JP6719399B2 - Analysis device, analysis method, and program - Google Patents

Description

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

2. Description of the Related Art Conventionally, a technique is known in which data relevance is automatically learned from past search behaviors and the learning is used to assist future search behaviors (for example, see Patent Document 1). Further, there is known a technique of collectively performing feature amount extraction and machine learning using the feature amount using a neural network (for example, see Patent Document 2).

JP, 2006-285982, A Japanese Patent Publication No. 2010-516642

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

The present invention has been made in view of the above problems, and an object of the present invention is to provide an analysis device, an analysis method, and a program that can improve learning accuracy.

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 multi-layered is the first teacher. After training the classification system so as 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. Unlearned data based on the learning processing unit for learning the classification system based on the learning result using the first teacher data and the classification system learned by the learning processing unit so as to approach the teacher data. And a classification processing unit that classifies the items into a predetermined category.

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

It is a figure showing an example of analysis system 1 containing analysis device 100 in an embodiment. It is a figure which shows an example of a structure of the analysis apparatus 100 in embodiment. It is a figure which shows 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 134. It is a figure which shows an example of the analysis information 136. It is a figure which shows an example of DNN structure information 138. It is a figure which shows typically the deep neural network DNN produced|generated based on the DNN structure information 138. 6 is a flowchart showing 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. 7 is a flowchart showing an example of processing executed by the classification processing unit 116. It is a figure which shows each learning precision 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 analysis apparatus 100 of embodiment.

Hereinafter, an analysis device, 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 inputting certain input data v (an example of the first data), the analysis device outputs the output data h output from the first classifier included in the classification system in which the classifiers that can be independently configured are included in the multilayer system. After learning the classification system so that (an example of the second data) approaches the first teacher data, when the input data v is input, the second classifier having a deeper hierarchy than the first classifier outputs 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 output approaches 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 layer 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. Of the layers of (1) to (3), the neural network is a layer deeper than at least the layer of the neural network used as the first classifier. The “deep layer” means a layer close to the output layer described later. Each of the “first classifier” and the “second classifier” may mean a neural network of a certain layer or a neural network of a plurality of layers.

Then, the analysis device classifies the unlearned data into a predetermined category based on the learned classification system. This can improve the accuracy of learning. Hereinafter, as an example, the classification system will be described as a deep neural network DNN configured by a plurality of neural networks.

[overall structure]
FIG. 1 is a diagram illustrating an example of an analysis system 1 including an analysis device 100 according to an embodiment. The analysis system 1 according to 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 the network NW. Note that some or all of the plurality of devices included in the analysis system 1 may be integrated in one analysis device 100.

Each device shown in FIG. 1 transmits and receives various information via the network NW. The network NW includes, for example, a wireless base station, a Wi-Fi access point, a communication line, a provider, the Internet and the like. Note that it is not necessary that all combinations of the respective devices illustrated in FIG. 1 can 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, for example, a mobile phone such as a smartphone, a tablet terminal, or a computer device such as a personal computer. For example, when the terminal device 10 receives a predetermined operation from the user, the terminal device 10 may communicate with the service providing device 20 via an application installed in advance to acquire the 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 or a search service, or an application that can enjoy an SNS (Social Networking Service), a mail service, an information providing service (for example, news or weather forecast). It may be.

In addition, when the terminal device 10 receives a predetermined operation from the user, the terminal device 10 may access the website provided by the service providing device 20 via a predetermined web browser. For example, the website provided by the service providing device 20 provides the same services as the services provided by the various applications described above.

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, or communicates with the terminal device 10 in which an application is activated to transfer various types of information. It may be an application server device that executes.

The log acquisition device 30 collects a cookie or a cache managed for each web browser, a log file of an HTTP (Hypertext Transfer Protocol) request, or the like from the terminal device 10 or the service providing device 20, and thereby the Acquire the behavior history of the user to use. The action history is, for example, purchase history in Internet shopping (type, price, number of purchased products or services, etc.), search history in search sites (search engine used, search query entered, etc.), It includes a viewing history on the moving image site (type of moving image, viewing time, etc.), an HTTP referrer (URL history) for identifying what kind of web page is accessed.

Further, 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 action history, similarly to the user's action history on the Internet via a web browser. You may get it.

