  When it is detected in step S507 that the access subject 202 has approached from the proximity wireless sensor 214 (yes), the intention analysis device 230 sends the access subject ID, position information, and detection start time to the database 324 in step S508. In step S509, it is determined whether or not the response of the proximity wireless sensor has disappeared. If the response of the proximity wireless sensor does not disappear (no), the process branches to step S508 and the response from the proximity wireless sensor 214 is inspected. On the other hand, if it is determined in step S509 that the response of the close proximity wireless sensor 214 has disappeared (yes), the disappearance time is sent to the database in step S510, and the subsequent response of the close proximity wireless sensor 214 is monitored again in step S506.

  FIG. 6 shows an embodiment of a data table of each processing unit managed by the database 324. The processing result data 600 is a data table for registering the access subject ID, position information, start time, and end time for registering the detection result of the proximity wireless sensor 214. For example, a name for uniquely identifying the access subject 202 is registered. Has been. In addition, as additional information, an address, a telephone number, an age, and the like may be acquired as data from the proximity wireless sensor. The data table 600 further creates a data table 600 by acquiring the position where the proximity wireless sensor 214 is installed, the start time and the end time, respectively, and entering them in the data table 600.

  In addition, the data table 610 manages video images from the camera. For the recognition target IDs such as the T-shirt 1 and the T-shirt 2, the position information, the transmission start time and the transmission end time of the video image. An image ID that uniquely identifies a video image received within that period is registered in association with it. By referring to the data tables 600 and 610, the access subject, the recognition target, the position, the contact time, and the contact time to the recognition target are managed in association with each other.

  Furthermore, the intention analysis server 340 uses the image analysis program included in the intention analysis unit 320 to determine the contact state of the video image using the criteria in Table 1 described above. For example, specifically, when the approaching time length and the time length when the image has changed overlap for 10 seconds or more, it is determined that the action mode is “touch”, and the data table 620 Fill out. Further, when the approach is detected but the video image is not registered, it is determined that the action mode = “view” and is entered in the data table 620.

  Then, for example, if the state where the access subject approaches and the recognition target disappears from the video image continues for a set time, for example, 30 seconds, it is determined that the action mode is picked up and entered in the data table 620 To do. In addition, if the action style = “hold in hand”, the access subject may purchase the recognition target. In this case, if the recognition target is lost, an appropriate criterion is set and the time length is set in advance. can do. Through the processing described above, data necessary for analyzing the interest is stored in the database 324, and an action vector related to a specific recognition subject of the access subject is generated based on the above formula (1).

  FIG. 7 shows a flowchart of processing executed by the behavior cluster unit of the present embodiment. The processing in FIG. 7 starts from step S700, and in step S701, the access subject ID, location information, video image, and each time are stored in the database 324. In step S702, the behavior of the access subject is determined from the analysis of the video image. And registered in the data table 620.

  In step S703, it is determined whether or not the data collection period has ended. If the data collection period has not ended (no), the process returns to step S701 to continue data collection. The data collection period can be appropriately set as an appropriate length. If the access frequency is high, the data collection period may be set every day. If the access frequency is not so high, data may be collected in units of one week or one month. Can be accumulated and processed. Data in the data collection period can be accumulated to improve the accuracy of interest analysis, and the results can be summarized for each data collection period to determine the temporal transition of the interest distribution.

  When the data collection period ends in step S703 (yes), an action vector is created in step S704, and the action vector of the access subject is normalized for each recognition target. Thereafter, in step S705, the behavior of the access subject with respect to a specific recognition target is cluster-classified from the element value of the behavior vector. Thereafter, in step S706, it is determined whether or not the process has been completed for all recognition targets. If the process has not been completed (no), the process is branched to step S704 and the process is repeated.

  On the other hand, if it is determined in step S706 that the process has been completed for all recognition targets (yes), it is determined in step S707 whether the process has been completed for all access subjects, and the process has not been completed (no). The process is branched to step S704 and the process is repeated. If it is determined in step S707 that the processing has been completed for all access subjects (yes), cluster data creation is completed in step S708, control is passed to the action rank attaching unit 316 in step S709, and the processing is terminated.

