JP6515438B2 - Importance calculation apparatus, importance calculation apparatus method, and importance calculation apparatus system - Google Patents

Description

  The present invention relates to an importance calculation apparatus, an importance calculation apparatus method, and an importance calculation apparatus system, and in particular, an importance calculation apparatus that ranks the importance of an attribute related to an object using object information, the importance The present invention relates to a calculation device method and an importance degree calculation device system.

  Patent Document 1 describes an example of a store information providing system for searching for a menu with high evaluation based on accumulation of evaluations made by a user with respect to a menu provided by a restaurant.

  The store information search system described in Patent Document 1 stores store information and product information in the store information database in association with each other. The user posts the evaluation of the menu provided at the visited store through the terminal to the system. The system stores the posted data in the store information database as information associated with the aforementioned product information. The system determines the user level according to the number of posts of the user and the number of reference votes from other users. The system also provides the ability to search for menus. For a search based on a certain menu, a menu with a high evaluation is presented based on the number of evaluations corrected based on the user level.

  Patent Document 2 describes an example of a store information collection / providing device that presents store information to a reader, evaluates a location according to the interest trend of the reader, and presents the result to a store operator, a store opening candidate, etc. ing.

  In the shop information collection and provision device described in Patent Document 2, information (shop position, available number of persons, business hours, etc.) regarding the shop is registered in advance in the registered shop information database. The collection and provision device collects information on reaction behavior (such as browsing detailed information) performed on a search result when the user searches for a store via the terminal. The collection and provision device requests the user's evaluation of the store position based on the reaction behavior.

  By the way, in opening a new store, it is necessary to estimate (or analyze) important attributes (such as taste and cleanliness) required for the new store.

  However, when the inventions disclosed in Patent Document 1 and Patent Document 2 are used, although user evaluation and reaction behavior can be obtained for existing stores, any information can be obtained for new stores that have not yet opened. I can not. This is because it is premised that stores actually exist in order to obtain evaluation and reaction behavior. That is, even if information on known stores and information about them are collected, it is impossible to estimate important attributes of unknown stores that do not exist there. In other words, the analysis processing technique which estimates the important attribute of unknown data (data about new stores) based on known data (data about existing stores) is not disclosed.

  Although the store is shown as a specific example, the present invention is not limited to this, and the information of known objects (services, transaction objects such as movable property and real estate, human resources desired to be collected by recruiting, etc.) and the related objects In the case where information (consumer or purchaser's evaluation comment, adopter's evaluation comment, etc.) is related, it is possible to replace with estimating the important attribute in the information of the unknown object.

JP, 2013-73577, A Unexamined-Japanese-Patent No. 2003-345962

  As in the techniques described in Patent Document 1 and Patent Document 2 described above, there is a problem that it is not possible to estimate the important attribute of an unknown object that does not exist.

  The object of the present invention is to solve the above-mentioned problems, and to calculate the importance degree of an important object of an unknown object based on the relevancy between the information of the known object and the related information of the known object, A degree calculation device method and an importance degree calculation device system are provided.

  The importance degree calculation apparatus of the present invention performs predetermined processing on a first information group in which a predetermined attribute is set to a predetermined value among information groups consisting of a plurality of attributes and a second information group consisting of an input query. And a calculation unit for calculating the importance of the predetermined attribute in the query based on the similarity between the estimation score and the target score.

  An advantage of the present invention is that important attributes of an object can be estimated even if the object is not existing.

