JP2018077671A - Information processing apparatus, information processing method, apparatus for generating prediction models, method for generating prediction models and program - Google Patents

Abstract

An information processing apparatus, an information processing method, a prediction model generation method, and a program capable of further improving prediction accuracy are provided. A plurality of first prediction units for generating a plurality of first prediction results representing the probability of occurrence of a predetermined event by analyzing target data by executing different algorithms, respectively, A plurality of first prediction units that generate the first prediction result based on a prediction model that has learned a data group, and a plurality of results obtained by analyzing a blending data group including correct data different from the correct data group using a plurality of prediction models Based on the prediction result, an adjustment unit that adjusts an algorithm for generating the second prediction result and an algorithm adjusted by the adjustment unit cause a predetermined event to occur based on the first prediction result And a second prediction unit that generates a second prediction result representing the probability. [Selection] Figure 3

Description

  The present invention relates to an information processing device, an information processing method, a prediction model generation device, a prediction model generation method, and a program.

  2. Description of the Related Art Conventionally, a technique is known in which a user predicts whether to click distribution information such as an advertisement or a news article, and selects distribution information based on a prediction result. In relation to this, a technique called stacking for generating a final prediction result based on a plurality of prediction results generated by a plurality of learners is known (see Non-Patent Document 1).

Sill Joseph, et al. “Feature-Weighted Linear Stacking” arXiv preprint arXiv: 0911.0460 (2009)

  However, in the conventional technology, even if information such as advertisements and news articles is distributed based on the prediction result, information that is not of interest to the user is rarely clicked, so the information distribution effect is not improved. There was a problem.

  The present invention has been made in consideration of such circumstances, and an information processing device, an information processing method, a prediction model generation device, a prediction model generation method, and a program that can further improve prediction accuracy. One of the purposes is to provide.

  One aspect of the present invention is a plurality of first prediction units that generate a plurality of first prediction results representing the probability that a predetermined event occurs by analyzing target data by executing different algorithms. Then, a plurality of first prediction units that generate the first prediction result based on a prediction model that has learned the correct data group, and a blending data group that includes correct data different from the correct data group are analyzed using the plurality of prediction models. An adjustment unit that adjusts an algorithm for generating a second prediction result based on a plurality of prediction results as a result, and an algorithm adjusted by the adjustment unit, thereby executing the algorithm based on the first prediction result And a second prediction unit that generates the second prediction result representing the probability that the predetermined event will occur.

  According to one embodiment of the present invention, prediction accuracy can be further improved.

1 is a configuration diagram illustrating an example of a content providing system 1. FIG. 2 is a configuration diagram illustrating an example of a content distribution apparatus 200. FIG. 2 is a configuration diagram illustrating an example of an information analysis apparatus 400. FIG. 5 is a configuration diagram illustrating an example of a prediction unit 430. FIG. 12 is a flowchart illustrating an example of a flow of algorithm adjustment processing in a prediction result generation unit 450. It is a figure which shows an example of the prediction process of the target data by linear separation. It is a figure which shows an example of the prediction process of the target data by linear separation and nonlinear separation. 2 is a diagram illustrating an example of a hardware configuration of an information analysis apparatus 400. FIG.

  DESCRIPTION OF EMBODIMENTS Hereinafter, embodiments of an information processing apparatus, an information processing method, a prediction model generation method, a prediction model generation apparatus, and a program according to the present invention will be described with reference to the drawings.

<Outline of Embodiment>
In the content providing system 1 according to the embodiment, a plurality of first prediction units that predict the probability of occurrence of a predetermined event by analyzing target data and a result predicted by the plurality of first prediction units And a second prediction unit that predicts the probability that an event will occur. The content providing system 1 first acquires a plurality of probabilities by predicting the probability of occurrence of a predetermined event by each of a plurality of first prediction units having different prediction algorithms, and then, by a second prediction unit, Based on the plurality of probabilities, the probability that a predetermined event occurs is further predicted.