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

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

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

The advertisement distribution server device 40 distributes an advertisement as the service providing device 20 provides a service to the terminal device 10, based on the analysis result by the analysis device 100, for example. For example, when the terminal device 10 receives a web page including an advertisement frame or a style sheet for an application corresponding thereto from the service providing device 20, based on an ad tag or SDK (Software Development Kit) provided in advance in the advertisement frame. , And sends an advertisement distribution request 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) capable of accessing the website of the advertisement client who has contracted in advance. It is transmitted to the terminal device 10. As a result, the advertisement image is displayed on the screen of the terminal device 10.

The analysis device 100 uses the deep neural network DNN to classify the user corresponding to the action history into a predetermined category based on the user's action history 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 device 100 provides the classification result of the category for each user to, for example, the advertisement distribution server device 40. Thereby, the advertisement distribution server device 40 can change the distribution content of the advertisement according to the classified categories.

[Analysis device configuration]
Hereinafter, the configuration of the analysis device 100 will be described with reference to the drawings. FIG. 2 is a diagram illustrating an example of the configuration of the analysis device 100 according to the embodiment. As illustrated, the analysis device 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 reception unit 102 is an input interface such as a touch panel, a keyboard, and a mouse that receives an operation input from the 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) and a DMA (Direct Memory Access) controller. The communication unit 106 communicates with the log acquisition device 30, the advertisement distribution server device 40, or the like via the network NW. For example, the communication unit 106 receives the user's action history from the log acquisition device 30 and stores it in the storage unit 130 as the action history information 132.

FIG. 3 is a diagram showing an example of the action history information 132. As in the illustrated example, the action history information 132 is associated with the time to which address the position of the terminal device 10 corresponds to, or is input to the search engine name used and its search engine. It is the information that is associated with the search query, or that the title name of the browsed website and the URL of the website are associated with each other. Further, the action history information 132 may be, for example, information in which the purchased product name or service name is associated with the URL of the website from which they were purchased with respect to the time. As a matter of course, the user does not take all imaginable actions with respect to a certain event, but usually takes one or a few actions, at most tens to hundreds actions. is there. For example, it is extremely rare that one user purchases all the products out of the products sold at a shopping site, and usually only purchases some products. Therefore, the behavior history of the user 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 history.

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, by a processor such as a CPU (Central Processing Unit) or a GPU (Graphics Processing Unit) executing a program stored in the storage unit 130. Further, some or all of the constituent elements of the control unit 110 may be realized by hardware such as an LSI (Large Scale Integration), an ASIC (Application Specific Integrated Circuit), or an FPGA (Field-Programmable Gate Array). , May be realized by the 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), and the like. The storage unit 130 stores various programs such as firmware and application programs and the action history information 132 described above, as well as questionnaire information 134, analysis information 136, DNN configuration information 138, and parameter information 140 for each layer. These pieces of information will be described later.

The acquisition unit 112 acquires information indicating the result of the questionnaire (survey) stored in the external server by communicating with an external server (not shown) using the communication unit 106, and acquires the information from the questionnaire information 134. Is stored in the storage unit 130.

When the service providing apparatus 20 provides a questionnaire monitor site or a questionnaire monitor service capable of answering a questionnaire 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.

Further, for example, when the administrator or the like who manages the analysis device 100 inputs the result of the questionnaire 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 showing an example of the questionnaire. As in the illustrated example, the questionnaire is prepared with predetermined options for the standardized survey items (Q1 to 4 etc. in the figure). In the illustrated example, five options are prepared, but the number is not limited to this, and may be two, three, four, or six or more.

FIG. 5 is a diagram showing an example of the questionnaire information 134. As in the illustrated example, the questionnaire information 134 is, for example, information in which a score corresponding to the option selected by the user is associated with the survey item. At least part of the user of the questionnaire information 134 and the user of the action history information 132 are the same. That is, the action history of the user who is the respondent to the questionnaire is acquired as the action history information 132.

For example, in the above-mentioned questionnaire, the higher the answer option for the survey item, the higher the score. For example, a positive answer option such as "I think so" or "I think so" has a higher score than a negative answer option such as "I don't think so" or "I don't think so at all". To be done. Note that this relationship is an example, and a negative answer option for a survey item may be set to have a higher score.

