Learning device, learning method, and learning program

Abstract

This invention provides a learning device, learning method, and learning program for appropriately assigning labels to unlabeled data in groups. [Solution] In a learning system in which a business terminal and a learning device are connected to each other via a network, the learning device comprises an acquisition unit and a learning unit. The acquisition unit acquires a first data set in which labels are not missing and a second data set in which labels are missing, regarding the user's behavior. The learning unit performs training for supervised regression based on the first data set and training for semi-supervised regression based on the second data set for a predictor that predicts behavior based on input data.

Description

Technical Field

[0001] The present invention relates to a learning device, a learning method, and a learning program.

Background Art

[0002] Conventionally, a technique for assigning labels to unlabeled data is known. For example, a technique for learning including unlabeled data and applying it to unlabeled data for prediction is known.

Prior Art Documents

Non-Patent Documents

[0003]

Non-Patent Document 1

Non-Patent Document 2

Summary of the Invention

Problems to be Solved by the Invention

[0004] However, conventional technologies were unable to properly label unlabeled data, for example, at the group level.

[0005] This application was made in view of the above, and aims to appropriately assign labels to unlabeled data on a group basis. [Means for solving the problem]

[0006] The learning device according to the present invention is characterized by comprising: an acquisition unit that acquires a first data set in which labels are not missing and a second data set in which labels are missing regarding the user's behavior; and a learning unit that performs learning for supervised regression based on the first data set and learning for semi-supervised regression based on the second data set for a predictor that predicts behavior based on input data. [Effects of the Invention]

[0007] According to one embodiment, the effect is achieved that labels can be appropriately assigned to unlabeled data on a group basis. [Brief explanation of the drawing]

[0008] [Figure 1] Figure 1 shows an example of the configuration of a learning system according to an embodiment. [Figure 2] Figure 2 is an explanatory diagram for illustrating the conventional technology. [Figure 3] Figure 3 is an explanatory diagram illustrating the dataset according to the embodiment. [Figure 4] Figure 4 is Figure (1), which shows an example of data for calculation according to the embodiment. [Figure 5] Figure 5 is Figure (2), which shows an example of data for calculation according to the embodiment. [Figure 6] Figure 6 is Figure (3) showing an example of data for calculation according to the embodiment. [Figure 7A]FIG. 7A is a diagram (1) showing experimental results. [Figure 7B] FIG. 7B is a diagram (2) showing experimental results. [Figure 8] FIG. 8 is a diagram showing an example of information processing according to an embodiment. [Figure 9] FIG. 9 is a diagram showing a configuration example of a business operator terminal according to an embodiment. [Figure 10] FIG. 10 is a diagram showing a configuration example of a learning device according to an embodiment. [Figure 11] FIG. 11 is a diagram showing an example of a user information storage unit according to an embodiment. [Figure 12] FIG. 12 is a diagram showing an example of a predictor storage unit according to an embodiment. [Figure 13] FIG. 13 is a flowchart showing an example of a procedure of information processing according to an embodiment. [Figure 14] FIG. 14 is a hardware configuration diagram showing an example of a computer for realizing the functions of a learning device.

Mode for Carrying Out the Invention

[0009] Hereinafter, embodiments for implementing a learning device, a learning method, and a learning program according to the present application (hereinafter referred to as "embodiments") will be described in detail with reference to the drawings. Note that the learning device, the learning method, and the learning program according to the present application are not limited by this embodiment. Also, in the following embodiments, the same parts are denoted by the same reference numerals, and redundant explanations are omitted.

[0010] (Embodiment) [1. Configuration of Learning System] The learning system 1 shown in FIG. 1 will be described. As shown in FIG. 1, the learning system 1 includes a business operator terminal 10 and a learning device 100. The business operator terminal 10 and the learning device 100 are communicably connected by wire or wirelessly via a predetermined communication network (network N). FIG. 1 is a diagram showing a configuration example of the learning system 1 according to an embodiment.