  A 1st prediction part produces | generates a 1st prediction result based on the prediction model which learned the correct data group. The content providing system 1 further includes an algorithm for generating a second prediction result based on a plurality of prediction results as a result of analyzing a blending data group including correct data different from the correct data group using a plurality of prediction models. adjust. A 2nd prediction part produces | generates the 2nd prediction result showing the probability that a predetermined event will occur based on a 1st prediction result by performing the adjusted algorithm. Thereby, the content provision system 1 can improve the precision of prediction.

  The predetermined event is, for example, when a news title or advertising content is clicked when a news title or advertising content is posted on a web page. The probability that a predetermined event occurs is, for example, the probability that a news title or advertisement content is clicked. The predetermined event may be the number of times a news title or advertising content is clicked during a predetermined period, and the probability that a predetermined event occurs is the number of times a news title or advertising content is clicked during a predetermined period. May be a probability that exceeds a predetermined number. The content providing system 1 predicts the probability that a news title or advertising content will be clicked when posting a news title or advertising content in a limited area of a web page, Advertising content can be distributed.

<Configuration of content providing system 1>
FIG. 1 is a configuration diagram illustrating an example of a content providing system 1. The content providing system 1 includes, for example, a terminal device 100, a content distribution device 200, a data server 300, and an information analysis device 400. In the example of FIG. 1, a terminal device 100, a content distribution device 200, a data server 300, and an information analysis device 400 are connected to a network NW. Each device connected to the network NW includes a communication interface such as a NIC (Network Interface Card) or a wireless communication module (not shown in FIG. 1). 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 for all combinations of these components to be able to communicate with each other, and the network NW may partially include a local network. Further, it is not necessary that all combinations of these components can communicate with each other, and the network NW may partially include a local network. The content distribution device 200, the data server 300, and the information analysis device 400 may be partly or wholly integrated into a single device.

  The terminal device 100 is a device used by a user (general user). The terminal device 100 is a computer device (communication device) such as a mobile phone such as a smartphone, a tablet terminal, or a personal computer, for example. The terminal apparatus 100 transmits an HTTP (Hypertext Transfer Protocol) request to the content distribution apparatus 200 when a user agent (UA) such as a web browser is activated and a user performs a predetermined operation. And the terminal device 100 produces | generates a web screen based on the content sent in response from the content delivery apparatus 200, and displays it on a display part. An arbitrary application program may be used as a means for acquiring the content instead of the web browser. The terminal device 100 selects one of the contents included in the web screen by receiving a user operation. The terminal device 100 transmits a content browsing request associated with the selected content to the content distribution device 200.

  The content distribution device 200 is a server device that provides content such as news articles and advertisements, for example. The content distribution device 200 provides content in response to the content browsing request received from the terminal device 100. FIG. 2 is a configuration diagram illustrating an example of the content distribution apparatus 200. The content distribution device 200 includes, for example, a content storage unit 210, a request reception unit 220, and a content selection unit 230.

  The content storage unit 210 is realized by, for example, an HDD (Hard Disk Drive), a flash memory, a RAM (Random Access Memory), a ROM (Read Only Memory), or a hybrid storage device combining a plurality of these. Further, part or all of the content storage unit 210 may be an external device accessible by the content distribution device 200, such as NAS (Network Attached Storage) or an external storage server. The content storage unit 210 stores content data to be provided to the terminal device 100.

  Part or all of the functional units including the request reception unit 220 and the content selection unit 230 are performed by hardware such as LSI (Large Scale Integration), ASIC (Application Specific Integrated Circuit), or FPGA (Field-Programmable Gate Array). It may be realized or may be realized by cooperation of software and hardware. The request reception unit 220 receives a browsing request received from the terminal device 100. In response to receiving the browsing request, the request receiving unit 220 transmits information included in the browsing request to the information analysis apparatus 400 as target data. Information included in the browsing request is, for example, user information or a query. The content selection unit 230 selects content to be distributed to the terminal device 100 among the content stored in the content storage unit 210 based on the prediction result of the information analysis device 400.