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

For example, in the analysis of the questionnaire results, (1) how the survey items included in the questionnaire were designed, (2) what kind of analysis method to quantitatively evaluate, and (3) the questionnaire as a result of the analysis. It is carried out in consideration of matters such as how to classify the users who answered to. For example, with respect to (1), whether the response method to the survey item is the option format or the description format, and if it is the option format, is the option expressed in text or numerical value? , Etc. need to be considered. Regarding (2), by using a score according to the user's answer, for example, by performing factor analysis, what is the factor that contributes to the characteristic of the user who answered the questionnaire, or in the characteristic of the user, It is necessary to consider how much the factors contribute. Further, when performing the principal component analysis instead of the factor analysis, it is necessary to consider which variable is the principal component among a plurality of variables that may represent the characteristics of the user. Regarding (3), as a final analysis result, which user and which user are considered to have the same tendency based on 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, analysts who analyze the results of the questionnaire perform a proper analysis for the purpose of conducting the questionnaire, and a parameter that is a statistical classification index for classifying the users who answered the questionnaire with a certain degree of granularity. Need to be derived. Such parameters are handled as hyperparameters that analysts and the like need to set appropriately.

For example, when the categories are “high annual income” and “low annual income” as the hyperparameter of (3), the hyperparameter (analysis index) of (2) has a high annual income among the users who answered the questionnaire. A parameter common to users or a factor common to users with low annual income may be used as a parameter.

FIG. 6 is a diagram showing an example of the analysis information 136. In the illustrated example, the result of the questionnaire is shown when the factor analysis is performed in consideration of the items (1) to (3) described above. For example, the analysis information 136 is information in which a factor score (FAC1 or the like in the drawing) considered in the factor analysis is associated with a factor score for each user who answers the questionnaire. The factor score is, for example, an index value indicating the degree of correlation between the factor and the characteristics of the user. In the illustrated example, with respect to the factor FAC1, 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. Represents. These factor scores may be calculated, for example, based on the scores corresponding to the answer options selected by the user among the options prepared for the survey items of the questionnaire.

Further, as in the illustrated example, the analysis information 136 may indicate a classification result indicating what category each user who has answered the questionnaire has been classified. The category to be classified by the user may be derived by comprehensively considering a plurality of classification indexes such as the factor score described above. For example, in the case of a user in which all the factor scores of factors FAC1 to FAC4 are positive values, the category of the user is set to type 1, and in the case of a user in which any one of the factor scores is a negative value, the category of the user is If the user has type 2 and any two factor scores have a negative value, the category of the user is type 3, and if any of the users has three or more factor scores with a negative value, the category of the user is You may classify as type 4.

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

FIG. 7 is a diagram showing 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. Generally, the deep neural network DNN is composed of three layers of 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. The 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 represented by a function called an activation function (transfer function) and returns an output according to an 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 to this, and an arbitrary function may be used. The input layer neural network is an example of an “input device”, and the output layer neural network is an example of an “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 number of hidden layers is at least two or more, and in the illustrated example, the number of hidden layers is n (n≧2). In addition, the number of units in each hidden layer is r. Note that the number of units in each hidden layer does not have to be uniformly r, and may be different in each layer. The number of layers of hidden layers and the number of units of 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. Further, the number of hidden layer units may be increased in order to improve learning accuracy.

The learning processing unit 114 refers to such DNN configuration information 138 to generate a deep neural network DNN having a predetermined configuration during pre-learning.

FIG. 8 is a diagram schematically showing a deep neural network DNN generated based on the DNN configuration information 138. As in the illustrated example, the deep neural network DNN is a neural 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. The edges 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 proximity to the input layer side is expressed as “shallow”, and the proximity to the output layer side is expressed as “deep”. In the above example, 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 next to the first layer (hereinafter, referred to as a second layer) among the n hidden layers. To be done. An edge is connected from each unit included in the second layer to each of a plurality of units of the shallowest layer next to the second layer (hereinafter referred to as the third layer) among the n layers of the hidden layer. To be done. Similarly, for the units of the third and subsequent layers, the edges are connected to the units of the deeper layers. An edge is connected from each unit of the deepest nth layer to one or more units of the output layer. In this way, the deep neural network DNN is represented as a state probability model in which units in each layer are connected by edges, such as a constrained Boltzmann machine (RBM). The deep neural network DNN may be expressed as a state probability model in which units belonging to the same layer are further connected by an edge like a Boltzmann machine without restriction.

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

[Pre-learning]
The pre-learning process will be described below with reference to the flowchart. FIG. 9 is a flowchart showing an example of processing executed by the learning processing unit 114.

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

For example, the learning processing unit 114 determines the first layer (i=1) as the processing target layer. The learning processing unit 114 may determine the second layer, the third layer, or the like other than the first layer as the layer to be processed. In the following description, it is assumed that the first layer is first determined as the layer to be processed.