[0011] The operator terminal 10 is an information processing device used by, for example, an operator who analyzes the behavior of users on a given service. The operator of the operator terminal 10 is, for example, an operator specialized in analyzing user behavior. Also, the operator terminal 10 is an information processing device used by, for example, an operator who predicts the behavior of users (specifically, predicts as a classification problem or a regression problem) and provides a service suitable for the users. The operator of the operator terminal 10 is, for example, an operator specialized in providing services to users.

[0012] The operator terminal 10 may be any device as long as it can realize the processing in the embodiment. Also, the operator terminal 10 may be a device such as a smartphone, a tablet terminal, a notebook PC, a desktop PC, a mobile phone, or a PDA. FIG. 8 shows the case where the operator terminal 10 is a smartphone.

[0013] The operator terminal 10 is, for example, a smart device such as a smartphone or a tablet, and is a portable terminal device that can communicate with an arbitrary server device via a wireless communication network such as 4G - 5G (Generation) or LTE (Long Term Evolution). Also, the operator terminal 10 has a screen such as a liquid crystal display and has a screen with a touch panel function, and may receive various operations on display data such as content, such as tap operations, slide operations, and scroll operations, by a finger or a stylus from the operator. In FIG. 8, the operator terminal 10 is used by the operator P1.

[0014] The learning device 100 is an information processing device aimed at appropriately attaching labels to label - free data in group units, and may be any device as long as it can realize the processing in the embodiment. The learning device 100, for example, acquires the behavior information of users and predicts the behavior of users. For example, the learning device 100 predicts the behavior of users by performing learning on a predictor that predicts the behavior of users.

[0015] As will be explained in detail later, there are two types of training performed on the predictor. One is training based on a dataset that already has labels assigned to it (i.e., no missing data), and the other is training based on a dataset that does not yet have labels assigned to it (i.e., data is missing). The former uses a dataset that already has labels assigned to it, so it is training for supervised regression, and the latter uses a dataset that does not yet have labels assigned to it, so it is training for semi-supervised regression (unsupervised regression may also be used). In the following embodiments, supervised regression and semi-supervised regression may be appropriately replaced with supervised classification and semi-supervised classification, respectively.

[0016] The learning device 100 is implemented, for example, by the server equipment or cloud system of the service provider terminal 10. For example, the learning device 100 is implemented by the server equipment or cloud system of a service provider specializing in analyzing user behavior. Alternatively, for example, the learning device 100 is implemented by the server equipment or cloud system of a service provider specializing in providing services to users.

[0017] [2. An example of information processing] Conventionally, techniques for assigning labels to unlabeled data are known. For example, a technique is known in which a model is trained on unlabeled data and then applied to the unlabeled data to be predicted.

[0018] Conventional techniques like these are limited to cases where there is only one group, and data is obtained from randomly selected samples within that single group.

[0019] Furthermore, examples of methods for labeling unlabeled data include methods for solving classification problems and methods for solving regression problems. Classification problems involve solving binary values ​​such as 0 or 1, while regression problems involve solving numerical values. This reveals the class (category) of the data.

[0020] An example of a classification problem is to conduct a sample survey on whether or not a person supports a political party and predict which users did not participate in the survey. Another example of a regression problem is to conduct a sample survey on the number of times a person has been involved in a traffic accident and predict which users did not participate in the survey.

[0021] Figure 2 illustrates the application of such classification and regression problems. Figure 2 is an explanatory diagram for illustrating the conventional technique, specifically the case where data is obtained randomly from a single group.

[0022] The symbols "〇," "×," and "△" in Figure 2 represent data sampled from the population. Of these, "〇" and "×" are labeled data, while "△" is unlabeled data. "〇" represents a positive example in a classification problem or a value greater than 0 in a regression problem, for example, indicating support for a political party. "×" represents a negative example in a classification problem or 0 in a regression problem, for example, indicating non-support for a political party. The numbers in parentheses accompanying "〇" represent positive values ​​in the regression problem.

[0023] To predict unlabeled data, such as the "△" data, there are two approaches: training the predictor using only labeled data, or training the predictor using both labeled and unlabeled data. The latter approach, including unlabeled data in the training, is expected to improve accuracy.

