JP2018128942A - Analyzing apparatus, analyzing method, and program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide an analyzing apparatus, an analyzing method, and a program which can improve accuracy of the learning.SOLUTION: An analyzing apparatus has a learning processing unit for, when a first data is input, in order that a second data output from a first sorter included in a classification system formed of multi-layer sorter each of which can be formed independently becomes closer to a first teacher data, making the classification system learn and for, then, when the first data is input, in order that a third data output from a second sorter in a deeper hierarchy as compared with the first sorter becomes closer to a second teacher data, making the classification system learn on the basis of a learning result based on the first teacher data, and a classification processing unit for classifying unlearned data into a predetermined category on the basis of the classification system which are made to learn by the learning processing unit.SELECTED DRAWING: Figure 2

Description

  The present invention relates to an analysis apparatus, an analysis method, and a program.

  2. Description of the Related Art Conventionally, a technology is known that automatically learns data relevance from past search behaviors and uses the learning to assist future search behaviors (see, for example, Patent Document 1). In addition, a technique is known in which a feature amount is extracted and machine learning using the feature amount is performed at once using a neural network (see, for example, Patent Document 2).

JP 2006-285882 A JP 2010-519642 A

  However, in the conventional technique, the accuracy of learning may not be sufficient.

  SUMMARY An advantage of some aspects of the invention is that it provides an analysis apparatus, an analysis method, and a program that can improve the accuracy of learning.

  According to one aspect of the present invention, when the first data is input, the second data output from the first classifier included in the classification system in which the classifiers that can be independently configured are configured in multiple layers is the first teacher. After learning the classification system to approach the data, when the first data is input, the third data output from the second classifier having a deeper hierarchy than the first classifier is the second data. A learning processing unit that learns the classification system based on a learning result using the first teacher data so as to approach the teacher data, and an unlearned data based on the classification system learned by the learning processing unit And a classification processing unit that classifies the data into a predetermined category.

  According to one embodiment of the present invention, it is possible to provide an analysis device, an analysis method, and a program that can improve the accuracy of learning.

It is a figure which shows an example of the analysis system 1 containing the analysis apparatus 100 in embodiment. It is a figure which shows an example of a structure of the analyzer 100 in embodiment. It is a figure showing an example of action history information 132. It is a figure which shows an example of a questionnaire. It is a figure which shows an example of the questionnaire information. It is a figure which shows an example of the analysis information 136. FIG. It is a figure which shows an example of the DNN structure information 138. It is a figure which shows typically the deep neural network DNN produced | generated based on DNN structure information 138. FIG. 6 is a flowchart illustrating an example of processing executed by a learning processing unit 114. It is a figure which shows an example of the parameter information 140 for every layer. It is a figure which shows typically the content of the loop process of the flowchart shown in FIG. 5 is a flowchart illustrating an example of processing executed by a classification processing unit 116. It is a figure which shows each learning accuracy of the learning method of the deep neural network DNN in this embodiment, and the method illustrated as a comparative example. It is a figure which shows an example of the hardware constitutions of the analyzer 100 of embodiment.

  Hereinafter, an analysis apparatus, an analysis method, and a program to which the present invention is applied will be described with reference to the drawings.

[Overview]
The analysis device is realized by one or more processors. When the analysis apparatus inputs certain input data v (an example of the first data), output data h output from the first classifier included in the classification system in which classifiers that can be configured independently are configured in multiple layers. When the input data v is input after learning the classification system so that (an example of the second data) approaches the first teacher data, it is output from the second classifier that is deeper than the first classifier. The classification system is trained based on the learning result using the first teacher data so that the output data h (an example of the third data) to be approximated to the second teacher data.

  For example, the “classifier” is a neural network, and the “classification system” is a deep neural network DNN composed of a plurality of neural networks. In this case, the “first classifier” is a neural network of an arbitrary hierarchy among a plurality of layers included in the hidden layer of the deep neural network DNN, and the “second classifier” is a plurality of layers included in the hidden layer. The neural network of a hierarchy deeper than the hierarchy of the neural network at least as the first classifier. “Deep hierarchy” means a hierarchy close to an output layer described later. Each of the “first classifier” and the “second classifier” may mean a neural network of a certain hierarchy, or may mean a neural network of a plurality of hierarchies.

  Then, the analysis device classifies the unlearned data into a predetermined category based on the learned classification system. Thereby, the accuracy of learning can be improved. Hereinafter, as an example, a description will be given assuming that the classification system is a deep neural network DNN composed of a plurality of neural networks.

[overall structure]
FIG. 1 is a diagram illustrating an example of an analysis system 1 including an analysis apparatus 100 according to the embodiment. The analysis system 1 in the embodiment includes, for example, one or more terminal devices 10, a service providing device 20, a log acquisition device 30, an advertisement distribution server device 40, and an analysis device 100. These devices are connected via a network NW. Note that some or all of the plurality of devices included in the analysis system 1 may be integrated into one analysis device 100.

  Each apparatus shown in FIG. 1 transmits and receives various information via the network NW. The network NW includes, for example, a radio base station, a Wi-Fi access point, a communication line, a provider, the Internet, and the like. Note that it is not necessary for all combinations of the devices shown in FIG. 1 to be able to communicate with each other, and the network NW may partially include a local network.