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

For example, the learning processing unit 114 refers to the action history information 132 and derives the number of views of the web page, the number of times a predetermined search query is input, and the like, and sets this as input data v. Further, the learning processing unit 114 vectorizes a character string (which may include alphabets, numbers, symbols, etc.) related to each action history such as the product code of the purchased product or the URL of the browsing site, and this vector is used as the vector. The input data v for the input layer may be used. The learning processing unit 114 derives the 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 represented by the following mathematical expression (1), for example.

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 the unique bias component of each layer.

Next, the learning processing unit 114 extracts data to be the teacher data t from the questionnaire information 134 and the analysis information 136 according to the depth of the hierarchy selected as the 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-described example, the neural network including three layers of the input layer, the first layer, and the output layer). Data.

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

Based on such a premise, for example, when the first processing layer 114 is a processing target, the learning processing unit 114 refers to the questionnaire information 134 that is raw data that has not been analyzed, and returns the questionnaire. The score of the option selected for is extracted as teacher data t. In addition, for example, when the second layer is the processing target, the learning processing unit 114 refers to the analysis information 136 that is the data on which the factor analysis has been performed, and extracts the factor score as the teacher data t. Further, 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 the category determined based on the factor score as the teacher data t. That is, when the first layer is the processing target, the learning processing unit 114 sets the hyperparameter of the item (1) considered in the analysis of the above-mentioned questionnaire result as the teacher data t, and when the second layer is the processing target. , The hyperparameter of the item (2) considered in the analysis of the questionnaire result is the teacher data t, and when 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 t. t. In this way, 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 advancing the analysis can be learned stepwise by the neural network of each hidden layer.

Next, the learning processing unit 114 evaluates the error between the output data h output from the output layer and the teacher data t extracted according to the depth of the layer to be processed in the neural network including the extracted layer. It is derived as a 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 the errors (squared errors) between the data output from all the units in the output layer and the teacher data t.

Next, the learning processing unit 114 uses the error backpropagation method to determine the weight W and bias b of each layer so as to minimize the evaluation function I (S110). For example, the learning processing unit 114 derives the gradient ∂I/∂W of the evaluation function I by calculating a partial differential with respect to the weight W of the evaluation function I, and the weight W and the bias based on the gradient ∂I/∂W. Determine b.

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

FIG. 10 is a diagram showing an example of the layer-by-layer parameter information 140. 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 layer to be processed.

Next, the learning processing unit 114 determines whether or not all the hierarchies 1 to n of the generated deep neural network DNN have been decided as hierarchies to be processed (S114).

When all the hierarchies are not determined as the hierarchies to be processed, the learning processing unit 114 adds 1 to the temporary parameter i (S116), and moves the process to S100 described above. As a result, a layer different from any of the layers determined up to the previous time is determined as the layer to be processed. In the above-described example, the first layer is determined as the layer to be processed at the time of the first processing, and thus the second layer is determined as the layer to be processed at the time of the second processing, for example.

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

In the process of the flowchart described above, the learning processing unit 114 determines, in the process of S100, a certain i-th layer among the n-layers of the hidden layers as the process target layer, 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) out of the n hidden layers as a layer to be processed. In this case, the learning processing unit 114 may collectively learn the parameters of a plurality of layers determined as the layer to be processed.

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 the number of units of each layer is one, and only the weight W is expressed as a parameter of each layer. For example, in the first processing, the input layer, the first hidden layer, and the output layer are extracted. For example, the input data v input to the input layer is data that digitizes the behavior history of the user. In this case, in order to determine the weight W ₁ of the first layer, the first teacher data t ₁ according to the degree of the depth of the first layer is extracted. For example, as described above, the score of the option selected for answering the questionnaire may be adopted as the first teacher data t ₁ . At this time, the questionnaire result is assumed to be the answer result by the user who was the subject of the action history input as the input data v. For example, when the behavior history of the user A is input to the input layer, the questionnaire result of the user A is used as the first teacher data t ₁ .