[0024] Here, we give an example of calculation formulas when applying them to a classification problem. The following calculation formulas (1A), (1B), and (1C) represent PN classification, PU classification, and PNPU ​​classification, respectively. PN classification is a classification based only on data with labels, PU classification is a classification based on data including data without labels, and PNPU ​​classification is a classification that combines PN classification and PU classification. Furthermore, the following calculation formula (1D) shows the notation for marginal probability used in the following calculation formulas (1A), (1B), and (1C). In addition, "assuming..." in the following calculation formula (1B) indicates that the data without labels are randomly sampled from the positive and negative data.

[0025]

number

number

number

number

[0026] Here, we consider an example of a calculation formula when applying regression problems. The calculation formulas (2A), (2B), and (2C) below represent PN regression, PU regression, and PNPU ​​regression, respectively. All of these are regressions that increase the probability of a value of zero. PN regression is a regression based only on data with labels, PU regression is a regression based on data including unlabeled data, and PNPU ​​regression is a regression that combines PN regression and PU regression. Furthermore, the calculation formula (2D) below shows the notation for marginal probabilities used in the calculation formulas (2A), (2B), and (2C) below. In addition, "assuming..." in the calculation formula (2B) below indicates that the unlabeled data is data randomly sampled from positive and negative examples.

[0027]

number

number

number

number

[0028] In this way, the calculation formula for classification problems can be extended to be applied to regression problems. However, this is the case when sample data is obtained randomly within a single group, and it is not possible to appropriately assign labels to unlabeled data on a group-by-group basis when there are multiple groups.

[0029] This application was made in view of the above, and aims to appropriately assign labels to unlabeled data on a group basis.

[0030] One example of a classification problem involving labeling unlabeled data at the group level is when you want to judge the merits of a new business at the user level. In such a case, you might have access to data on the effectiveness of similar businesses, and you've been able to get feedback from a few users using a prototype, but you want to predict the effects on a wider range of users.

[0031] Furthermore, an example of a regression problem for assigning labels to unlabeled data at the group level is predicting the number of conversions for samples where conversions are missing in a given operating system (OS) (or given web browser).

[0032] Figure 3 illustrates the application of such classification and regression problems. Figure 3 is an explanatory diagram illustrating the dataset according to the embodiment. In Figure 3, the dataset contains multiple groups.

[0033] Group PN is a group to which a label has been assigned, while Group PU is a group to which no label has been assigned. Note that even within Group PU, some data may be labeled. For example, positive and negative examples may be predetermined for some data. In this case, there may be a bias in the number of positive and negative examples.

[0034] Here, since the distribution of group PU differs from that of group PN, applying a predictor trained on group PN to group PU may result in lower accuracy. For example, if group PN is a group of men and group PU is a group of women, it is conceivable that applying a predictor trained on a group of different genders will not yield accurate predictions. Therefore, training of both group PN and group PU is necessary.

[0035] Here, we give an example of a calculation formula when applying the classification problem according to the embodiment. The calculation formulas (3A), (3B), and (3C) below represent PN classification, PU classification, and PNPU ​​classification, respectively. PN classification is a classification based on group PN, PU classification is a classification based on group PU, and PNPU ​​classification is a classification that combines PN classification and PU classification. Furthermore, the calculation formula (3D) below shows the notation for the probabilities used in the calculation formulas (3A), (3B), and (3C) below. In addition, "assuming..." in the calculation formula (1B) below indicates that although the distributions of PN and PU are different, it is assumed that the ratio of positive examples to negative examples for PU as a whole is the same as that of PN. Note that this assumption may be changed at will.

[0036]

number

number

number

number

[0037] Furthermore, we will give an example of a calculation formula when applying the regression problem according to the embodiment. The calculation formulas (4A), (4B), and (4C) below represent PN regression, PU regression, and PNPU ​​regression, respectively. All of them are regressions in which the probability of a value of zero is high. PN regression is a regression based on group PN, PU regression is a regression based on group PU, and PNPU ​​regression is a regression that combines PN regression and PU regression. In addition, the calculation formula (4D) below shows the notation of probabilities used in the calculation formulas (4A), (4B), and (4C) below. Furthermore, "assuming..." in the calculation formula (4B) below indicates that although the distributions of PN and PU are different, it is assumed that the ratio of positive examples to negative examples for PU as a whole is the same as that of PN. Note that this assumption may be changed at will.