  FIG. 8 is a diagram illustrating the cluster classification process described with reference to FIG. 7 using a data structure. As shown in FIG. 8, the behavior determination unit 310 issues a SQL sentence based on the image analysis result and the data received from the intention analysis device 230, and sets a set of behavior vectors associated with the access subject ID. , Generated for each recognition target. In the behavior matrix 800 of FIG. 8, the behavior vector is generated from the behavior pattern (seeing, touching, picking up by hand), and the behavior vector is added as recognition targets α, β, γ,. By going, an M × N action matrix of the access subject is formed (M is the number of recognition targets, and N is the number of action patterns). The behavior matrix 800 described above can be determined independently for each recognition target. If the recognition target is highly homogeneous, the values of the recognition targets α, β, γ,... It is also possible to perform statistical processing as an intention for the same recognition object.

The element values of the behavior matrix 800 shown in FIG. 8 are still expressed as times (s) obtained from the observed time information, and are inconvenient for comparison between different access subjects. The behavior normalization unit 312 normalizes the element values in units of behavior vectors using the above formula (1). In the present embodiment, the behavior matrix is a positive half-value matrix, and can be normalized using Equation (1). In this embodiment, the term “half positive value” means a real number of 0 or more, and is synonymous with a non-negative real number. In another embodiment, the norm of the action vector may be calculated and normalized so that | norm | ² becomes 1. After the behavior normalization process, a behavior matrix 810 with each element normalized is generated.

  In the present embodiment, the above-described normalized action vectors are classified into clusters in units of recognition targets. Cluster classification can be performed by adding a field to a specific recognition target record in the behavior matrix 800 and entering the result of the cluster classification as, for example, Type I, Type II, or Type III. The cluster classification of the present embodiment is preferably performed using the Ward method from the viewpoint of not applying a large overhead to the calculation process. In the Ward method, cluster classification is performed so that the increase in the sum of squares within a cluster is minimized. In this embodiment, for the purpose of converting a behavior vector into a condition attribute in a rough set, the cluster classification is defined as a condition attribute. It is generated in association with the number of. In addition to the Ward method, other clustering methods such as K-means and the shortest distance method are not particularly limited as long as they can be used as resources / libraries of the intention analysis server 340.

  The cluster classification 820 is performed by using the value of the element value of each behavior vector 830. For example, as shown in FIG. 8, attention is paid to the closest access behavior, that is, in this embodiment, behavior mode = hand. Then, the action vector = the action vector having the largest value to be picked up is classified into a type III cluster, the action vector closest to the average value is classified into a type II cluster, and the action vector having the minimum value is classified into a type I cluster. Then, other action vectors according to the Ward method, the sum of the sum of squares in the county [MIN {sum of squares in the county (I) + sum of squares in the county (II) + sum of squares in the county (III)] is minimized. By clustering in the same way, it is possible to classify all action vectors into clusters.

  This cluster classification process is performed for each identical recognition target, and it is an essential process to generate type I to type III clusters as many as the number of recognition targets. However, when the recognition target is the same quality, it is possible to perform the cluster classification process uniformly for the same behavior mode regardless of the difference of the recognition target. In this embodiment, the recognition target is assumed to be the same quality for the access subject ID. Process. However, in other embodiments, when the recognition target is not homogeneous or has low homogeneity, the rule generation is performed by classifying the action of the access subject for each identical recognition target.

  FIG. 9 is a diagram illustrating processing executed by the action ranking unit 316 according to the present embodiment using the data structure 900. In FIG. 9, for convenience of explanation, the ranking process for a type III cluster having a high closest access action ratio will be described. The same processing can be applied to the other clusters I and II only by different judgment threshold values for rank A, rank B, and rank C. A plurality of action vectors 910 are classified in the type III cluster. 9 shows a set element diagram, but in terms of data structure, a cluster structure is defined by describing a cluster attribute in an additional field of the behavior matrix shown in FIG. The process is performed by creating a view or table.

  Here, it is assumed that the action vector 910 has a value of (see, touch, take in the hand) = (0.22, 0.44, 0.44), for example. The action ranking unit 316 looks up the data table 410 shown in FIG. 4 and performs ranking in the type III cluster. For example, referring to the data table 410, the behavior mode = “see” = 0.22 is rank = B in the type III cluster, and the behavior mode = “touch” = 0.44 is rank A in the type III cluster. Similarly, the action mode = “take in hand” = 0.44 is also set as rank A.