In the second processing, for example, the input layer, the first and second hidden layers, and the output layer are extracted. For example, as the input data v input to the input layer, similarly to the data input to the input layer in the first process, data that digitizes the behavior history of the user is adopted. For example, in the first process, when the numerical data of the action history of a certain 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 will be digitized data. In the second process, the learning processing unit 114 reduces the error between the second teacher data t ₂ and the output data h output from the output layer, and reduces the error in the first layer determined in the first process. The newly added second layer weight W ₂ is determined while maintaining the parameters. The second teacher data t ₂ is data according to the degree of the depth of the second layer, and for example, as described above, the factor score obtained by the factor analysis may be used. 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 processing, for example, the input layer, the first to third hidden layers, and the output layer are extracted. For example, as the input data v input to the input layer, the data input to the input layer in the first and second processes is adopted. In the third processing, the learning processing unit 114 reduces the error between the third teacher data t ₃ and the output data h output from the output layer, and reduces the error of the first layer determined in the first processing. The newly added weight W ₃ of the third layer is determined while maintaining the parameters and the parameters of the second layer determined in the second processing. The third teacher data t ₃ is data according to the degree of the depth of the hierarchy 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 By using the determined weight W ₂ , the parameters of the first layer and the second layer are maintained.

As described above, the hidden layer is added every time the number of times of processing is increased, and the data corresponding to the depth of the added hidden layer is used as the teacher data, so that each 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 processing are maintained in the next processing, the occurrence of internal covariate shift can be suppressed. This makes it possible to prevent the learning time from becoming longer due to the occurrence of the internal covariate shift. In addition, when allowing the occurrence of the internal covariate shift, the learning processing unit 114 may redetermine the parameters determined in the previous processing by the error back propagation in the current processing. In this case, the learning processing unit 114 may update the parameters of each layer in the layer-specific parameter information 140.

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

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

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

Next, the classification processing unit 116 digitizes and inputs the action history information 132 indicating the action history of the user who has 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 of the deep neural network DNN to the service providing device 20 or the advertisement distribution server device 40 using the communication unit 106. Thereby, for example, the service providing apparatus 20 can transmit the sale information according to the category of the user classified by the deep neural network DNN to the terminal device 10 used by the user. Further, the advertisement distribution server device 40 can transmit, for example, an advertisement according to the category of the user classified by the deep neural network DNN to the terminal device 10 used by the user.

FIG. 13 is a diagram showing the learning accuracy of each of the learning method of the deep neural network DNN in this embodiment and the method exemplified as a comparative example. The method of the comparative example is a method of learning the parameters of each layer such that the output data h of each layer of the hidden layer approaches the input data v of the layer. That is, the method of the comparative example uses the hidden layer so that 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 of learning the parameters of the i-th layer of. The uppermost record and the middle record in the figure 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 unit in the input layer, 1000 units in the first layer, 1000 units in the second layer, and 5 units in the output layer. The learning accuracy is shown when the deep neural network DNN is used for each case. In the method of the comparative example shown in the middle record, the input layer unit is 1, the first layer unit is 1000 units, the second layer unit is 1000 units, and the X layer unit is Is a predetermined number and represents the learning accuracy when using the deep neural network DNN in the case where the number of units in the output layer is 5. In the method of the present embodiment shown in the record at the bottom, the number of units in the input layer is 1, the number of units in the first layer is 1000, the number of units in the second layer is 1000, and the number of units in the Y layer is Shows the learning accuracy when using the deep neural network DNN in the case where the number of units is the same as the number of units in the Xth layer described above and the number of units in the output layer is five.

As in the example shown in the figure, comparing the uppermost stage and the middle stage shows that the learning accuracy does not necessarily improve even when the number of hidden layers is increased. In general, the deep neural network DNN is apt to cause problems such as over-learning and learning information (error) not being transmitted to deeper layers. In order to deal with these, it is known that enormous data (for example, at least tens of thousands to hundreds of thousands of dense data in the case of image data) is used as the input data v, and pre-learning is performed.

However, when sparse data as information is used as the input data v like the action history on the Internet, a model that returns an appropriate learning result tends not to be generated. Further, if the learning data that is the source of the input data v is sparse, the accuracy of the error backpropagation method and the like tends to decrease. As described above, in the case of the method of the comparative example, the learning accuracy tends to depend on the input data v used for the pre-learning.

On the other hand, in the method of the present embodiment, in the pre-learning stage, the deep neural network DNN does not perform learning so that the output data h and the input data v come close to each other (the input data v is set as the teaching data t However, the sparse data is used as the input data v by changing the index value separately obtained from the questionnaire result to the teacher data t while changing the index value according to the depth of each layer. Also, learning accuracy can be improved.