  The data server 300 is a server device including, for example, an HDD, a flash memory, a RAM, a ROM, or a hybrid storage device that combines a plurality of these. A part or all of the data server 300 may include an external device accessible by the data server 300, such as NAS (Network Attached Storage) or an external storage server. The data server 300 accumulates target data to be analyzed in the information analysis apparatus 400. Specifically, the data server 300 includes user information such as an attribute of a user who operates the terminal device 100, a query transmitted from the terminal device 100 to the content distribution device 200, content distributed from the content distribution device 200 to the terminal device 100, distribution Various data such as the clicked content of the selected content and the history data such as the number of clicks of each content are stored. The content as the target data is, for example, a word that appears in a news article.

"Information analysis device 400"
FIG. 3 is a configuration diagram illustrating an example of the information analysis apparatus 400. The information analysis apparatus 400 includes, for example, a correct data collection unit 410, a data storage unit 420, a plurality of prediction units 430-1, ... 430-N, a target data reception unit 440, and a prediction result generation unit 450. , And an adjustment unit 460. In the following description, when the prediction unit is not distinguished from other prediction units, it is described as “prediction unit 430”. N is a natural number of 2 or more. Some or all of the functional units including the correct data collection unit 410, the plurality of prediction units 430-1,... 430-N, the target data reception unit 440, and the prediction result generation unit 450 are LSI, ASIC, or FPGA. It may be realized by hardware such as, or may be realized by cooperation of software and hardware.

  The correct data collection unit 410 collects correct data from the data server 300. The correct answer data is data obtained by adding tag information as a correct answer to the same kind of data as the target data. Specifically, the correct answer data is data in which tag information indicating whether or not the title of the news article is clicked is added to the content data indicating the title of the news article. The correct data collection unit 410 causes the data storage unit 420 to store the collected correct data.

  The data storage unit 420 is realized by, for example, an HDD, a flash memory, a RAM, a ROM, or a hybrid storage device that combines a plurality of these. Further, part or all of the data storage unit 420 may be an external device accessible by the information analysis device 400, such as a NAS or an external storage server. The data storage unit 420 stores the correct data collected by the correct data collection unit 410. The correct data stored in the data storage unit 420 is classified into correct data for learning the prediction unit 430, blending data for adjusting the combination of the prediction units 430, and test data for evaluating the prediction results. Is done.

“Configuration of Prediction Unit 430”
The prediction unit 430 predicts the probability that a predetermined event will occur by analyzing the target data using a prediction algorithm. The prediction algorithm of a certain prediction unit 430 is different from the prediction algorithms of other prediction units 430. Examples of the prediction algorithm include machine learning algorithms such as a neural network, GBDT (Gradient Boosting Decision Tree), logistic regression, SVM (Support Vector Machine), and random forest.

  FIG. 4 is a configuration diagram illustrating an example of the prediction unit 430. The prediction unit 430 includes a learning unit 432, a model parameter storage unit 434, and a prediction processing unit 436. The learning unit 432 generates a prediction model based on the correct answer data. For example, when the prediction model is a model using a neural network, the learning unit 432 generates and updates processing parameters in the neural network. The learning unit 432 stores the generated and updated processing parameters in the model parameter storage unit 434 as the processing parameters of the prediction model. A prediction model is software that executes a prediction algorithm according to a processing parameter as a learning result. The model parameter storage unit 434 stores the processing parameters of the prediction model. The model parameter storage unit 434 may also be used as a storage unit that stores the processing parameters generated by the other prediction units 430. The prediction processing unit 436 is supplied with the target data. The prediction processing unit 436 reads out the processing parameters of the prediction model from the model parameter storage unit 434. The prediction processing unit 446 predicts the probability that a predetermined event will occur for the target data according to the read processing parameter. Specifically, the prediction processing unit 436 predicts the probability that the title of the news article is clicked when content data representing the title of the news article is supplied as target data.

  Returning to FIG. The target data receiving unit 440 receives target data. Specifically, the target data receiving unit 440 receives a plurality of content data from the content distribution device 200 as target data.