  As a result, the rank set 920 of (B, A, A) is assigned to the action vector of (see, touch, pick in hand) = (0.22, 0.44, 0.44). The rank of the action vector described above is used as a condition attribute in the rough set by the rule generation unit 318 to create a rule based on the rough set. The rule generation unit 318 contracts the condition attribute, and registers, as the maximum likelihood rule, a condition expression or a set of condition expressions that gives the largest number of cluster classification attribute elements among the condition expressions given by the condition attribute reduction. . Note that the type I to type III of the cluster generally indicate the intention state as it is, and therefore the rule generation unit 318 can create a rule by reducing the elements of the rank set of the action vector of each cluster as a condition attribute. . It is also possible to use more complicated type attribute determination based on the specific recognition target and the nature of the access subject.

FIG. 10 shows an embodiment of the rules generated by the rule generation unit 318 described in FIG. 9 by the reduction process. In the embodiment shown in FIG. 10, seven rules are extracted and are indicated by τ _ij . The rules shown in FIG. 10 are managed in the database 324 as a rule table. For example, TOUCH = B and TAKE = B corresponding to the SQL statement “SELECT INTEREST WHERE TOUCH = B and TAKE = B” from the intention analysis unit 320. The value corresponding to the interest column (INTEREST) registered corresponding to is returned to the intention analysis unit 320. In the above SQL sentence, “INTEREST” means the value of the intention level registered in the interest column. In the described embodiment, INTEREST = HIGH is returned according to the rules of FIG. The result of the inquiry is displayed on the display device 234 from the intention analysis unit 320, and is displayed as a response to the SQL sentence sent by the intention analysis unit 320.

  The displayed response can be further analyzed with other conditions and used as higher order marketing information. Hereinafter, the present embodiment will be specifically described with specific examples.

(A) System Configuration The intention analysis system of this embodiment is configured from a personal computer and a intention analysis server. The personal computer uses LINUX (registered trademark) as an OS and Java (registered trademark) as a programming language to describe a program for executing the intention analysis method according to the present embodiment, thereby forming the intention analysis apparatus according to the exemplary embodiment. . The personal computer is connected to the network via Ethernet (registered trademark), uses JDBC to request SQL processing from the intention analysis server, obtains the response, and obtains the result of interest analysis. The intention analysis server was configured by installing PostgreSQL (registered trademark) as a database application on the server.
(B) Experiment FIG. 11 shows a simulated store and sensor configuration used in this experiment. In the simulated store 1110, a cash register 1120 and display shelves 1130 and 1140 were configured, and twelve types of T-shirts having different colors and neck shapes were displayed side by side for each display shelf. In order to detect RFID, a proximity wireless sensor 1160 indicated by a character ● is arranged on each side of the display shelf closer to the subject. In addition, a camera 1150 indicated by a character ■ was installed so that the state of the T-shirt on each display shelf could be observed. In this experiment, in order to cut off the influence of the external environment on the behavior, the camera 1150 was placed as one on two display shelves because each subject entered the simulated store.

We asked 13 men in their 20s to give their subjects, gave them an RFID to identify them individually, had them select a T-shirt on the display shelf, and brought what they wanted to the cash register. In addition, for the purpose of eliminating the influence of the crowded state of the display shelf and the influence of other people's behavior, we asked each subject to enter the simulated store and cut the influence of external information on the behavior. In the same store configuration, all types of T-shirts were changed and the same test subject set was used, and experiments were performed on 24 types of T-shirts twice in total. After the experiment, ask each subject to fill out a questionnaire with a five-level evaluation of the degree of interest for the 24 types of T-shirts used in the two experiments. Obtained willingness. In order to associate the acquired intention state with the cluster classification, it was classified into three classifications 1-2, 3, and 4-5, and the association was performed.
(C) Results FIG. 12 shows the results of behavior patterns obtained by this experiment. In FIG. 12, the result table is arranged from the left hand side to the right hand side in the descending order of intention level. The result table shown in FIG. 12 is obtained by using the above formula (1) for the total 312 data of actions taken by 13 subjects on 24 types of T-shirts in the two experiments described in (B). One of the results of classifying 281 items using hierarchical cluster analysis based on the Ward method, excluding data that was converted into normalized action vectors and extracted by using 10% (31 items) of data randomly for evaluation. It is the table which showed the part. As described above, the relationship between the three types of cluster classification and each stage of the decision-making process is determined based on the behavior mode = “take in hand” as a representative action and the behavior mode = “take in hand” value.