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

Further, according to the above-described embodiment, since the parameters are derived for each classifier at the stage of pre-learning, when changing or adding some classifiers of the classification system, the classification that is not changed The already derived parameters (learned parameters) can be used for the container. For example, in the deep neural network DNN, when changing the classification result from 5 patterns to 10 patterns, only the number of units in the output layer needs to 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 device 100 in the above-described embodiment uses the text data forming the web page as the input data v to generate the summary of the web page provided as the search result of a certain search query, and uses the deep neural network DNN. Trains the parameters of each hidden layer to output a summary of the text data. At this time, the learning processing unit 114 of the analysis device 100 sets the teacher data t of the output data h of each layer to text data in which the number of words gradually decreases 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, which 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 used as the input data v. Further, the learning processing unit 114 sets the teacher data t ₂ for the output data h of the second layer as the second predetermined number of text data having the number of words smaller than the first predetermined number. In this way, by reducing the number of words of the text data to be the teacher data t as the hierarchy becomes deeper, it is possible to more accurately generate the abstract sentence of the web page.

Further, the analysis device 100 in the above-described embodiment uses each image as an input data v when recognizing an image such as a character or a photograph, and outputs each of the hidden layers so as to output an appropriate recognition result to the deep neural network DNN. Parameters may be learned. At this time, the learning processing unit 114 of the analysis device 100 does not use the teacher data t of the output data h of each layer as the correct answer data, but the recognition result by an unspecified large number of users obtained by crowdsourcing or the like. For example, a predetermined image (for example, “cat image”) whose correct answer is determined in advance is shown to an unspecified number of users, and the recognition result of the image is used as the 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 a large number of unspecified users, some users may make a different recognition from the correct recognition. Therefore, the teacher data t tends to be data that has fluctuations with respect to correctness, and tends to be data that includes both correct recognition results and erroneous recognition results. For example, the learning processing unit 114 recognizes the teacher data t ₁ with respect to the output data h of the first layer as a recognition result having the largest fluctuation of correctness among recognition results by an unspecified number of users obtained by crowdsourcing or the like (for example, , Correct: false=0.5:0.5), and the recognition result of the teacher data t ₂ for the output data h of the second layer is smaller than the teacher data t _{1 in} the fluctuation of correctness (for example, correct: false=0. .:6:0.4) and the teacher data t _{3 corresponding} to the output data h of the third layer has a smaller recognition fluctuation than the teacher data t ₂ (for example, correct:error=0.7:0.3). ). In this way, the data having more correct recognition results as the hierarchy becomes deeper is the teacher data t, and the teacher data t _n for the output data h of the deepest n layer is the correct answer data (correct: false=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, for example, the hardware configuration shown in FIG. FIG. 14 is a diagram illustrating an example of the hardware configuration of the analysis device 100 of 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 or an HDD, and a drive device 100-. 7 are connected to each other by an internal bus or a dedicated communication line. A portable storage medium such as an optical disk is attached to the drive device 100-7. 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. By being executed by the control unit 110, the control unit 110 is realized. The program referenced by the control unit 110 may be downloaded from another device via the network NW.

Although the embodiments for carrying out the present invention have been described above with reference to the embodiments, the present invention is not limited to these embodiments, and various modifications and substitutions are made without departing from the gist of the present invention. Can be added.

1... Analysis system, 10... Terminal device, 20... Service providing device, 30... Log acquisition device, 40... Advertisement distribution server device, 100... Analysis device, 102... Reception unit, 104... Display unit, 106... Communication unit, 110 ... control unit, 112... acquisition unit, 114... learning processing unit, 116... classification processing unit, 130... storage unit, 132... action history information, 134... questionnaire information 136... analysis information, 138... DNN configuration information, 140... Parameter information for each layer

Claims (7)

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 included in the multi-layered system is 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;
Based on the classification system learned by the learning processing unit, a classification processing unit that classifies unlearned data into a predetermined category,
An analysis device having. The first data is data indicating a behavior history of the user,
The first teacher data is data based on the answer of the user,
The second teacher data is data obtained by analyzing the answer of the user,
The analysis device according to claim 1 . 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 1 . 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 has more correct recognition results than the first teacher data.
The analysis device according to claim 1 . The learning processing unit,
Using the input device and the output device of the plurality of classifiers included in the classification system and the first classifier, the error of the second data output from the output device with respect to the first teacher data is calculated. 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, to reduce an 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 4 . 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 included in the multi-layered system is 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. Train the classification system based on a learning result using the first teacher data,
Classifying unlearned data into predetermined categories 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 included in the multi-layered system is 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 to execute.

Images