  The terminal device 10 is a device used by a user. The terminal device 10 is a computer device such as a mobile phone such as a smartphone, a tablet terminal, or a personal computer. For example, when receiving a predetermined operation from the user, the terminal device 10 may communicate with the service providing device 20 via a preinstalled application and acquire content to be displayed or reproduced on the application. The content is, for example, moving image data, image data, audio data, text data, or the like. For example, the application may be an application that can enjoy Internet shopping and a search service, or an application that can enjoy SNS (Social Networking Service), mail service, information providing service (for example, news and weather forecast), and the like. There may be.

  Further, when receiving a predetermined operation from the user, the terminal device 10 may access a website provided by the service providing apparatus 20 via a predetermined web browser. For example, on the website provided by the service providing apparatus 20, services similar to the services provided by the various applications described above are provided.

  The service providing device 20 may be a web server device that provides a website such as a shopping site or a search site on the Internet, and communicates with the terminal device 10 on which the application is started to exchange various information. The application server apparatus to perform may be sufficient.

  The log acquisition device 30 collects, for example, cookies and caches managed for each web browser, log files of HTTP (Hypertext Transfer Protocol) requests, and the like from the terminal device 10 or the service providing device 20, thereby causing the terminal device 10 to Acquire user action history. Behavior history includes, for example, purchase history on internet shopping (such as the type, price, and number of products or services purchased), search history on search sites (such as search engines used and search queries entered), A viewing history (moving image type, viewing time, etc.) on a moving image site, an HTTP referer (URL history) for identifying what kind of web page is accessed.

  In addition, the log acquisition device 30 uses the purchase history, the search history, the viewing history, and the like for each application activated in the terminal device 10 as the user's behavior history in the same manner as the user's behavior history on the Internet via the web browser. You may get it.

  In addition, for example, an application that is activated on the terminal device 10 introduces a nearby real store or uses the location information of the terminal device 10 (for example, GPS (Global Positioning System positioning coordinates)) When providing various services such as providing information, the log acquisition device 30 may acquire position information stored in the terminal device 10 as a user's action history.

  The user action history acquired by the log acquisition device 30 may be for each terminal device 10, may be for each account in OS (Operating System) unit in one terminal device 10, or may be a web browser. Alternatively, it may be per account in application units.

  Note that the log acquisition device 30 may acquire the user's behavior history from another service provision device (not shown) instead of or in addition to acquiring the user's behavior history from the service provision device 20. The log acquisition device 30 transmits information indicating the acquired user action history to the analysis device 100.

  For example, the advertisement distribution server device 40 distributes advertisements as the service providing device 20 provides a service to the terminal device 10 based on the analysis result of the analysis device 100. For example, when the terminal device 10 receives a web page including an advertising space or a style sheet for an application corresponding thereto from the service providing device 20, the terminal device 10 is based on an ad tag or SDK (Software Development Kit) provided in advance in the advertising space. Then, an advertisement distribution request is transmitted to the advertisement distribution server device 40. In response to this, the advertisement distribution server device 40 uses, as a response to the advertisement distribution request, an advertisement image (for example, an image in which a URL is embedded) that can access the website of the advertisement requester who has contracted in advance. It transmits to the terminal device 10. As a result, the advertisement image is displayed on the screen of the terminal device 10.

  Using the deep neural network DNN, the analysis device 100 classifies the user corresponding to the behavior history into a predetermined category based on the behavior history of the user acquired by the log acquisition device 30. As described above, the deep neural network DNN is a neural network including at least two layers. Then, the analysis apparatus 100 provides the category distribution result for each user to the advertisement distribution server apparatus 40, for example. Thereby, the advertisement delivery server device 40 can change the delivery content of the advertisement according to the classified category.

[Configuration of analyzer]
Hereinafter, the configuration of the analysis apparatus 100 will be described with reference to the drawings. FIG. 2 is a diagram illustrating an example of a configuration of the analysis apparatus 100 according to the embodiment. As illustrated, the analysis apparatus 100 includes, for example, a reception unit 102, a display unit 104, a communication unit 106, a control unit 110, and a storage unit 130.

  The accepting unit 102 is an input interface such as a touch panel, a keyboard, and a mouse that accepts an operation input from a user. The display unit 104 is a display device such as a liquid crystal display device.

  The communication unit 106 includes, for example, a communication interface such as a NIC (Network Interface Card) or a DMA (Direct Memory Access) controller. The communication unit 106 communicates with the log acquisition device 30 or the advertisement distribution server device 40 via the network NW. For example, the communication unit 106 receives a user's action history from the log acquisition device 30 and stores this in the storage unit 130 as action history information 132.

  FIG. 3 is a diagram illustrating an example of the action history information 132. As shown in the example in the figure, the action history information 132 is associated with the address to which address the position of the terminal device 10 corresponds, or input to the search engine name used and the search engine. The search query is associated with each other, or the title name of the browsed website and the URL of the website are associated with each other. Moreover, the action history information 132 may be information in which, for example, the purchased product name or service name and the URL of the website from which the product is purchased are associated with the time. Of course, a user does not take all possible actions for a certain event, but usually takes one, a few, or even a few tens to hundreds of actions. is there. For example, it is extremely rare for one user to purchase all of the products sold at a shopping site, and usually only a few products are purchased. Therefore, the user's behavior history indicated by the behavior history information 132 tends to be sparse data with a small amount of information in the sense that it does not comprehensively include all possible behavior histories.