[0038]

number

number

number

number

[0039] Next, to give a concrete example, we will use the data shown in Figure 4 to calculate the following formula (5). In the following formula (5), R PU Find (g). Then, R PUThe model is trained to minimize (g). Furthermore, the following calculation formula (6) is calculated using the data shown in Figure 5. In the following calculation formula (6), the expected values ​​of the PN and PU labels are calculated, and the model is trained to have the same expected values, R PU (g) is determined. Then, the following calculation formula (7) is calculated. In the following calculation formula (7), R r PNPU Find (g). Then, R r PNPU The system is trained to minimize (g).

[0040]

number

number

number

[0041] The entire PU is trained to meet the assumptions (the expected values ​​match). When limited to labeled data, the prediction may deviate downwards, and when limited to unlabeled data, the prediction may deviate upwards. By compensating for the downward deviation of the prediction when labeling with unlabeled data, the upward deviation of the prediction when unlabeled data can be suppressed. Finally, using the data shown in Figure 6, the following calculation formula (8) is calculated.

[0042]

number

[0043] Next, I will explain the actual experimental results. In certain operating systems (or certain web browsers), ITP (Intelligent Tracking Prevention) is enabled by default, and conversions for actions prior to a certain number of days may not be obtained. For example, conversions can be obtained if ITP is disabled. The problem to be solved is to compensate for conversions that could not be measured by ITP.

[0044] Here, the first operating system is designated as group PN, and the second operating system as group PU. The dataset consists of combinations of features such as "demographic data," "distribution platform," "ad category," and "number of conversions up to a specified number of days prior," and labels for "number of conversions up to a specified number of days prior."

[0045] Figure 7(A) is Figure (1) showing the experimental results. In PN learning, PN was "0.35783" and PU was "0.20389", showing a significant difference in prediction strength between PN and PU. On the other hand, in PNPU ​​learning, PN was "0.36015" and PU was "0.36508", showing no significant difference in prediction strength between PN and PU, indicating that PU can predict with a relatively similar magnitude. The prediction strength is calculated using the following formula (9).

[0046]

number

[0047] Figure 7(B) is Figure (2) showing the experimental results. PN learning is represented by a filled bar graph, while PNPU ​​learning is represented by a transparent bar graph. It can be seen that PN learning cannot account for differences in distribution, resulting in unstable allocations relative to dates. On the other hand, in PNPU ​​learning, the distribution is average relative to dates.

[0048] Figure 8 shows an example of information processing according to the embodiment. The learning device 100 acquires user information (user information of multiple users) (step S101). The learning device 100 also acquires (or generates, predicts, extracts, and identifies) a data set with labels and a data set without labels regarding the actions of the users (multiple users) based on the acquired user information (step S102). The learning device 100 also simultaneously performs training on a predictor that predicts actions for supervised regression based on the data set with labels and for semi-supervised regression (unsupervised regression may also be used) based on the data set without labels (step S103). When the learning device 100 receives input data (step S104), it provides the output result from the predictor (step S105).

[0049] [3. Configuration of the carrier terminal] Next, the configuration of the carrier terminal 10 according to the embodiment will be described using Figure 9. Figure 9 is a diagram showing an example of the configuration of the carrier terminal 10 according to the embodiment. As shown in Figure 9, the carrier terminal 10 has a communication unit 11, an input unit 12, an output unit 13, and a control unit 14.

[0050] (Communications Section 11) The communication unit 11 is implemented, for example, by a NIC (Network Interface Card). The communication unit 11 is connected to a predetermined network N by wire or wireless connection and sends and receives information to and from the learning device 100 via the predetermined network N.

[0051] (Input section 12) The input unit 12 receives various operations from the service provider. In Figure 8, it receives various operations from service provider P1. For example, the input unit 12 may receive various operations from the service provider via a touch panel display. Alternatively, the input unit 12 may receive various operations from buttons on the service provider terminal 10, or from a keyboard or mouse connected to the service provider terminal 10.