  As shown in FIG. 12, when looking from the evaluation stage (cluster = type III) to the problem recognition stage (cluster = type I), in the problem recognition stage, most of the behavioral patterns are “ At the evaluation stage, the subject must evaluate the recognition target using all of the behavior patterns that are the targets of the evaluation of “see”, “touch”, and “take in hand”. I understand. Further, in the intermediate information retrieval stage, the behavior of the subject shows diversification at an intermediate level between the problem recognition stage and the evaluation stage.

  FIG. 13 shows a result table 1300 that is statistically processed from the category classification data shown in FIG. 12, assuming a normal distribution for the appearance frequency of the behavior pattern in each cluster classification indicating the intention stage of the subject. As shown in the statistical table 1310, the behavior mode = “see” is dominant in the problem recognition stage cluster (type I), and the behavior mode = “see” in the information search stage cluster (type II). And the ratio of “touch” are almost the same, and in the cluster at the evaluation stage (type III), the behavior style = “see”, “touch”, “take in hand” are all the same, Observing only a specific behavioral pattern cannot analyze the behavior, and it is necessary to evaluate each cluster in an integrated manner.

  In order to integrate the cluster data shown in FIG. 12 as a rough set, the values for the statistics table 1310 are used to rank each of the behavior styles = “see”, “touch”, and “take in hand”. And the range table 1320 is generated. The range table 1320 corresponds to the data table 410 in FIG. Using the range table 1320, rules were generated by using the rough results from the experimental results shown in FIG. 12, using the ranking results as condition attributes, and reducing the condition attributes using each cluster classification as a decision attribute. The generated rules are registered in the database as the rule table 1400 in FIG.

  FIG. 14 is a table in which a part of the result of comparing the stage evaluation of the decision making and the questionnaire result to the subject is extracted based on the rule attached by the rough set. As shown in FIG. 14, it was shown that the estimation of the decision making stage based on the rule and the agreement between the subject's intention stage based on the questionnaire result are sufficiently good. That is, by using the intention analysis system of the present embodiment, the intention level of the access subject can be estimated from the behavior of the access subject acquired over a long period of time, and the estimation result can be applied to the behavior estimation and marketing of the access subject. .

  As described above, according to the present invention, by analyzing the behavior of the access subject to the recognition target and quantifying the intention level of the access subject, the interest / preference of the access subject is quantified and specified based on the quantification. It is possible to provide an intention analysis device, an intention analysis method and program, and an intention analysis system that enable an access subject's intention prediction with respect to the recognition target of the object.

DESCRIPTION OF SYMBOLS 100 ... Data definition, 110 ... Problem recognition axis, 120 ... Information retrieval axis, 130 ... Evaluation axis, 140 ... Action vector, 200 ... Action observation system, 202 ... Access subject, 204 ... Path, 208 ... Display shelf, 210 ... Store , 212 ... recognition target, 214 ... proximity wireless sensor, 216 ... RFID, 218 ... camera, 220 ... network, 230 ... intention analysis device, 232 ... hard disk device (HDD), 234 ... display device, 236 ... keyboard, 238 ... mouse , 300 ... intention analysis system, 302 ... network, 304 ... network interface, 306 ... access detection unit, 308 ... image change detection unit, 310 ... action determination unit, 312 ... action normalization part, 314 ... action cluster part, 316 ... Action ranking part, 318 ... rule generation part, 320 ... intention analysis part, 3 22 ... Database access unit, 324 ... Database, 330 ... Second network, 340 ... Intention analysis server,

Claims (12)