  The control unit 110 includes, for example, an acquisition unit 112, a learning processing unit 114, and a classification processing unit 116. These components are realized, for example, when a processor such as a CPU (Central Processing Unit) or a GPU (Graphics Processing Unit) executes a program stored in the storage unit 130. In addition, some or all of the components of the control unit 110 may be realized by hardware such as a large scale integration (LSI), an application specific integrated circuit (ASIC), or a field-programmable gate array (FPGA). It may be realized by cooperation of software and hardware.

  The storage unit 130 is realized by, for example, an HDD (Hard Disc Drive), a flash memory, an EEPROM (Electrically Erasable Programmable Read Only Memory), a ROM (Read Only Memory), a RAM (Random Access Memory), or the like. The storage unit 130 stores questionnaire information 134, analysis information 136, DNN configuration information 138, layer-by-layer parameter information 140, and the like in addition to various programs such as firmware and application programs and the action history information 132 described above. Such information will be described later.

  For example, the acquisition unit 112 communicates with an external server (not shown) by using the communication unit 106 to acquire information indicating a result of a questionnaire (survey) stored in the external server, and obtains the questionnaire information 134. Is stored in the storage unit 130.

  When the service providing apparatus 20 provides a questionnaire monitor site that can answer a questionnaire or a questionnaire monitor service via an application, the acquisition unit 112 uses the communication unit 106 to obtain the questionnaire information 134 from the service providing apparatus 20. You may get it. The questionnaire information 134 may be stored in the storage unit 130 in advance.

  For example, when an administrator who manages the analysis apparatus 100 inputs a questionnaire result to the reception unit 102, the acquisition unit 112 may use the information input to the reception unit 102 as the questionnaire information 134. .

  FIG. 4 is a diagram illustrating an example of a questionnaire. As in the example shown in the figure, the questionnaire is prepared with predetermined options for the standardized survey items (Q1 to 4 in the figure). In the illustrated example, five options are prepared, but the present invention is not limited to this, and may be two, three, four, or six or more.

  FIG. 5 is a diagram illustrating an example of the questionnaire information 134. Like the example of illustration, the questionnaire information 134 is the information by which the score according to the choice selected by the user was matched with the investigation item, for example. The user of the questionnaire information 134 and the user of the action history information 132 are at least partially the same. That is, it is assumed that the action history of the user who is the answerer of the questionnaire is acquired as the action history information 132.

  For example, in the above-described questionnaire, the higher the answer option for the survey item, the higher the score is set. For example, compared to negative answer options such as “I do not think so” or “I do not think so”, positive answer options such as “I think so” or “I think so” are set higher. Is done. Note that this relationship is an example, and a negative answer option may be set higher for a survey item.

  In addition, the acquisition unit 112 acquires information indicating the analysis result of the questionnaire described above, and stores this in the storage unit 130 as analysis information 136. For example, when the device from which the questionnaire information 134 is acquired performs analysis, the acquisition unit 112 may acquire the analysis information 136 from the device that has performed the analysis. For example, when an analyst who analyzes a questionnaire result inputs an analysis result to the reception unit 102, the acquisition unit 112 may use the information input to the reception unit 102 as analysis information 136.

  For example, the analysis of the questionnaire results includes (1) how each survey item included in the questionnaire is designed, (2) what analysis method is used for quantitative evaluation, and (3) the results of the analysis This is done in consideration of matters such as how to classify users who responded to. For example, for (1), whether the answer method for the survey item is an option format or a description format, or if the option format is an option format, is the option expressed in text or numerically? It is necessary to consider such things. For (2), by using a score according to the user's answer, for example, by performing factor analysis, what is the factor contributing to the characteristics of the user who answered the questionnaire, It is necessary to consider how much the factor contributes. In addition, when principal component analysis is performed instead of factor analysis, it is necessary to consider which variable is the principal component from a plurality of variables that may represent the characteristics of the user. For (3), as a final analysis result, which user and which user are considered to have the same tendency from the characteristics of each of the plurality of users who answered the questionnaire, or the category of the user's classification destination is It is necessary to consider whether there is one or more for each user. In this way, parameters that serve as statistical classification indicators for analysts who analyze questionnaire results to analyze appropriately from the purpose of conducting the questionnaire and classify users who answered the questionnaire to a certain degree of granularity Need to be derived. Such parameters are treated as hyperparameters that an analyst or the like needs to set appropriately.

  For example, when the category is “high annual income” and “low annual income” as the hyperparameter of (3), the hyperparameter (analysis index) of (2) has the highest annual income among the users who answered the questionnaire Factors common to users or factors common to users with low annual income may be used as parameters.

  FIG. 6 is a diagram illustrating an example of the analysis information 136. In the example shown in the figure, the result when the factor analysis is performed on the questionnaire result in consideration of the items (1) to (3) described above is shown. For example, the analysis information 136 is information in which a factor score is associated with each of the users who answered the questionnaire with respect to the factor considered in the factor analysis (such as FAC1 in the figure). The factor score is, for example, an index value that represents the degree of correlation between the factor and the user characteristics. In the illustrated example, the factor score of the user A is 1.113, the factor score of the user B is 0.380, and the factor score of the user C is 0.397 with respect to the factor FAC1. Represents. These factor scores may be calculated based on, for example, a score corresponding to the answer option selected by the user among the options prepared for the survey items of the questionnaire.