It is a block diagram which shows the characteristic structure of 1st embodiment of this invention. It is a figure which shows the operation example of the data processing part in 1st embodiment of this invention. It is a figure which shows the example of the attribute in 1st embodiment of this invention. It is a figure which shows the operation example of the data processing part in 1st embodiment of this invention. It is a flowchart which shows the process of importance calculation in 1st embodiment of this invention. It is a flowchart which shows the production | generation process of the temporary related vector in 1st embodiment of this invention. It is a figure which shows the example of the query screen in 1st embodiment of this invention. It is a figure which shows the example of the result screen in 1st embodiment of this invention. It is a block diagram which shows the characteristic structure of 2nd embodiment of this invention. It is a block diagram which shows the structure of the shop information storage part of 2nd embodiment of this invention, and an evaluation information storage part. It is a flowchart which shows the learning process in 2nd embodiment of this invention. It is a flowchart which shows the estimation process in 2nd embodiment of this invention. It is a figure which shows the example of the data structure of the evaluation attribute information storage part in 2nd embodiment of this invention. It is a figure which shows the example of the data structure of the evaluation attribute master memory | storage part in 2nd embodiment of this invention. It is a figure which shows the example of the data structure of the evaluation attribute dictionary memory | storage part in 2nd embodiment of this invention. It is a flowchart which shows the evaluation attribute dictionary production | generation process in 2nd embodiment of this invention. It is a figure which shows the example of the data structure of the shop attribute information storage part in 2nd embodiment of this invention. It is a figure which shows the example of the data structure (shop classification) of the shop attribute master memory | storage part in 2nd embodiment of this invention. It is a figure which shows the example of the data structure (location area) of the shop attribute master memory | storage part in 2nd embodiment of this invention. It is a figure which shows the example of the data structure of the shop attribute dictionary memory | storage part in 2nd embodiment of this invention. It is a flowchart which shows the shop attribute dictionary generation process in 2nd embodiment of this invention. It is a figure which shows the example of the evaluation attribute vector in 2nd embodiment of this invention. It is a flowchart which shows the evaluation attribute vector production | generation process in 2nd embodiment of this invention. It is a figure which shows the example of the shop attribute vector in 2nd embodiment of this invention. It is a flowchart which shows the shop attribute vector generation process in 2nd embodiment of this invention. It is a block diagram which shows the detailed structure of the attribute learning part of 2nd embodiment of this invention. It is a flowchart which shows the attribute learning process in 2nd embodiment of this invention. It is a figure which shows the example of the temporary attribute vector in 2nd embodiment of this invention. It is a flowchart which shows the temporary attribute vector generation process in 2nd embodiment of this invention. It is a figure which shows the example of the shop attribute information of the new shop in 2nd embodiment of this invention. It is a figure which shows the example of the shop attribute vector of the new shop in 2nd embodiment of this invention. It is a figure which shows the example of the data structure of the important attribute score memory | storage part in 2nd embodiment of this invention. It is a block diagram which shows the detailed structure of the important attribute score calculation part of 2nd embodiment of this invention. It is a flowchart which shows the important attribute score calculation process in 2nd embodiment of this invention. It is a flowchart which shows the important attribute display process in 2nd embodiment of this invention. It is a figure which shows the example of the display result of the important attribute display part in 2nd embodiment of this invention. It is a block diagram which shows the characteristic structure of 3rd embodiment of this invention. It is a figure which shows the example of the data structure of the important attribute score time-sequential memory | storage part in 3rd embodiment of this invention. It is a flowchart which shows the important attribute score time-series calculation process in 3rd embodiment of this invention. It is a flowchart which shows the important attribute time-sequential display in 3rd embodiment of this invention. It is a figure which shows the example of the display result of the important attribute time-sequential display part in 3rd embodiment of this invention. It is a flow chart which shows store attribute improvement of the existing store in a 4th embodiment of the present invention.

(First embodiment)
Next, a first embodiment of the present invention will be described.

  First, the configuration of the first embodiment of the present invention will be described. FIG. 1 is a block diagram showing the configuration of the importance degree calculation apparatus according to the first embodiment of the present invention.

  Referring to FIG. 1, the importance degree calculation apparatus according to the first embodiment of the present invention includes a data processing unit 204 and a calculation unit 205.

  First, definitions of target attribute information and target related attribute information in the present invention will be described. The targets are services, trading targets such as movable property and real estate, and human resources that you would like to collect by recruiting. The target attribute information is information representing the target attribute. The target attribute is, for example, a price, an area, a distance from a station, etc. in the case of real estate, and is, for example, an age, sex, educational background, etc. in the case of human resources. The target related attribute information includes evaluation comments by consumers and buyers, and evaluation comments by adopters. Each is complex data, and is exchanged in the form of structured or unstructured data. The target attribute information and the target related attribute information may be converted to a vector representation for a predetermined process. In this case, target attribute information is converted as a vector of length m, and target related attribute information is converted as a vector of length n. In addition, a score (scalar value) between target attribute information and target related attribute information is also given. For example, if it is a high rating or rating, it is a high score.

  The data processing unit 204 is a program executable on computer hardware that generates a query vector based on attribute information of a target input as a query, generates a provisional vector, and calculates a score based on two vectors. is there. In generating a query vector, target attribute information is converted into a vector as a query vector (length m). In the generation of the temporary vector, a temporary vector (length n) is generated in which the attribute to be estimated in the target related attribute information has a predetermined value. In the calculation of the score, a vector of length m and a vector of length n are processed by a predetermined process to calculate an estimated score.

  The calculation unit 205 is a program executable on computer hardware that calculates the importance based on the similarity between the calculated estimated score and a target score defined in advance by the calculation unit or set from the outside. It is.

  FIG. 2 shows a specific example of the processing of the data processing unit 204. The data processing unit 204 extracts information from the target attribute information using morphological analysis or table reference. If the target is, for example, an object, information such as the price and the number of years of production, and the word appearance frequency of the introductory sentence, and if the target is, for example, a person, an age group and gender are extracted. The data processing unit 204 performs vectorization by storing the extracted information at each defined position of a vector having a predetermined length. For example, the first element of the vector is age, the second element is gender, and the third is the number of occurrences of word A. As a result, the query is a vector of length m. Furthermore, a temporary vector of length n is generated in which the value of the attribute to be estimated is set to 1 and the values of the other attributes are set to 0 among the target related attribute information. The element values 0 and 1 are shown as an example, and may be real numbers. m and n may be equal or different.

  The data processing unit 204 calculates a numerical value such as 0.528 as an estimated score from the query vector and the temporary vector. The numerical value is calculated by the data processing unit based on, for example, a certain rule. A certain rule takes, for example, a vector of target attribute information and a vector of target related attribute information as inputs, and a tendency or correlation between input and output data when taking a score between the two pieces of information as output. It is a rule learned mechanically. Based on this rule, even when data not included in learning data is input as well as input data included in learning data, there is a characteristic that a score to be output can be obtained from the tendency or correlation.