A intention analysis server that analyzes the intention of the access subject based on the behavior of the access subject,
When the access subject approaches the recognition target, the access subject identifies the closeness of the access subject's behavior mode performed on the recognition target while the access subject approaches the recognition target, and the behavior An action determination unit that generates an action vector that associates the duration of the action with the action pattern;
An action normalization unit that normalizes element values of the action vector and normalizes actions of the access subject with respect to the recognition target;
An action cluster unit that classifies the action vector normalized by the action normalization unit into a plurality of clusters based on a closest access action of the access subject in the action mode;
The element values of the action vector attributed to the cluster generated by the action cluster unit are ranked based on the numerical range assigned to the assigned cluster, and the ranking value is a condition attribute, An action ranking unit that generates a rough set having a classification attribute as a decision attribute;
And a rule generation unit that generates a maximum likelihood rule that reduces the condition attribute of the rough set and gives the decision attribute.   The intention according to claim 1, wherein the behavior determination unit levels the behavior of the access subject with respect to the recognition target in association with closeness using a duration of the behavior and image analysis. Analysis server.   The intention analysis server according to claim 1, wherein the behavior cluster unit executes cluster classification by hierarchical clustering using a Ward method.   The intention according to any one of claims 1 to 3, wherein the intention analysis server includes a data table used by the behavior determination unit, the behavior ranking unit, and a database that manages a rule generated by the rule generation unit. Analysis server.   The intention analysis server acquires information on the behavior of the access subject from a intention analysis device connected via a network, and the intention analysis device detects the approach of the access subject to the recognition target An access detection unit that detects that the access subject has acted on the recognition target and sends a video image to the intention analysis server, and issues an inquiry to the intention analysis server The intention analysis server according to claim 1, further comprising an intention analysis unit that executes an analysis of the intention of the access subject. An intention analysis method executed by an information processing device that analyzes the intention of the access subject from the behavior of the access subject, the intention analysis method comprising:
Identifying the closeness of the behavior of the access entity performed on the recognition object while the access entity is approaching the recognition object when the access entity approaches the recognition object;
Associating the duration of the behavior with the behavior, generating a behavior vector;
Normalizing element values of the behavior vector to normalize the behavior of the access subject with respect to the recognition target;
Classifying the normalized behavior vectors into a plurality of clusters based on the access subject's closest access behavior of the behavior modes;
Ranking the element values of the action vector attributed to the generated cluster based on the numerical range assigned to the attributed cluster;
Generating a rough set with the ranking value as a condition attribute and the classification attribute of the cluster as a decision attribute;
And a step of generating a maximum likelihood rule that reduces the condition attribute of the rough set and gives the decision attribute.   The step of identifying the closeness includes the step of leveling the behavior of the access subject with respect to the recognition target in association with the closeness using a duration of the behavior and image analysis. Item 7. The intention analysis method according to Item 6.   The intention analysis method according to claim 6 or 7, wherein the step of classifying into clusters includes a step of executing cluster classification by hierarchical clustering using a Ward method.   Identifying the closeness, ranking comprises determining the closeness and querying a database managing respective data tables that register numerical ranges for the ranking, and The intention analysis method according to claim 6, further comprising a step of registering a maximum likelihood rule in the database. A intention analysis system including a intention analysis server that analyzes the intention of the access subject from the behavior of the access subject, and a intention analysis device connected to the intention analysis server via a network,
The intention analysis server
When the access subject approaches the recognition target, the access subject identifies the closeness of the access subject's behavior mode performed on the recognition target while the access subject approaches the recognition target, and the behavior An action determination unit that generates an action vector that associates the duration of the action with the action pattern;
An action normalization unit that normalizes element values of the action vector and normalizes actions of the access subject with respect to the recognition target;
An action cluster unit that classifies the action vector normalized by the action normalization unit into a plurality of clusters based on a closest access action of the access subject in the action mode;
The element values of the action vector attributed to the cluster generated by the action cluster unit are ranked based on the numerical range assigned to the assigned cluster, and the ranking value is a condition attribute, An action ranking unit that generates a rough set having a classification attribute as a decision attribute;
A rule generation unit that generates a maximum likelihood rule that reduces the condition attribute of the rough set and gives the decision attribute;
The intention analysis device includes:
An access detection unit that detects an approach of the access subject to the recognition target and sends it to the intention analysis server via the network;
An image change detection unit that detects that the access subject has acted on the recognition target and sends a video image to the intention analysis server via the network;
A intention analysis unit that issues a query to the intention analysis server via the network and executes an analysis of the intention of the access subject.   The behavior determination unit levels the behavior mode of the access subject with respect to the recognition target in association with closeness using the duration of the behavior mode and image analysis, and the behavior cluster unit The intention analysis system according to claim 10, wherein the cluster classification is performed by hierarchical clustering using a Ward method.   An apparatus-executable program for an information processing apparatus to execute the intention analysis method according to any one of claims 6 to 9.