[0052] (Output section 13) The output unit 13 is a display screen for a tablet device, for example, which is implemented using a liquid crystal display or an organic EL (Electro-Luminescence) display, and is a display device for displaying various types of information. For example, the output unit 13 displays information transmitted from the learning device 100. For example, the output unit 13 displays prediction results.

[0053] (Control Unit 14) The control unit 14 is, for example, a controller, and is implemented by a CPU (Central Processing Unit) or MPU (Micro Processing Unit) executing various programs stored in the internal memory of the business terminal 10 using RAM (Random Access Memory) as the working area. For example, these various programs include application programs installed on the business terminal 10. For example, these various programs include application programs that display information transmitted from the learning device 100. The control unit 14 is also implemented by an integrated circuit such as an ASIC (Application Specific Integrated Circuit) or an FPGA (Field Programmable Gate Array).

[0054] As shown in Figure 9, the control unit 14 has a receiving unit 141 and a transmitting unit 142, and realizes or executes the information processing operations described below.

[0055] (Receiver 141) The receiving unit 141 receives, for example, information transmitted from the learning device 100. For example, the receiving unit 141 receives prediction results. For example, the receiving unit 141 receives label assignment results.

[0056] (Transmitter 142) The transmitting unit 142 transmits information to, for example, the learning device 100. For example, the transmitting unit 142 transmits user information. For example, the transmitting unit 142 transmits user behavior information on a predetermined service. The transmitting unit 142 also transmits input data from the service provider.

[0057] [4. Configuration of the learning device] Next, the configuration of the learning device 100 according to the embodiment will be described using Figure 10. Figure 10 is a diagram showing an example of the configuration of the learning device 100 according to the embodiment. As shown in Figure 10, the learning device 100 has a communication unit 110, a storage unit 120, and a control unit 130. The learning device 100 may also have an input unit (for example, a keyboard or mouse) that accepts various operations from the administrator of the learning device 100, and a display unit (for example, a liquid crystal display) for displaying various information.

[0058] (Communications Department 110) The communication unit 110 is implemented, for example, by a NIC. The communication unit 110 is connected to the network N by wire or wireless connection and sends and receives information to and from the carrier terminal 10 via the network N.

[0059] (Storage unit 120) The storage unit 120 is implemented by, for example, semiconductor memory elements such as RAM and flash memory, or storage devices such as hard disks and optical discs. As shown in Figure 10, the storage unit 120 has a user information storage unit 121 and a predictor storage unit 122.

[0060] The user information storage unit 121 stores user information. Here, Figure 11 shows an example of the user information storage unit 121 according to this embodiment. The information stored in the user information storage unit 121 is used, for example, for training a predictor. As shown in Figure 11, the user information storage unit 121 has items such as "User ID" and "User Information".

[0061] "User ID" indicates identification information used to identify a user. "User Information" indicates user information. In the example shown in Figure 11, conceptual information such as "User Information #1" and "User Information #2" is stored in "User Information," but in reality, text data indicating the actions performed by the user and the date and time of those actions is stored.

[0062] The predictor storage unit 122 stores the predictors. Here, Figure 12 shows an example of the predictor storage unit 122 according to this embodiment. The information stored in the predictor storage unit 122 is used, for example, to retrieve predictors. As shown in Figure 12, the predictor storage unit 122 has items such as "predictor ID" and "predictor".

[0063] The "Predictor ID" indicates identification information for identifying the predictor. The "Predictor" indicates the predictor. In the example shown in Figure 12, conceptual information such as "Predictor #1" and "Predictor #2" is stored in "Predictor," but in reality, parameter information of the predictor is stored there.

[0064] (Control unit 130) The control unit 130 is a controller, and is implemented, for example, by a CPU or MPU executing various programs stored in the memory device inside the learning device 100 using RAM as the working area. Alternatively, the control unit 130 can be implemented by an integrated circuit, such as an ASIC or FPGA.