  A specific example is a case where the data processing unit 204 is realized by using different neural networks for two input vectors. The data processing unit 204 performs feature extraction of a vector of target attribute information with a first neural network, and outputs a vector dimensionally compressed to a specific dimensionality N, and in parallel, generates a vector of target related attribute information. The second neural network performs feature extraction, outputs a vector dimensionally compressed to the same dimensionality N as the feature extracted query vector, and further calculates an estimated score from the inner product of both vectors (N dimensions) after feature extraction Then, the estimation error is calculated from the difference between the calculated estimated score and the score between the two pieces of information, and by adjusting both neural networks by the error back propagation method, the value of each element of the neural network as a learning result is calculated. Get the weight value of

  FIG. 3 shows specific examples of target attribute information and its vector representation. In this example, 10 attributes are defined. For example, it is assumed that the attribute information of the object is information representing “man in 20's working adult”. The data processing unit 204 determines whether or not the extracted information is applicable to each defined attribute, and replaces the element of the applicable attribute with the extracted information. As a result, the attribute information of the target “man in the twenties” is converted into a vector in the fourth column.

  FIG. 4 shows a specific example of processing in which the data processing unit 204 generates a provisional vector and calculates an estimated score. The temporary vector is a vector expression of length n, and the k (k is 1 to n) temporary related information is set such that the k-th attribute is 1 and the other attributes are 0. Also here, the value of the attribute may be a real number other than 0 or 1. The provisional related information generated in this manner and the information of the unknown object are input to the data processing unit to obtain the k-th provisional score in each k-th trial. In the example of FIG. 3, the first provisional score is 0.92, and the n-th provisional score is 0.11.

  Here, estimation of the important attribute for the query by the calculation unit 205 will be described. The estimation is realized by obtaining a temporary vector k that outputs a score close to the target score, for example 1.00. In the example of FIG. 4, when the first temporary vector is input, a relatively high score is obtained, and when the nth temporary vector is input, a relatively low score is obtained. From this, it can be said that the first attribute of the related attribute information corresponding to k = 1 in the related information is more important than the n-th attribute in the query. In this way, assuming that the score when using the k-th temporary vector is the temporary score (k), the k-th attribute of the related attribute information for the query is closer to the target score and the temporary score (k). It can be judged as important. For example, the smaller the difference between the two, the better. The target score may be given in advance or may be input each time.

  In this way, important attributes for the query are estimated. The estimated result is notified to the user who made the query through a screen or the like.

  The importance degree calculation apparatus 100 may be a computer that includes a CPU (Central Processing Unit) and a storage medium storing a program, and operates under control based on the program. The data processing unit 204 and the calculation unit 205 may be separate storage media or may be configured by one storage medium.

  FIG. 5 is a flowchart showing a flow of processing of importance calculation of the importance calculation device 100 according to the first embodiment of the present invention.

  In step S101, the query data input unit 202 receives an input of a query and a target score. It is also possible to omit the input in this step as the target score is determined in advance.

  Next, the loop is started by the number n of the length of the vector representation of the related information (step S102). Let us refer to the index as k in the loop.

  In step S103, the data processing unit 204 converts the query into a query vector of length m, and further generates a k-th tentative vector. This is a vector in which the k-th attribute is, for example, 1 and the other elements are, for example, a length n of 0.

  Next, the data processing unit 204 acquires a provisional score (k) from the query vector and the generated k-th provisional vector (step S104).

  In step S105, the calculation unit 205 compares the provisional score (k) with the target score to calculate an important attribute score (k) (). The important attribute score (k) is a score relatively evaluated for each k with respect to the relation between the provisional score (k) and the target score. For example, the important attribute score (k) may be the difference between the two (in this case, higher rank or rank closer to zero) or the division result (high rank or rank closer to 1 in this case) . As yet another example, the important attribute score (k) may be ranked as important, ordinary, non-important.

  The above steps S103 to S106 are repeated until k reaches n (step S106).

  The above-described result obtained for k may be notified to the user through any method such as being displayed on a display device. This is the end of the description of the flow of importance level calculation.

  FIG. 6 is a detail of generation of a provisional vector performed in step S103.

  In step S201, an area of a provisional vector of length n is generated.

  Next, all the attributes of the temporary vector area are initialized to 0 (step S202).

  In step S203, the k-th attribute is set to 1. This is the end of the description of the flow of generation of the provisional vector.

  A screen which is a user interface in the present embodiment will be described with reference to FIGS. 7 and 8.

  FIG. 7 shows a screen W101 for receiving an input of a query from the user. This interface comprises a screen component for inputting a target score and a screen component for inputting information of an unknown target. The input information is delivered to the degree-of-importance calculating apparatus 100 via the query input unit 202. The target score can be input in a slider form between 0 and 1. The information of the unknown object is input by operating the check box from the list displayed by inputting characters and words in the search box. For example, when "salt ramen" is input, the elements related to it are displayed in a list, and the user selects the required one from those displayed in the list. When the setting is completed, the OK button is pressed to start calculating the degree of importance.

  FIG. 8 is a screen W102 for displaying the obtained degree of importance. This screen is processed by the respective units in response to the user's input to the screen W101, and finally the result calculated by the calculation unit 205 is presented to the user. The upper part shows a target score set in advance. At the bottom, a tabular screen part including important attribute scores, attributes, and attribute IDs is displayed. The obtained results are displayed as rows on this screen part. The Ascending and Descending buttons allow you to control the sort order of this table, from top to bottom, or vice versa, from top to bottom.