  Further, as in the example shown in the figure, the analysis information 136 may indicate a classification result indicating what category the user has classified for each user who answered the questionnaire. The category to be classified by the user may be derived, for example, by comprehensively considering a plurality of classification indexes such as the above-described factor score. For example, if all factor scores from factor FAC1 to FAC4 are positive values, the user category is type 1, and if any one of the factor scores is negative, the user category is In the case of a user who is a type 2 and any two factor scores are negative, the category of the user is a type 3, and in the case of a user where any three or more factor scores are negative, the category of the user is Classification may be made such as type 4.

  The learning processing unit 114 generates (constructs) a deep neural network DNN to be learned based on the DNN configuration information 138, and the behavior history information 132, the questionnaire information 134, and the analysis described above for the deep neural network DNN. Pre-training is performed using information 136.

  FIG. 7 is a diagram illustrating an example of the DNN configuration information 138. As in the illustrated example, the DNN configuration information 138 is information in which the number of units representing the number of neural networks is associated with each layer. In general, the deep neural network DNN includes three layers: an input layer, a hidden layer (intermediate layer), and an output layer. Data to be learned by the deep neural network DNN is input to the input layer. A result learned by the deep neural network DNN is output from the output layer. The hidden layer performs processing that is the core of learning. For example, the hidden layer is expressed by a function called an activation function (transfer function) and returns an output corresponding to the input. For example, the activation function is a normalized linear function (ReLU function), a sigmoid function, a step function, or the like, but is not limited thereto, and an arbitrary function may be used. The neural network in the input layer is an example of “input device”, and the neural network in the output layer is an example of “output device”.

  The DNN configuration information 138 defines the number of units in each of these layers. In the illustrated example, the number of units in the input layer is one, and the number of units in the output layer is s. Further, the hidden layer is at least two layers, and in the illustrated example, the hidden layer represents n layers (n ≧ 2). Moreover, it represents that the number of units of each layer of a hidden layer is r. Note that the number of units in each layer of the hidden layer is not necessarily r, and may be different in each layer. The number of hidden layers and the number of units in each layer described above are hyperparameters and may be arbitrarily changed. For example, when it is decided to classify into five categories by analyzing the questionnaire results, the number of units in the output layer may be set to five. Moreover, in order to improve the learning accuracy, the number of hidden layer units may be increased.

  The learning processing unit 114 refers to such DNN configuration information 138 to generate a deep neural network DNN having a predetermined configuration at the time of prior learning.

  FIG. 8 is a diagram schematically showing the deep neural network DNN generated based on the DNN configuration information 138. As shown in FIG. As in the example shown in the figure, the deep neural network DNN has a nerve transmission network from one or more units (neural network) of the input layer to each of a plurality of units of the shallowest layer among the n layers of the hidden layer. Edges that are not connected are connected. The “shallowest layer” is a layer closest to the input layer among a plurality of layers included in the hidden layer. In the present embodiment, the fact that it is close to the input layer side is expressed as “shallow”, and the case that it is close to the output layer side is expressed as “deep”. In the example described above, the shallowest layer is the first layer. From each unit included in the first layer, an edge is connected to each of a plurality of units of the shallowest layer (hereinafter referred to as the second layer) after the first layer among n layers of the hidden layer. Is done. From each unit included in the second layer, an edge is connected to each of a plurality of units of the shallowest layer (hereinafter referred to as the third layer) after the second layer among the n layers of the hidden layer. Is done. Similarly, for the units in the third and subsequent layers, the edge is connected to the units in the deeper layer. An edge is connected to each unit of the output layer from each deepest nth layer unit. In this way, the deep neural network DNN is expressed as a state probability model in which units of each layer are connected by edges, such as a constrained Boltzmann machine (RBM). Note that the deep neural network DNN may be expressed as a state probability model in which units belonging to the same layer are connected by edges, like a Boltzmann machine without restrictions.

  As pre-learning, the learning processing unit 114 performs initial setting of the deep neural network DNN by learning each layer of the hidden layer step by step in the generated deep neural network DNN.

[Pre-learning]
Hereinafter, the prior learning process will be described with reference to a flowchart. FIG. 9 is a flowchart illustrating an example of processing executed by the learning processing unit 114.

  First, the learning processing unit 114 determines the i-th layer among the n layers of the hidden layers as a processing target layer (S100). i is a temporary parameter temporarily calculated during processing, and a natural number in the range of 1 to n is given.

  For example, the learning processing unit 114 determines the first layer (i = 1) as a processing target layer. Note that the learning processing unit 114 may determine the second layer or the third layer other than the first layer as a processing target layer. In the following description, first, the first layer is determined as a processing target layer.

  Next, the learning processing unit 114, from the generated deep neural network DNN, the input layer and the output layer, all the layers determined as processing target layers in the previous processing, and the processing target layer in the current processing The determined i-th layer is extracted (S102). In the case of the first process, for example, the learning processing unit 114 extracts the input layer, the output layer, and the first layer from the deep neural network DNN.

  Next, the learning processing unit 114 inputs the action history information 132 indicating the action history of the user who answered the questionnaire in the input layer in the extracted neural network (S104). At this time, the learning processing unit 114 may convert the user's action history into information that can be processed by the processor, thereby obtaining input data v for the input layer.