[0065] As shown in Figure 10, the control unit 130 includes an acquisition unit 131, a learning unit 132, a prediction unit 133, and a provision unit 134, and realizes or executes the information processing operations described below. Note that the internal configuration of the control unit 130 is not limited to the configuration shown in Figure 10, and other configurations are also acceptable as long as they perform the information processing described later.

[0066] (Acquisition part 131) The acquisition unit 131 acquires various information from the storage unit 120. The acquisition unit 131 also stores the acquired information in the storage unit 120.

[0067] The acquisition unit 131 acquires various information from external information processing devices. The acquisition unit 131 also acquires various information from other information processing devices such as the business terminal 10.

[0068] The acquisition unit 131 acquires, for example, user information. For example, the acquisition unit 131 acquires user behavior information. For example, the acquisition unit 131 acquires user conversion behavior information.

[0069] The acquisition unit 131 acquires, for example, a data set (corresponding to the first data set) in which labels (e.g., labels indicating whether the user converted or not, labels indicating user information, etc.) are not missing regarding the user's behavior, and a data set (corresponding to the second data set) in which such labels are missing. For example, the acquisition unit 131 acquires the first data set in which labels are not missing because certain conditions are met (e.g., the user has allowed tracking), and the second data set in which labels are missing because such conditions are not met. Alternatively, for example, the acquisition unit 131 acquires the first data set and the second data set simultaneously. Alternatively, for example, the acquisition unit 131 acquires the second data set in which the sample size is the same as that of the first data set.

[0070] The acquisition unit 131 acquires, for example, input data. For example, the acquisition unit 131 acquires input data received from a business operator.

[0071] (Learning Section 132) The learning unit 132 trains a predictor that predicts behavior as a classification or regression problem. For example, based on the information acquired by the acquisition unit 131, the learning unit 132 trains a predictor that predicts behavior based on input data, performing supervised regression based on the first data set and semi-supervised regression (or unsupervised regression) based on the second data set. For example, the learning unit 132 simultaneously performs supervised regression based on the first data set and semi-supervised regression based on the second data set. The predictor takes user behavior information as input and outputs the user's prediction result (label assignment result). As mentioned above, regression in these processes can be rephrased as classification.

[0072] The learning unit 132 performs training to perform semi-supervised regression based on the second data set so that the number of positive examples (or negative examples) is the same as that of the first data set. The learning unit 132 also performs training to perform semi-supervised regression based on the second data set so that the distribution of positive examples (or negative examples) is the same as that of the first data set.

[0073] (Prediction unit 133) The prediction unit 133, for example, inputs the input data acquired by the acquisition unit 131 into a predictor trained by the learning unit 132 to make predictions about the user's behavior. The prediction unit 133 also assigns labels based on the prediction results, for example.

[0074] (Provider 134) The providing unit 134 provides, for example, the prediction results from the prediction unit 133. For example, the providing unit 134 provides the label assignment results based on the prediction results.

[0075] [5. Information Processing Flow] Next, the information processing procedure by the learning system 1 according to the embodiment will be explained using Figure 13. Figure 13 is a flowchart showing the information processing procedure by the learning system according to the embodiment.

[0076] As shown in Figure 13, the learning device 100 acquires user information (step S201). The learning device 100 acquires a first data set with complete labels and a second data set with missing labels regarding the user's behavior (step S202). The learning device 100 performs training on the predictor for supervised regression based on the first data set and for semi-supervised regression based on the second data set (step S203).

[0077] [6. Effects] As described above, the learning device 100 according to the embodiment includes an acquisition unit 131 and a learning unit 132. The acquisition unit 131 acquires a first data set in which labels are not missing and a second data set in which labels are missing, regarding the user's behavior. The learning unit 132 performs training on a predictor that predicts behavior based on input data, for supervised regression based on the first data set and for semi-supervised regression based on the second data set.