  In this manner, the user can input the target score and the information of the unknown object as a query through the screen, and obtain the importance of the attribute (in the above example, the number of important attribute scores) through the screen. In the present embodiment, an example in which importance information is obtained through the screen has been described, but obtaining through the screen is only an example. For example, the result is output by a PDF file or a TXT file, for example. They may be attached to the email.

  Thus, the operation of the first embodiment of the present invention is completed.

  According to the first embodiment of the present invention, even if an object does not exist yet, the important attribute of the object can be estimated.

  The reason is that the calculation unit calculates a temporary score for both the input query and the k-th temporary related information vector generated sequentially based on the rules of the calculation unit, and the higher the target score, the k-th Since the important attribute score of the attribute of is determined, using this score, it is possible to judge the relative importance between the attributes.

Second Embodiment
Next, a second embodiment of the present invention will be described.

  The present embodiment is an embodiment to which the first embodiment is applied. Store information (existing and new) corresponds to target attribute information in the first embodiment, and evaluation information corresponds to target related attribute information.

  The configuration of the embodiment of the present invention will be described in detail with reference to FIGS. 9 and 10.

  Since the embodiment of the present invention is composed of two phases of a learning phase and an estimation phase, they will be separately described.

  The learning phase performs attribute learning from evaluation information and store information.

  The part related to evaluation information includes an evaluation attribute master storage unit 311 a holding evaluation attribute master information, an evaluation attribute information storage unit 311 b holding evaluation attribute information of existing stores, and (311 a and 311 b collectively (Also referred to as 311), an evaluation attribute dictionary 313 for generating and storing an evaluation attribute dictionary from the evaluation information storage unit 311, and information from the evaluation information storage unit 311 and the evaluation attribute dictionary 313 to generate an evaluation attribute vector And an evaluation attribute vector generation unit 314.

  The part related to the store information is a store attribute master storage unit 301a holding store attribute master information of the existing store, a store attribute information storage unit 301b storing store attribute information of the existing store, and (301a and 301b collectively Store information dictionary 301), store attribute dictionary storage unit 303 for creating and storing a store attribute dictionary from the store information storage unit 301, and information from the store information storage unit 301 and the store attribute dictionary storage unit 303 And a store attribute vector generation unit 304 that generates a store attribute vector.

  Further, the part related to attribute learning is a set of three vectors generated by the store attribute vector generation unit 304 and the evaluation attribute vector generation unit 314, and an evaluation score stored in the evaluation attribute information storage unit 311b. And a learning result storage unit 322 that holds learning results.

  In the estimation phase, a store attribute information storage unit (new store) 301c holding store attribute information of a new store, a target score input unit 334 receiving an input of a target score which is an evaluation score desired to be assigned to a new store, store attribute vector An important attribute score calculation unit 331 that calculates an importance score for each evaluation attribute of a new store using information from the generation unit 304, evaluation attribute vector generation unit 314, learning result storage unit 322, and target score input unit 334 as input The important attribute score storage unit 332 that holds the importance score for each important attribute calculated by the important attribute score calculation unit 331 and the importance score from the important attribute score storage unit 332 are input to display the important attributes of the new store The important attribute display unit 333 is included.

  The operation of the present embodiment will be described based on the drawings such as flowcharts.

  The operation of the present embodiment will be described in detail with reference to FIG.

First, the operation of the learning phase will be described. The evaluation attribute dictionary storage unit 313 selects and repeats the records held by the evaluation attribute information storage unit 311b one by one (S301, S303), and refers to the data held by the evaluation attribute master storage unit 311a, while referring to the evaluation attribute A dictionary generation process is performed (S302), and the generated dictionary is stored in the storage area of the evaluation attribute dictionary storage unit 313 (S302). In parallel, the shop attribute dictionary generation unit 303 repeats the records held by the shop attribute information storage unit 301b one by one (S304, S306), and refers to the data held by the shop attribute master storage unit 301a. An attribute dictionary generation process (S305) is performed, and the generated dictionary is stored in the storage area of the store attribute dictionary storage unit 303 (S305).
Next, the evaluation attribute vector generation unit 314 selects records of evaluation attribute information held by the evaluation attribute information storage unit 311 b one by one (S 307), and selects shop attribute information targeted by the selected evaluation attribute information. Then, the evaluation score corresponding to the selected evaluation attribute information (existing store) is selected, and the selected evaluation attribute information is converted into an evaluation attribute vector while referring to the evaluation attribute dictionary held by the evaluation attribute dictionary storage unit 313 (S308) ). In parallel, the shop attribute vector generation unit 304 converts the selected shop attribute information into a shop attribute vector while referring to the information held by the shop attribute dictionary storage unit 301b (S309). The attribute learning unit 321 machine-learns the relationship between both attribute vectors and the evaluation score, and sequentially stores the learning result in the learning result storage unit 322 (S310).