  For example, the learning processing unit 114 refers to the action history information 132 to derive the number of views of the web page, the number of times of input of a predetermined search query, and the like as input data v. In addition, the learning processing unit 114 vectorizes a character string (which may include alphabets, numbers, symbols, and the like) related to each action history such as the product code of the purchased product and the URL of the browsing site. The input data v for the input layer may be used. The learning processing unit 114 derives output data h to be output from the output layer based on the extracted activation function of each unit of the hidden layer. The output data h can be expressed by, for example, the following formula (1).

h = σ (W ^T v + b) (1)

  σ represents an activation function of each unit of each layer, W represents a weight given to output data when data is output from a unit of a certain layer to a unit of a deeper layer, b represents a unique bias component of each layer.

  Next, the learning processing unit 114 extracts data as teacher data t from the questionnaire information 134 and the analysis information 136 according to the depth of the hierarchy selected as a processing target in the deep neural network DNN (S106). The teacher data t is the norm of the output data output from the output layer in the extracted neural network (in the above example, the neural network including the input layer, the first layer, and the output layer). Is the data.

  For example, the learning processing unit 114 extracts, as the teacher data t, the index value obtained when the analysis of the questionnaire result is further advanced as the layer to be processed is deeper. For example, a state where the multivariate analysis is performed on the questionnaire result can be regarded as a state where the analysis of the questionnaire result is further advanced as compared with a state where the multivariate analysis is not performed. In addition, the state in which the category for categorizing users based on the results of multivariate analysis has been determined can be regarded as a state in which the analysis of the questionnaire results has been further advanced compared to the state immediately after the multivariate analysis was performed. .

  Under such a premise, for example, when the first processing layer is the processing target, the learning processing unit 114 refers to the questionnaire information 134 that is raw data that has not been analyzed, for answering the questionnaire. The score of the selected option is extracted as teacher data t. In addition, for example, when the second layer is a processing target, the learning processing unit 114 refers to the analysis information 136 that is data on which factor analysis has been performed, and extracts a factor score as teacher data t. For example, when the third layer is a processing target, the learning processing unit 114 refers to the analysis information 136 and extracts a numerical value indicating a category determined based on the factor score as the teacher data t. That is, when the first layer is to be processed, the learning processing unit 114 sets the hyperparameter of the item (1) considered at the time of the above-described questionnaire result analysis as the teacher data t and the second layer as the processing target. When the hyperparameter of the item (2) considered in the analysis of the questionnaire result is the teacher data t and the third layer is the processing target, the hyperparameter of the item (3) considered in the analysis of the questionnaire result is the teacher data. Let t. As described above, the learning processing unit 114 sets different types of data (index values) as the teacher data t of each layer. As a result, the knowledge obtained by proceeding with the analysis can be learned step by step by the neural network in each layer of the hidden layer.

Next, the learning processing unit 114 evaluates an error between the output data h output from the output layer and the teacher data t extracted according to the depth of the processing target hierarchy in the extracted neural network. Derived as function I (S108). For example, the evaluation function I can be expressed as (1/2) × (ht) ² . The evaluation function I may be the sum of errors (square errors) between the data output from each of all units in the output layer and the teacher data t.

  Next, the learning processing unit 114 determines the weight W and the bias b of each layer so as to minimize the evaluation function I using the error back propagation method (S110). For example, the learning processing unit 114 derives the gradient ∂I / ∂W of the evaluation function I by calculating a partial derivative of the evaluation function I with respect to the weight W, and the weight W and the bias based on the gradient ∂I / ∂W. b is determined.

  Next, the learning processing unit 114 associates the weight W and bias b of each layer with the hierarchy to be processed, and stores the associated information in the storage unit 130 as layer-specific parameter information 140 (S112).

  FIG. 10 is a diagram illustrating an example of the parameter information 140 for each layer. As in the illustrated example, the layer-by-layer parameter information 140 is information in which the determined weight W and the bias b are associated with the processing target layer.

  Next, the learning processing unit 114 determines whether or not all layers 1 to n of the generated deep neural network DNN are determined as processing target layers (S114).

  If all hierarchies have not been determined as processing target hierarchies, the learning processing unit 114 adds 1 to the temporary parameter i (S116), and moves the process to S100 described above. Thereby, a hierarchy different from any hierarchy determined up to the previous time is determined as a hierarchy to be processed. In the above-described example, the first layer is determined as the processing target layer at the time of the first processing, and therefore the second layer is determined as the processing target layer at the second processing.

  On the other hand, when all the hierarchies are determined as processing target hierarchies, the learning processing unit 114 ends the process of this flowchart.

  In the process of the flowchart described above, the learning processing unit 114 determines a certain i-th layer among the n layers of the hidden layers in the process of S100, but the present invention is not limited to this. For example, the learning processing unit 114 may collectively determine a plurality of layers (for example, the first layer and the second layer) among the n layers of the hidden layers as a processing target layer. In this case, the learning processing unit 114 may learn the parameters of a plurality of layers determined as the processing target hierarchy.