[0078] As a result, the learning device 100 according to this embodiment can train its predictor to appropriately assign labels to unlabeled data using, for example, two types of data sets: a data set with labels intact and a data set with labels missing. Furthermore, because the learning device 100 according to this embodiment can appropriately assign labels to unlabeled data, for example, the cost of labeling can be reduced. In addition, because the learning device 100 according to this embodiment does not require strong assumptions about the distribution as in the prior art, for example, it can appropriately predict any data without a decrease in accuracy.

[0079] Furthermore, the acquisition unit 131 acquires a first data group and a second data group regarding actions that include actions related to conversions.

[0080] As a result, the learning device 100 according to this embodiment can train the predictor so that it can appropriately assign labels even in the case of actions related to conversions.

[0081] Furthermore, the acquisition unit 131 acquires a first data set in which labels indicating whether or not the user has converted are not missing, and a second data set in which labels indicating whether or not the user has converted are missing.

[0082] As a result, the learning device 100 according to this embodiment can train the predictor so that it can appropriately assign labels indicating whether or not a user has converted.

[0083] Furthermore, the acquisition unit 131 acquires a first data group in which labels indicating user information are not missing, and a second data group in which labels indicating user information are missing.

[0084] As a result, the learning device 100 according to this embodiment can train a predictor so that it can appropriately assign labels indicating user information, for example.

[0085] Furthermore, the acquisition unit 131 acquires a first data set in which labels are not missing because the user has permitted tracking, and a second data set in which labels are missing because the user has not permitted tracking.

[0086] As a result, the learning device 100 according to this embodiment can train its predictor to appropriately assign labels based on, for example, two sets of data determined by the user's tracking settings.

[0087] Furthermore, the learning unit 132 performs training to perform semi-supervised regression based on a second data set that has the same number of samples as the first data set.

[0088] As a result, the learning device 100 according to this embodiment can, for example, train a predictor that takes into account the number of samples.

[0089] Furthermore, the learning unit 132 performs training to perform semi-supervised regression based on the second data set so that the number of positive examples is the same as that of the first data set.

[0090] As a result, the learning device 100 according to this embodiment can, for example, train a predictor that takes into account the number of positive examples.

[0091] Furthermore, the learning unit 132 performs learning to perform semi-supervised regression based on the second data set so that the distribution of the first data set and the distribution of positive examples become the same.

[0092] As a result, the learning device 100 according to this embodiment can, for example, train a predictor that takes into account the distribution of positive examples.

[0093] [7. Hardware Configuration] Furthermore, the learning device 100 according to the above-described embodiment is realized by a computer 1000 having the configuration shown in Figure 14. Figure 14 is a hardware configuration diagram showing an example of a computer that realizes the functions of the learning device 100. The computer 1000 has a CPU 1100, RAM 1200, ROM 1300, HDD 1400, communication interface (I / F) 1500, input / output interface (I / F) 1600, and media interface (I / F) 1700.

[0094] The CPU 1100 operates based on programs stored in the ROM 1300 or HDD 1400, and controls various parts. The ROM 1300 stores boot programs executed by the CPU 1100 when the computer 1000 starts up, as well as programs that depend on the computer 1000's hardware.

[0095] The HDD1400 stores programs executed by the CPU1100, as well as data used by such programs. The communication interface1500 acquires data from other devices via a predetermined communication network and sends it to the CPU1100, and transmits data generated by the CPU1100 to other devices via the predetermined communication network.

[0096] The CPU 1100 controls output devices such as displays and printers, and input devices such as keyboards and mice, via the input / output interface 1600. The CPU 1100 acquires data from input devices via the input / output interface 1600. The CPU 1100 also outputs the generated data to output devices via the input / output interface 1600.

[0097] The media interface 1700 reads a program or data stored in the recording medium 1800 and provides it to the CPU 1100 via the RAM 1200. The CPU 1100 loads the program from the recording medium 1800 onto the RAM 1200 via the media interface 1700 and executes the loaded program. The recording medium 1800 can be, for example, an optical recording medium such as a DVD (Digital Versatile Disc) or PD (Phase Change Rewritable Disk), a magneto-optical recording medium such as an MO (Magneto-Optical disk), tape media, magnetic recording medium, or semiconductor memory.