  The evaluation attribute dictionary generation / storage (S302) will be described in detail with reference to FIGS. 13, 14, 15, and 16. FIG. The data structure of the evaluation attribute information storage unit 311b is shown in FIG. 13. The data structure of the evaluation attribute master storage unit 311a is shown in FIG. 14. The data structure of the evaluation attribute dictionary storage unit 313 is shown in FIG. 15. The operation of the evaluation attribute dictionary storage unit 313 is shown in FIG. Each is shown in. The evaluation attribute dictionary storage unit 313 acquires data for each column of the acquired evaluation attribute information (S501, S502), determines whether the data is text or not (S503), and performs morphological analysis if it is text. Attribute is extracted (S504), and if it is not text, the attribute is extracted with reference to the evaluation attribute master held by the evaluation attribute master storage unit 311a (S505), and the extracted attribute is added to the evaluation attribute dictionary storage unit 313 (S506). In FIG. 15, in order to distinguish the text format attribute extracted from the evaluation attribute information from the attribute of the other format, the evaluation attribute dictionary is “column name: evaluation attribute” (However, in the case of the attribute extracted from text) , Hold the evaluation attribute in the form of "word" column name.

  Store attribute dictionary generation and storage (S305) will be described in detail with reference to FIGS. 17, 18, 19, 20 and 21. FIG. The data structure of the store attribute information storage unit 301b is shown in FIG. 17, the data structure of the store attribute master storage unit 301a is shown in FIGS. 18 and 19, the data structure of the store attribute dictionary storage unit 303 is shown in FIG. Are respectively shown in FIG. The store attribute dictionary storage unit 303 acquires data of each column of the acquired store attribute information (existing store) (S601, S602), and when the data is text, performs morphological analysis to extract an attribute (S604) If the text is not text, the attribute is extracted with reference to the store attribute master held by the store attribute master storage unit 301a (S605), and the extracted attribute is added to the storage area of the store attribute dictionary storage unit 303 (S606). In order to distinguish between text format attributes and other format attributes, the store attribute dictionary is "column name: store attribute" (However, in the case of the attribute extracted from the text, the column name is "word") Store attributes in the form of.

  The evaluation attribute vector generation (S308) will be described in detail with reference to FIGS. 13, 15, 22 and 23. FIG. The data structure of the evaluation attribute vector is shown in FIG. 22, and the operation of the evaluation attribute vector generation unit 105 is shown in FIG. The evaluation attribute vector generation unit 314 generates an evaluation attribute vector having the same dimension as the number of evaluation attributes of the evaluation attribute dictionary held by the evaluation attribute dictionary storage unit 313, and initializes all elements of the vector to 0 (S701) Data of each column of the acquired evaluation attribute information is acquired (S702, S703), it is determined whether the data is text or not (S704), and if it is text, morphological analysis is performed to extract an attribute (S705) If it is not text, the attribute is extracted with reference to the evaluation attribute master held by the evaluation attribute master storage unit 311a (S706), and the extracted attribute is referred to with reference to the evaluation attribute dictionary held by the evaluation attribute dictionary storage unit 313. The matching evaluation attribute is searched (S 807), and when the matching attribute is found, the element of the evaluation attribute vector corresponding to the attribute is set to 1. (S808).

  Store attribute vector generation (S309) will be described in detail with reference to FIG. 17, FIG. 20, FIG. 24 and FIG. The data structure of the store attribute vector is shown in FIG. 24, and the operation of the store attribute vector generation unit 304 is shown in FIG. The shop attribute vector generation unit 304 generates a shop attribute vector having the same dimension as the number of shop attributes in the shop attribute dictionary held by the shop attribute dictionary storage unit 303, and initializes all elements of the vector to 0 (S801) Data for each column of the acquired shop attribute information (existing shop) is acquired (S802, S803). In step S804, it is determined whether the data is text. If the data is text in step S805, morphological analysis is performed to extract an attribute. If it is not text, the attribute is extracted with reference to the store attribute master held by the store attribute master storage unit 301a (S806), and the store attribute dictionary stored with the store attribute dictionary 303 is matched with the extracted attribute A store attribute to be searched is searched (S807), and when a matching attribute is found, the element of the store attribute vector corresponding to the attribute is set to 1 (S808).

  The attribute correlation learning / storage (S310) will be described in detail with reference to FIGS. 13, 22, 24, 26, and 27. FIG.

  The configuration of the attribute learning unit 321 is shown in FIG. The attribute learning unit 321 performs feature extraction from the store attribute vector and performs dimension compression on the specific dimension N by the store attribute vector feature extraction unit 321a, and feature extraction from the evaluation attribute vector, and the number of dimensions of the store attribute vector after feature extraction An estimated evaluation score calculation unit 321c that calculates an estimated evaluation score from an evaluation attribute vector feature extraction unit 321b that performs dimension compression to the same number of dimensions as N, a shop attribute vector that has been subjected to feature extraction and the evaluation attribute vector, The difference (estimated error) between the estimated evaluation score calculated by the unit 321c and the actual evaluation score held by the evaluation attribute information storage unit 311b is calculated, and the difference is estimated to the store attribute vector feature extraction unit 321a and the evaluation attribute vector feature extraction unit 321b. The estimation error calculation unit 321 d is configured to back propagate an error.

  In addition, although arbitrary methods (principal component analysis, Latent Semantic Indexing etc.) which perform feature extraction are applicable to store attribute vector feature extraction part 321a and evaluation attribute vector feature extraction part 321b, it implemented using a neural network after that Give an example. In addition, although an arbitrary metric (L1 norm, L2 norm, etc.) for calculating a scalar value from two vectors is applicable to the estimated evaluation score calculation unit 321c, an example implemented using an inner product will be described hereinafter.