FIG. 11 is a diagram schematically showing the contents of the loop processing of the flowchart shown in FIG. In the example of FIG. 10, it is assumed that there is one unit for each layer, and only the weight W is expressed as a parameter for each layer. For example, in the first process, the input layer, the first layer of the hidden layer, and the output layer are extracted. For example, the input data v input to the input layer is data obtained by digitizing a user's action history. In this case, in order to determine the weight W ₁ of the first layer, the first teacher data t ₁ corresponding to the depth of the first layer is extracted. For example, the first training data t _1, as described above, the score of the selected alternatives may be employed for the survey responses. At this time, it is assumed that the questionnaire result is an answer result by the user who was the operation subject of the action history input as the input data v. For example, when applied to the input layer to the action history of the user A, the first training data t _1, the survey results of the user A is utilized.

In the second process, for example, the input layer, the first and second layers of the hidden layer, and the output layer are extracted. For example, as the input data v input to the input layer, data obtained by quantifying the user's action history is employed, similarly to the data input to the input layer in the first process. For example, in the first process, when data that digitizes the action history of a user A is input to the input layer, the input data v input to the input layer in the second process is the action history of the user A. It becomes numerical data. In the second process, the learning processing unit 114, a second training data t _2, so as to reduce the error between the output data h which is output from the output layer, the first layer determined in the first process of The newly added weight W ₂ of the second layer is determined while maintaining the parameters. The second training data t ₂ is the data corresponding to the degree of depth of the hierarchy of the second layer, for example, as described above, factor score may be used which is obtained by factor analysis. For example, the learning processing unit 114 in the processing of the second time, by a weight W ₁ to the determined weight W _{1 #} of the first layer during the first process, to maintain the parameters of the first layer.

In the third process, for example, the input layer, the first to third layers of the hidden layer, and the output layer are extracted. For example, as the input data v input to the input layer, data input to the input layer in the first and second processing is adopted. In the third process, the learning processing unit 114, a third training data t _3, so as to reduce the error between the output data h which is output from the output layer, the first layer determined in the first process of parameters and to determine a third layer weights W ₃ of the newly added while maintaining the parameters of the second layer determined in the processing of the second time. The third teacher data t ₃ is data corresponding to the depth of the third layer, and for example, as described above, a numerical value indicating a category determined based on the factor score may be used. For example, the learning processing unit 114, in the third process, the weighting W _{1 #} of the first layer first and a weight W _1, which is determined at the time of treatment, the weight W _{2 #} of the second layer during the second process with determined weight W _2, to maintain the parameters of the first and second layers.

  Thus, every time the number of processes is repeated, a hidden layer is added, and data corresponding to the depth of the added hidden layer is used as teacher data, so that each layer of the hidden layer is added to the deep neural network DNN. Can be learned with high accuracy. In addition, since the parameters of each layer determined in the previous process are maintained in the next process, the occurrence of an internal covariate shift can be suppressed. This can prevent the learning time from becoming longer due to the occurrence of the internal covariate shift. Note that when the occurrence of the internal covariate shift is allowed, the learning processing unit 114 may re-determine the parameter determined in the previous process by the error back propagation of the current process. In this case, the learning processing unit 114 may update the parameters of each layer in the layer-by-layer parameter information 140.

  By referring to the DNN configuration information 138 and the layer parameter information 140, the classification processing unit 116 generates (re-creates) a deep neural network DNN in which the initial values of the weight W and the bias b of each layer are determined by prior learning. The user corresponding to the behavior history is classified into a predetermined category based on the behavior history of the unlearned user. “Unlearned” means, for example, that it is not used for learning at the stage of prior learning.

  FIG. 12 is a flowchart illustrating an example of processing executed by the classification processing unit 116. The processing of this flowchart is repeatedly performed at a predetermined cycle, for example.

  First, the classification processing unit 116 determines whether or not new action history information 132 has been acquired by the acquisition unit 112 (S200). When the new action history information 132 is acquired, the DNN configuration information 138 and the layer A deep neural network DNN in which initial values of parameters (weight W and bias b) for each layer are determined is generated based on the parameter information 140 (S202).

  Next, the classification processing unit 116 digitizes and inputs the action history information 132 indicating the action history of the user who answered the questionnaire in the input layer in the generated deep neural network DNN (S204).

  Next, the classification processing unit 116 outputs the output data h output from the output layer of the deep neural network DNN as a learning result by the deep neural network DNN (S206).

  For example, the classification processing unit 116 transmits the learning result by the deep neural network DNN to the service providing apparatus 20 and the advertisement distribution server apparatus 40 using the communication unit 106. Thereby, for example, the service providing apparatus 20 can transmit, for example, sale information corresponding to the category of the user classified by the deep neural network DNN to the terminal apparatus 10 used by the user. Moreover, the advertisement delivery server apparatus 40 can transmit the advertisement according to the user's category classified by the deep neural network DNN to the terminal apparatus 10 used by the user, for example.

  FIG. 13 is a diagram illustrating the learning accuracy of the deep neural network DNN learning method according to the present embodiment and the method exemplified as a comparative example. The method of the comparative example is, for example, a method of learning the parameters of each layer so that the output data h of each layer of the hidden layer approaches the input data v of that layer. That is, in the method of the comparative example, the output data h output from the i-th layer and the output data h of the (i−1) -th layer (= input data v to the i-th layer) are close to each other. This is a method for learning the parameters of the i-th layer. In the figure, the uppermost record and the middle record represent learning accuracy by the method of the comparative example, and the lowermost record represents learning accuracy by the method of the present embodiment.