“Prediction Result Generation Unit 450 and Adjustment Unit 460”
The prediction result generation unit 450 is based on the prediction results of the plurality of prediction units 430-1 to 430-N (hereinafter referred to as a plurality of first prediction results), and indicates a prediction result (hereinafter referred to as the first). 2 prediction results). The prediction result generation unit 450 executes an algorithm for generating the second prediction result. The algorithm for generating the second prediction result is adjusted by the adjustment unit 460. The algorithm for generating the second prediction result represents, for example, a calculation procedure for generating the second prediction result by weighting and adding the plurality of first prediction results. Further, the prediction result generation unit 450 may generate the second prediction result by calculating a plurality of first prediction results by a predetermined prediction algorithm such as SVM.

  The adjustment unit 460 adjusts an algorithm for generating a second prediction result based on a plurality of prediction results as a result of analyzing a blending data group including correct data different from the correct data group using a plurality of prediction models.

“Adjustment of Algorithm in Prediction Result Generation Unit 450”
FIG. 5 is a flowchart illustrating an example of the algorithm adjustment processing flow in the prediction result generation unit 450. First, the information analysis apparatus 400 performs a learning process by the plurality of prediction units 430 using training data (S10). Next, the information analysis apparatus 400 performs a prediction process using blending data by each of the plurality of prediction units 430 that has performed the learning process (S20). Next, the information analysis apparatus 400 adjusts, for example, a combination of the prediction units 430 as the algorithm adjustment processing in the prediction result generation unit 450 using the prediction results of the plurality of prediction units 430 (S30). For example, the information analysis apparatus 400 changes the combination of the prediction unit 430 that outputs the first prediction result to the prediction result generation unit 450 and the prediction unit 430 that does not output the first prediction result to the prediction result generation unit 450. Next, the information analysis apparatus 400 performs a prediction process by using the test data by the plurality of prediction units 430 and the prediction result generation unit 450 whose combinations are adjusted, and performs a process of evaluating the prediction result (S40).

Prediction process
FIG. 6 is a diagram illustrating an example of target data prediction processing based on linear separation. It is assumed that the information analysis apparatus 400 has received target data D1 including a plurality of clicked news article titles and target data D2 including a plurality of news article titles not clicked. The information analysis apparatus 400 performs a prediction unit 430 that executes a prediction model capable of linearly separating the target data D1 and D2 with the straight line L1 as a boundary, and linearly separates the target data D1 and D2 with the straight line L2 as a boundary. A prediction unit 430 that executes a prediction model capable of Thereby, the prediction result production | generation part 450 can produce | generate the 2nd prediction result isolate | separated by both the straight lines L1 and L2. Specifically, the prediction result generation unit 450 can perform the prediction process according to the OR condition of both the straight lines L1 and L2.

  FIG. 7 is a diagram illustrating an example of target data prediction processing based on linear separation and nonlinear separation. The information analysis apparatus 400 performs a prediction unit 430 that executes a prediction model capable of linearly separating the target data D1 and D2 with the straight line L1 as a boundary, and nonlinearly separates the target data D1 and D2 with the curve L3 as a boundary. A prediction unit 430 that executes a prediction model capable of Thereby, the prediction result production | generation part 450 can produce | generate the 2nd prediction result isolate | separated by both the straight line L1 and the curve L3.

  Note that the information analysis apparatus 400 predicted whether or not the title of the news article was clicked, but the present invention is not limited to this. In addition, user information such as the gender of the user is added and the title of the news article is clicked by the female user. It may be predicted whether or not.

  As described above, according to the content providing system 1 of the embodiment, a plurality of prediction units 430 (first prediction unit) generated as a prediction model that learned a correct data group, and correct data different from the correct data group are obtained. Based on the plurality of prediction results predicted by each of the prediction units 430 with respect to the blending data group including the prediction data, the prediction unit 430 for generating the second prediction result is selected from the plurality of prediction units 430, and the selected prediction Based on the first prediction result generated by the unit 430, the prediction result generation unit 450 that generates the second prediction result representing the probability that a predetermined event occurs is provided, so that the single prediction unit 430 performs the prediction process. The prediction accuracy can be increased.