  The operation of the attribute learning unit 321 is shown in FIG. The attribute learning unit 321 receives the evaluation attribute vector generated by the evaluation attribute vector generation unit 314, performs feature extraction with a neural network, and outputs a vector dimensionally compressed to a specific number of dimensions N (S901). The store attribute vector generated by the store attribute vector generation unit 304 is input to perform feature extraction using a neural network, and a vector dimensionally compressed to the same dimension number N as the feature extracted evaluation attribute vector is output (S902). Furthermore, the estimated evaluation score is calculated from the inner product of the evaluation attribute vector (N-dimensional) and the shop attribute vector (N-dimensional) after feature extraction (S903), and the estimation error is calculated from the difference between the calculated estimated evaluation score and the actual evaluation score. Calculate (S904). The estimation error is fed back to the evaluation attribute vector feature extraction unit 321b, and the parameters of the evaluation attribute vector feature extraction unit 321b are adjusted by machine learning so as to reduce the estimation error (S905). The parameters of the store attribute vector feature extraction unit 321a are adjusted by machine learning so as to be fed back to the feature extraction unit so that the estimation error becomes smaller (S906). Finally, the learning result (the parameter group of the store attribute vector feature extraction unit 321a and the evaluation attribute vector feature extraction unit 321b) is stored in the learning result storage unit (existing store) 322 (S907).

  In addition, since the parameter adjustment method (reverse error propagation method) of the neural network based on an estimation error is known, detailed description is omitted. Also, a neural network that receives a store attribute vector and a neural network that receives an evaluation attribute vector each have a weight corresponding to the input so that the same or different vector is input to output a vector of the same dimension. You may do so.

  Next, the operation of the estimation phase will be described with reference to FIG. The attribute vector generation unit 304 selects one store attribute information (new store) from the store attribute information storage unit 301c that holds information on a new store, and stores the store attribute dictionary memory unit 303 and the store attribute master storage unit 301a. The selected shop attribute information (new shop) is converted into a shop attribute vector while referring to the information held by the (S401). The evaluation attribute vector generation unit 314 selects records (features) one by one from the evaluation attribute dictionary held by the evaluation attribute dictionary storage unit 313 generated in S302 (S402), and the evaluation attribute dictionary storage unit (existing store) 313 The selected record (feature) is converted into a temporary attribute vector while referring to the evaluation attribute dictionary to be held (S403). The important attribute score calculation unit 331 calculates an estimated evaluation score based on the shop attribute vector and the temporary attribute vector using the learning result held by the learning result storage unit 322 generated in S310 (S404), and the calculation result Are sequentially stored in the important attribute score storage unit 332 for each record (feature) (S404). When the target score is input (S406), the important attribute score is calculated based on the target score and the estimated evaluation score (S407), and the highly correlated evaluation attribute is displayed on the screen as the important attribute (S408).

  The details of the estimation phase will be described. First, temporary attribute vector generation (S403) by the evaluation attribute vector generation unit 314 will be described in detail with reference to FIG. 15, FIG. 28, and FIG. An example of the temporary attribute vector is shown in FIG. 28, and the operation of the evaluation attribute vector generation unit 314 is shown in FIG. The evaluation attribute vector generation unit 314 generates an evaluation attribute vector having the same dimension as the number of evaluation attributes of the evaluation attribute dictionary held by the evaluation attribute dictionary storage unit 313 (S1001), and initializes all elements of the vector to 0 ((1001) S002), 1 is set to the element corresponding to the evaluation attribute of the evaluation attribute dictionary selected in S402 (S1003). When explaining using FIG. 28, when the evaluation attribute of the evaluation attribute dictionary selected in S402 is the first attribute (EA1), the corresponding temporary attribute vector is (1, 0, 0,... 0) . When the second and third attributes (EA2, EA3) are selected, corresponding temporary attribute vectors are generated. Therefore, when there are 19 attributes, a total of 19 temporary attribute vectors are generated.

  Store attribute vector generation (S401) of the estimation phase will be described in detail with reference to FIG. 30 and FIG. The data structure of the store attribute information storage unit (new store) 301c is the same as that of the store attribute information storage unit (existing store) 301b. An example of shop attribute information held by the shop attribute information storage unit (new shop) 301c is shown in FIG. 30, and an example of a shop attribute vector generated from the information is shown in FIG. In addition, since the operation of the store attribute vector generation unit 304 is the same as the store attribute vector generation (S309), the detailed description of the operation is omitted.