  The method of the comparative example shown in the uppermost record has one input layer unit, 1000 first layer units, 1000 second layer units, and 5 output layer units. The learning accuracy is shown when the deep neural network DNN is used. The method of the comparative example shown in the middle record has one input layer unit, 1000 first layer units, 1000 second layer units, and X layer unit. Is a predetermined number and represents the learning accuracy when the deep neural network DNN is used when the number of units in the output layer is five. In the method of the present embodiment shown in the bottom record, there are one input layer unit, 1000 first layer units, 1000 second layer units, and Y layer. Represents the learning accuracy when the deep neural network DNN is used in the case where the number of units is the same as the number of units in the X-th layer and the number of units in the output layer is five.

  As in the example shown in the drawing, when the uppermost stage and the middle stage are compared, it can be seen that the learning accuracy is not necessarily improved even when the number of hidden layers is increased. In general, in the deep neural network DNN, problems such as overlearning easily occur and learning information (error) is not transmitted to a deeper layer are likely to occur. In order to deal with these problems, it is known that a large amount of data (for example, at least tens of thousands to several hundreds of thousands of dense data in the case of image data) is used as input data v and prior learning is performed.

  However, when sparse data as information is used as input data v, such as an action history on the Internet, a model that returns an appropriate learning result tends not to be generated. In addition, if the learning data that is the source of the input data v is sparse, the accuracy of the error back propagation method or the like is likely to decrease. Thus, in the case of the method of the comparative example, the learning accuracy tends to depend on the input data v used for the prior learning.

  On the other hand, the method of this embodiment does not cause the deep neural network DNN to learn so that the output data h and the input data v are close to each other in the preliminary learning stage (the input data v is changed to the teacher data t). In this case, the sparse data is used as the input data v by changing the index value obtained separately from the questionnaire results to the teacher data t while changing it according to the depth of each layer. Also, the learning accuracy can be improved.

  According to the embodiment described above, when the input data v is input, the first classifier (for example, hidden) included in the classification system (for example, deep neural network DNN) in which the classifiers that can be configured independently are configured in multiple layers. When the same input data v is input after the classification system is trained so that the output data h output from the first layer neural network) approaches the first teacher data t, the first classifier Learning result using the first teacher data so that the output data h output from the second classifier (for example, the neural network of the second layer of the hidden layer) deeper than the second hierarchy data approaches the second teacher data A learning processing unit 114 that learns the classification system based on the classification, and unlearned data based on the classification system learned by the learning processing unit 114 By providing a classification processing unit 116 classifies into categories, the parameters for each classifier can be set appropriately. As a result, the learning accuracy can be improved.

  In addition, according to the above-described embodiment, parameters are derived for each classifier in the pre-learning stage. Therefore, when some classifiers of the classification system are changed or added, classifications that do not change are performed. As for the container, the parameters already derived (learned parameters) can be used. For example, in the deep neural network DNN, when the classification result is changed from 5 patterns to 10 patterns, only the number of units in the output layer may be changed, and the hidden layer can be used as it is. As a result, a highly versatile deep neural network DNN can be constructed.

<Application example>
Hereinafter, application examples of the above-described embodiment will be described. For example, the analysis apparatus 100 according to the above-described embodiment uses the deep neural network DNN as text data constituting a web page as input data v in order to generate a summary sentence of the web page provided as a search result of a certain search query. Next, the parameters of each layer of the hidden layer are learned so as to output a summary sentence of the text data. At this time, the learning processing unit 114 of the analysis apparatus 100 sets the teacher data t of the output data h of each layer as text data in which the number of words decreases stepwise according to the depth of the hidden layer. For example, the learning processing unit 114 sets the teacher data t ₁ for the output data h of the first layer that is the shallowest (closest to the input layer) as text data having a first predetermined number of words. The first predetermined number is a number smaller than the number of words of the text data as the input data v. Also, the learning processing unit 114, the teacher data t ₂ for the output data h of the second layer, the number of words and a second predetermined number of the text data less than the first predetermined number. Thus, by reducing the number of words of text data used as the teacher data t as the hierarchy becomes deeper, it is possible to generate a web page summary sentence with higher accuracy.

Further, the analysis apparatus 100 according to the above-described embodiment uses each layer of the hidden layer so that an image is input data v and an appropriate recognition result is output to the deep neural network DNN when recognizing images such as characters and photographs. These parameters may be learned. At this time, the learning processing unit 114 of the analysis apparatus 100 does not use the teacher data t of the output data h of each layer as the correct answer data but as a recognition result by an unspecified number of users obtained by crowdsourcing or the like. For example, a predetermined image (for example, “cat image”) whose correct answer has been determined in advance is shown to a large number of unspecified users, and the recognition result of the image is set as teacher data t. At this time, a certain user A recognizes the visually recognized image as a “cat”, and a certain user B recognizes the visually recognized image as a “dog”. In this way, when the same image is shown to each of an unspecified number of users, some users may recognize differently from the correct recognition. Therefore, the teacher data t tends to be data that fluctuates with respect to correctness and data that includes both a correct recognition result and an incorrect recognition result. For example, the learning processing unit 114 recognizes the teacher data t ₁ for the output data h of the first layer among the recognition results obtained by an unspecified number of users obtained by crowdsourcing or the like (for example, the recognition result having the largest correctness / wrongness). positive: false = 0.5: 0.5), and the teacher data _{t 2} for the output data h of the second layer, a small recognition result of shaking of correctness than teacher data _{t 1} (e.g., positive: false = 0 .6: 0.4), and the teacher data _{t 3} for the output data h in the third layer, less recognition result of shaking of correctness than the teacher data _{t 2} (e.g., positive: false = 0.7: 0.3 ). In this way, data with more correct recognition results as the depth of the hierarchy becomes teacher data t, and teacher data t _n for the deepest n-layer output data h is correct data (correct: incorrect = 1: 0)). By doing so, the accuracy of image recognition can be improved.