[0098] For example, when the computer 1000 functions as a learning device 100 according to the embodiment, the CPU 1100 of the computer 1000 realizes the functions of the control unit 130 by executing a program loaded on the RAM 1200. The CPU 1100 of the computer 1000 reads and executes these programs from the recording medium 1800, but as another example, these programs may be obtained from other devices via a predetermined communication network.

[0099] [8. Other] Furthermore, among the processes described in the above embodiments, all or part of the processes described as being performed automatically can be performed manually, or all or part of the processes described as being performed manually can be performed automatically by known methods. In addition, the processing procedures, specific names, and information including various data and parameters shown in the above document and drawings can be arbitrarily changed unless otherwise specified. For example, the various information shown in each figure is not limited to the information shown.

[0100] Furthermore, the components of each illustrated device are functionally conceptual and do not necessarily need to be physically configured as shown. In other words, the specific forms of distribution and integration of each device are not limited to those shown, and all or part of them can be functionally or physically distributed and integrated in any unit according to various loads and usage conditions.

[0101] Furthermore, the embodiments described above can be combined as appropriate, as long as the processing content is not contradictory.

[0102] Although some embodiments of the present invention have been described in detail above with reference to the drawings, these are illustrative examples, and the present invention can be implemented in various other forms with modifications and improvements based on the knowledge of those skilled in the art, starting with the embodiments described in the disclosure section of the invention.

[0103] Furthermore, the terms "section, module, unit" mentioned above can be replaced with "means" or "circuit." For example, the acquisition unit can be replaced with acquisition means or acquisition circuit. [Explanation of Symbols]

[0104] 1. Learning System 10. Carrier terminals 11 Communications Department 12 Input section 13 Output section 14 Control Unit 100 Learning Devices 110 Communications Department 120 Storage section 121 User information storage unit 122 Predictor memory unit 130 Control Unit 131 Acquisition Department 132 Learning Department 133 Prediction Section 134 Provision Department 141 Receiving Unit 142 Transmitter N Network

Claims

1. Regarding user behavior, the acquisition unit acquires a first data set in which labels are not missing and a second data set in which labels are missing. A learning unit that performs learning for supervised regression based on the first data set and learning for semi-supervised regression based on the second data set for a predictor that predicts behavior based on input data, A learning device characterized by having the following features.

2. The acquisition unit is, Regarding the actions, including actions related to conversions, the first data set and the second data set are acquired. The learning device according to feature 1.

3. The acquisition unit is, Obtain the first data set, which does not have missing labels indicating whether or not a user converted, and the second data set, which does have missing labels indicating whether or not a user converted. The learning device according to feature 1.

4. The acquisition unit is, The first data set, in which labels indicating user information are not missing, and the second data set, in which labels indicating user information are missing, are obtained. The learning device according to feature 1.

5. The acquisition unit is, Obtain the first set of data where labels are not missing because the user has allowed tracking, and the second set of data where labels are missing because the user has not allowed tracking. The learning device according to feature 1.

6. The aforementioned learning unit, The learning process described above is performed to carry out semi-supervised regression based on a second data set having the same number of samples as the first data set. The learning device according to feature 1.

7. The aforementioned learning unit, Perform the learning described above for semi-supervised regression based on the second data set so that the number of positive examples is the same as that of the first data set. The learning device according to feature 1.

8. The aforementioned learning unit, The learning process described above is performed to perform semi-supervised regression based on the second data set so that the distribution of the first data set and the distribution of positive examples are the same. The learning device according to feature 1.

9. A learning method performed by a computer, The acquisition process involves obtaining a first data set in which labels are not missing and a second data set in which labels are missing, regarding user behavior. A learning process for a predictor that predicts behavior based on input data, which includes learning for supervised regression based on the first data set and learning for semi-supervised regression based on the second data set, A learning method characterized by including the following.

10. Regarding user behavior, the acquisition procedure involves obtaining a first data set with complete labels and a second data set with missing labels. A learning procedure for a predictor that predicts behavior based on input data, comprising learning for supervised regression based on the first data set and learning for semi-supervised regression based on the second data set, A learning program characterized by having a computer execute it.

Images