The important attribute score calculation / storage (S407) will be described in detail with reference to FIG. 32, FIG. 33 and FIG. The data structure of the important attribute score storage unit 332 is shown in FIG. The configuration of the important attribute score calculation unit 331 is shown in FIG. The important attribute score calculation unit 331 performs feature extraction from the store attribute vector and performs dimension compression on the number N of dimensions by the store attribute vector feature extraction unit 331a, and performs feature extraction from the evaluation attribute vector and stores the store attribute vector after feature extraction. It comprises an evaluation attribute vector feature extraction unit 331b that performs dimension compression to the same number of dimensions as the number of dimensions N, and an estimated evaluation score calculation unit 331c that calculates an estimated evaluation score from the store attribute vector for which feature extraction has been performed and the evaluation attribute vector. .
The operation of the important attribute score calculation unit 331 is shown in FIG. The important attribute score calculation unit 331 reads the learning result held by the learning result storage unit 322 generated by the attribute learning unit 321 (S1101), and inputs the temporary attribute vector generated by the evaluation attribute vector generation unit 314 as an input Feature extraction is performed by the network, and a vector compressed to a specific number of dimensions N is output (S1102), and in parallel, feature extraction is performed by the neural network using the store attribute vector generated by the store attribute vector generation unit 304 as input The evaluation evaluation score is estimated from the inner product of the evaluation attribute vector (N-dimensional) and the shop attribute vector (N-dimensional) after the feature extraction, outputting a vector dimensionally compressed to the same dimension number N as the evaluation attribute vector extracted feature Is calculated in the important attribute score storage unit 332 (S1104). The estimated evaluation score column of the corresponding element in the evaluation attributes-option, to save the calculated estimated evaluation score (S1105).

  The important attribute display (S406, S407) will be described in detail with reference to FIG. 35 and FIG. The operation of the important attribute display unit 333 is shown in FIG. Further, FIG. 36 shows a display result of the important attribute display portion 333 in a case where a target score (hereinafter, an example in which “1.0” is input by the user ”is described) is acquired from the user. The important attribute display unit 333 acquires a target score that the user wants to give to a new store from the target score input unit 334 (S1201), and one estimated evaluation score for each evaluation attribute ID held by the important attribute score storage unit 332 (S1202, S1203), calculate the absolute value of the difference between the target score and the estimated evaluation score, and calculate the important attribute score of the evaluation attribute ID (in this example, the smaller the score value, the stronger the degree of important attributes) Calculate (S1204), and after calculating the important attribute score for all evaluation attribute IDs, sort the evaluation attribute ID in ascending order by the important attribute score (S1206), and display the pair of the evaluation attribute ID and the important attribute score on the screen (S1207).

  In the present embodiment, extraction of attributes by morphological analysis (S504, S604) can be applied to any known method (rule-based, learning-based, etc.), and thus detailed description will be omitted.

  Thus, the operation of the second embodiment of the present invention is completed.

  According to the second embodiment of the present invention, it is possible to estimate an important attribute in a store that does not exist, even for unknown stores.

  The reason is that machine learning of the correlation between the two from the shop attribute information of the existing shop and the evaluation information for the shop, and based on the learned correlation, an important attribute to unknown shop attribute information not included in the information This is because the (feature) is estimated.

Third Embodiment
Next, the third embodiment of the present invention will be described.

  In the third embodiment of the present invention, as shown in FIG. 37, in the configuration of the estimation phase in the second embodiment (FIG. 9), the important attribute score calculation unit 331 is replaced by the important attribute score time series calculation unit 401; The attribute score storage unit 332 is replaced with the important attribute score time-series storage unit 402, and the important attribute display unit 333 is replaced with the important attribute time-series display unit 403.

  Referring to FIG. 39, the operation of the important attribute score time-series calculation unit 401 learns the operation of the important attribute score calculation unit 331 shown in FIG. 34 according to the store information and the date and time of the evaluation information as the basis. And extended for calculation. S1301, S1302, S1303, S1304, and S1305 in FIG. 39 correspond to S1101, S1102, S1103, S1104, and S1105 in FIG. 34, respectively. However, as the learning result read in S1301, the result obtained by learning based on the store information and the evaluation information corresponding to the date and time when the important attribute score time series calculation unit operates is used. Further, the important attribute score time series calculation unit 401 stores the estimated evaluation score (S1305), and additionally stores the date and time when the estimated evaluation score is calculated in the estimated date and time sequence (S1306).

  The data structure of the important attribute score time-series storage unit 402 is shown in FIG. The operation of the important attribute score time series calculation unit 401 is shown in FIG. As shown in FIG. 38, in addition to the evaluation attribute ID and the estimated evaluation score, the estimated date and time column is added, and the date and time when the estimated evaluation score on the left is calculated is recorded in the estimated date and time column.

  The operation of the important attribute time-series display unit 403 will be described with reference to FIG. The important attribute time-series display unit 403 acquires a target score that the user wants to give to a new store from the target score input unit 334 (S1401), and the evaluation attribute ID held by the important attribute score time-series storage unit 402, estimated evaluation score And select tuples of estimated date and time one by one (S1402, S1403), calculate the absolute value of the difference between the target score and the estimated evaluation score, and select the important attribute score of the evaluation attribute ID (in this example, the smaller the score value is, The importance attribute degree is strong) is calculated (S1404), the estimated date and time of the estimated evaluation score is acquired (S1405), and after the calculation of the important attribute score for all the evaluation attribute ID and estimated date and time pairs is completed, the evaluation attribute A tuple of the ID, the important attribute score, and the estimated date and time is displayed on the screen (S1408).

  The display result of the important attribute time-series display unit 403 is shown in FIG. Among the degrees of importance evaluated at different times and times, the attributes that have changed significantly when judged according to a predetermined criterion are highlighted in a list. For example, in FIG. 41, the display is highlighted so that it can be seen that the importance of EA 2 is greatly increased.

  Thus, the operation of the third embodiment of the present invention is completed.

  According to the third embodiment of the present invention, it is possible to specify the change tendency according to the date and time of the important attribute.