<Hardware configuration>
Among the plurality of devices included in the analysis system 1 of the above-described embodiment, at least the analysis device 100 is realized by a hardware configuration as illustrated in FIG. 14, for example. FIG. 14 is a diagram illustrating an example of a hardware configuration of the analysis apparatus 100 according to the embodiment.

  The analysis device 100 includes a NIC 100-1, a CPU 100-2, a GPU (Graphics Processing Unit) 100-3, a RAM 100-4, a ROM 100-5, a secondary storage device 100-6 such as a flash memory and an HDD, and a drive device 100-. 7 are connected to each other by an internal bus or a dedicated communication line. The drive device 100-7 is loaded with a portable storage medium such as an optical disk. A program stored in a portable storage medium mounted on the secondary storage device 100-6 or the drive device 100-7 is expanded in the RAM 100-4 by a DMA controller (not shown) or the like, and the CPU 100-2 or the GPU 100-3. As a result, the control unit 110 is realized. The program referred to by the control unit 110 may be downloaded from another device via the network NW.

  As mentioned above, although the form for implementing this invention was demonstrated using embodiment, this invention is not limited to such embodiment at all, In the range which does not deviate from the summary of this invention, various deformation | transformation and substitution Can be added.

  DESCRIPTION OF SYMBOLS 1 ... Analysis system, 10 ... Terminal device, 20 ... Service providing device, 30 ... Log acquisition device, 40 ... Advertisement distribution server device, 100 ... Analysis device, 102 ... Reception part, 104 ... Display part, 106 ... Communication part, 110 ... Control unit, 112 ... Acquisition unit, 114 ... Learning processing unit, 116 ... Classification processing unit, 130 ... Storage unit, 132 ... Behavior history information, 134 ... Question information, 136 ... Analysis information, 138 ... DNN configuration information, 140 ... Parameter information for each layer

Claims (9)

When the first data is input, the second data output from the first classifier included in the classification system in which the classifiers that can be configured independently are configured to be close to the first teacher data After learning the classification system, when the first data is input, the third data output from the second classifier having a deeper hierarchy than the first classifier approaches the second teacher data. A learning processing unit for learning the classification system based on a learning result using the first teacher data;
A classification processing unit that classifies unlearned data into a predetermined category based on the classification system learned by the learning processing unit;
An analysis apparatus comprising: The first teacher data is a different type of data from the first data and the second data,
The second teacher data is a different type of data from the first data and the third data.
The analysis device according to claim 1. The first data is data indicating a user's behavior history,
The first teacher data is data based on the user's answer,
The second teacher data is data obtained by analyzing the answer of the user.
The analysis device according to claim 2. The first data is text data,
The first teacher data is text data having a number of words smaller than the number of words of the first data,
The second teacher data is text data having a number of words smaller than the number of words of the first teacher data.
The analysis device according to claim 2. The first data is image data;
The first teacher data is data indicating a recognition result of the image data by each of an unspecified number of users,
The second teacher data is data indicating a recognition result of the image data by each of an unspecified number of users, and is data having more correct recognition results than the first teacher data.
The analysis device according to claim 2. The learning processing unit
An error of the second data output from the output device with respect to the first teacher data is obtained by using an input device and an output device of the plurality of classifiers included in the classification system and the first classifier. Determining the parameters of the first classifier to be small;
Using the input device and the output device, and the first classifier and the second classifier, so as to reduce the error of the third data output from the output device with respect to the second teacher data, Determining the parameters of the second classifier while maintaining the determined parameters of the first classifier;
The analysis device according to any one of claims 1 to 5. When the first data is input, the second data output from the classification system in which the classifiers that can be configured independently are configured in multiple layers approaches the first data and the third data different from the second data. A learning processing unit for learning the classification system;
A classification processing unit that classifies unlearned data into a predetermined category based on the classification system learned by the learning processing unit;
An analysis apparatus comprising: Computer
When the first data is input, the second data output from the first classifier included in the classification system in which the classifiers that can be configured independently are configured to be close to the first teacher data After learning the classification system, when the first data is input, the third data output from the second classifier having a deeper hierarchy than the first classifier approaches the second teacher data. Learning the classification system based on a learning result using the first teacher data;
Classifying unlearned data into a predetermined category based on the learned classification system;
analysis method. On the computer,
When the first data is input, the second data output from the first classifier included in the classification system in which the classifiers that can be configured independently are configured to be close to the first teacher data After learning the classification system, when the first data is input, the third data output from the second classifier having a deeper hierarchy than the first classifier approaches the second teacher data. A process of learning the classification system based on a learning result using the first teacher data;
A process of classifying unlearned data into a predetermined category based on the learned classification system;
A program that executes

Images