<Hardware configuration>
FIG. 8 is a diagram illustrating an example of a hardware configuration of the information analysis apparatus 400. The information analysis apparatus 400 has a configuration in which, for example, a NIC 401, a CPU 402, a RAM 403, a ROM 404, a secondary storage device 405 such as a flash memory or HDD, and a drive device 406 are connected to each other via an internal bus or a dedicated communication line. Yes. The drive device 406 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 405 or the drive device 406 is expanded in the RAM 403 by a DMA controller (not shown) or the like, and executed by the CPU 402, whereby the information analysis device 400 A functional part is realized.

  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 ... Content provision system, 100 ... Terminal device, 200 ... Content delivery apparatus, 230 ... Content selection part, 300 ... Data server, 400 ... Information analysis apparatus, 410 ... Correct data collection part, 420 ... Data storage part, 430 ... Prediction , 432 ... learning unit, 434 ... model parameter storage unit, 436 ... prediction processing unit, 440 ... target data reception unit, 446 ... prediction processing unit, 450 ... prediction result generation unit

Claims (6)

A plurality of first prediction units that generate a plurality of first prediction results representing the probability of occurrence of a predetermined event by analyzing target data by executing different algorithms, and learning a correct data group A plurality of first prediction units that generate the first prediction result based on the predicted model,
An adjustment unit for adjusting an algorithm for generating a second prediction result based on a plurality of prediction results as a result of analyzing a blending data group including correct data different from the correct data group by the plurality of prediction models;
A second prediction unit that generates the second prediction result representing the probability that the predetermined event occurs based on the first prediction result by executing the algorithm adjusted by the adjustment unit;
An information processing apparatus comprising: A plurality of prediction models that are executed to generate a plurality of first prediction results representing the probability that a predetermined event occurs by analyzing target data by executing different algorithms. Generate a prediction model that learned
Adjusting an algorithm for generating a second prediction result based on a plurality of prediction results as a result of analyzing a blending data group including correct data different from the correct data group by the plurality of prediction models,
Generating the second prediction result representing the probability that the predetermined event occurs based on the first prediction result by executing an adjusted algorithm;
Information processing method. On the computer,
A plurality of prediction models that are executed to generate a plurality of first prediction results representing the probability that a predetermined event occurs by analyzing target data by executing different algorithms. Generate a prediction model that learned
Based on a plurality of prediction results as a result of analyzing a blending data group including correct data different from the correct data group by the plurality of prediction models, an algorithm for generating a second prediction result is adjusted,
Executing the adjusted algorithm to generate the second prediction result representing the probability that the predetermined event occurs based on the first prediction result;
program. A plurality of prediction models that generate a plurality of first prediction results that represent the probability of occurrence of a predetermined event by analyzing target data by executing different algorithms, each of which is a prediction model that has learned a correct data group A generating unit for generating
An adjustment unit that adjusts an algorithm for generating a second prediction result based on a plurality of prediction results as a result of analyzing a blending data group including correct data different from the correct data group by the plurality of prediction models; An apparatus for generating a prediction model. A plurality of prediction models that generate a plurality of first prediction results that represent the probability of occurrence of a predetermined event by analyzing target data by executing different algorithms, each of which is a prediction model that has learned a correct data group Produces
Adjusting an algorithm for generating a second prediction result based on a plurality of prediction results as a result of analyzing a blending data group including correct data different from the correct data group by the plurality of prediction models;
How to generate a prediction model. On the computer,
A plurality of prediction models that generate a plurality of first prediction results that represent the probability of occurrence of a predetermined event by analyzing target data by executing different algorithms, each of which is a prediction model that has learned a correct data group To generate
Adjusting an algorithm for generating a second prediction result based on a plurality of prediction results as a result of analyzing a blending data group including correct data different from the correct data group by the plurality of prediction models,
program.

Images