  The reason is that by estimating the important attribute score of each evaluation attribute at each different date and time and collectively displaying them, it is possible to grasp the rise and fall in the day and time for each evaluation attribute.

Fourth Embodiment
Next, a fourth embodiment of the present invention will be described.
(Improved store attributes of existing stores)
According to the fourth aspect of the present invention, as shown in FIG. 42, in the operation of the estimation phase in the second embodiment (FIG. 12), a shop attribute is selected where one shop attribute information (new shop) is selected. This is an operation in which one item of information (improved existing store) is selected (S1502).

  In the configuration of the estimation phase (FIG. 12), the store attribute information storage unit (new store) regarding what changed some store attributes in the store attribute information of the existing store as store attribute information regarding a virtual new store In the operation of the estimation phase shown in FIG. 42, the store attribute information is selected (S1502), and the important attribute according to the store attribute information is displayed (S1509).

  Thus, the operation of the fourth embodiment of the present invention is completed.

  According to the fourth embodiment of the present invention, it is possible to estimate, for an existing store, the attribute that the store should improve.

  The reason is that important attributes of existing stores can be calculated in the same manner as new stores.

  The present invention can be applied to a degree-of-importance calculating device that ranks the degree of importance of a target attribute.

100 Importance calculation apparatus 204 Data processing unit 205 Calculation unit 301 Store information storage unit 301a Store attribute master storage unit 301b Store attribute information storage unit 301c Store attribute information storage unit (new store)
303 store attribute dictionary storage unit 304 store attribute vector generation unit 311 evaluation information storage unit 311a evaluation attribute master storage unit (existing store)
311b Evaluation attribute information storage unit (existing store)
313 evaluation attribute dictionary storage unit 314 evaluation attribute vector generation unit 321 attribute learning unit 321a store attribute vector feature extraction unit 312b evaluation attribute vector feature extraction unit 321c estimated evaluation score calculation unit 321d estimated error calculation unit 322 learning result storage unit 331 important attribute score Calculation unit 331a Store attribute vector feature extraction unit 331b Evaluation attribute vector feature extraction unit 331c Estimated evaluation score calculation unit 331d Important attribute score calculation unit 332 Important attribute score storage unit 333 Important attribute display unit 334 Target score input unit 401 Important attribute score time series Calculation unit 402 important attribute score time series storage unit 403 important attribute time series display unit

Claims (12)

The estimated score is calculated using a first information group in which a predetermined value is set to an attribute for which an estimated score is to be calculated among information groups consisting of a plurality of attributes and a second information group composed of an input query. A data processing unit to calculate
A calculation unit that calculates an important score indicating importance of an attribute to be calculated for the estimated score in the query by performing arithmetic processing using the estimated score and the target score;
Importance calculation device provided with. The first information group is a first vector, and the second information group is a second vector obtained by converting the query into a vector.
The importance calculation device according to claim 1. The predetermined processing is configured by different neural networks for respectively processing the first vector and the second vector,
The neural network calculates an estimated score based on a predetermined weight.
The importance calculation device according to claim 2. The weight of the neural network is optimized based on the difference between the estimated score calculated by inputting a data pair including the first data and the second data to the data processing unit and the teacher score corresponding to the data pair. The importance degree calculation apparatus according to claim 3, further comprising a learning unit that A time series calculation unit that holds the importance score calculated by the calculation unit for the predetermined attribute for each date and time when the importance score is calculated;
A time series display unit for displaying the importance score of the predetermined attribute for each of the date and time;
The importance degree calculation device according to any one of claims 1 to 4, further comprising: The important score is calculated by calculating a difference between the estimated score and the target score.
The importance degree calculation apparatus of any one of Claim 1 thru | or 5. The first data is data including attribute information on a store,
The second data is data including evaluation information on the store,
The teacher score is an evaluation score for the store,
The importance calculation device according to claim 4. The data processing unit uses a first information group in which a predetermined value is set to an attribute for which an estimated score is to be calculated among information groups consisting of a plurality of attributes and a second information group consisting of an input query. Calculate the estimated score,
The calculation unit performs an arithmetic process using the estimated score and the target score to calculate an important score indicating importance of an attribute to be calculated for the estimated score in the query.
How to calculate importance. The first information group is a first vector, and the second information group is a second vector obtained by converting the query into a vector.
The importance calculation method according to claim 8. The predetermined processing is configured by different neural networks for respectively processing the first vector and the second vector,
Calculating an estimated score based on predetermined weights of the neural network;
The importance calculation method according to claim 9. On the computer
The estimated score is calculated using a first information group in which a predetermined value is set to an attribute for which an estimated score is to be calculated among information groups consisting of a plurality of attributes and a second information group composed of an input query. Calculate
By performing arithmetic processing using the estimated score and the target score, an important score indicating the importance of an attribute for which the estimated score is to be calculated in the query is calculated.
Importance calculation program to execute processing. By performing predetermined arithmetic processing on attribute information of a new store and evaluation attribute information in a new store using learning results generated based on attribute information of the existing store and evaluation attribute information of the existing store , A data processing unit that evaluates each attribute of evaluation attribute information in a new store,
A calculation unit configured to calculate a value indicating the degree of importance of each attribute by performing arithmetic processing using the value indicating the evaluation and the value indicating the target;
An importance calculation device which causes a display unit to display an attribute important for a new store based on the importance.

Images