Training device, training method, and training program

Abstract

In the present invention, a prediction device (1) includes: a training data input unit (11) that acquires, as a related data set, a labelled data set of a related task which is related to a target task; an important feature extraction unit (13) that acquires an important feature amount of the related task that contributes to the prediction of the related task; and a prediction model generator learning unit (14) that uses a prediction model generator (151) for generating a prediction model to evaluate the performance of a first prediction model generated from the important feature amount of the related task, with the labelled data of the related data set, and executes learning of the prediction model generator so that an evaluation value increases.

Description

Learning device, learning method, and learning program

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

[0002] Supervised machine learning technology requires training data to train a predictive model. For example, if you want to create a spam email predictor, you need training data for both non-spam and spam. Here, to make each piece of data computer-readable, it is represented by a pair of an n-dimensional vector (feature vector: each component of the feature vector is called a feature) that converts the email body into an n-dimensional vector, and a label (in this example, spam or non-spam).

[0003] However, in real-world problems, collecting training data can be difficult due to issues such as high annotation costs and data scarcity.

[0004] In order to perform highly accurate learning even in such situations, a method has been proposed for learning highly accurate predictive models even from a small amount of labeled data using a dataset of related tasks (Non-Patent Document 1).

[0005] In addition, a method has been proposed in which a predictive model is trained from auxiliary information about new predicted classes using a dataset from a related task (Non-Patent Document 2). In Non-Patent Document 2, auxiliary information (such as a class description) is provided for each class.

[0006] Chelsea Finn, Pieter Abbeel, and Sergey Levine. "Model-Agnostic Meta-Learning for Fast Adaptation of Deep Networks." International conference on machine learning. PMLR, 2017. Lampert, C. H.; Nickisch, H.; and Harmeling, S. 2013. Attribute-Based Classification for Zero-Shot Visual Object Categorization. IEEE transactions on pattern analysis and machine intelligence, 36(3): 453-465.

[0007] However, the method described in Non-Patent Document 1 requires even a small amount of data on the task to be solved, and therefore cannot be applied to problems for which the data on the task to be solved is not given.

[0008] Furthermore, the method described in Non-Patent Document 2 is a method specialized for classification tasks, and cannot be applied to general supervised learning problems such as regression tasks (tasks for predicting real values ​​from feature vectors), and therefore has a limited range of application.

[0009] The present invention has been made in consideration of the above, and aims to provide a learning device, a learning method, and a learning program that can accurately learn a predictive model even for a problem in which data for the task to be solved is not provided.

[0010] In order to solve the above-mentioned problems and achieve the object, the learning device of the present invention is characterized by having a first acquisition unit that acquires, as a related dataset, a labeled dataset of a related task that is related to a target task; a second acquisition unit that acquires important features of the related task that contribute to the prediction of the related task; and a learning unit that uses a predictive model generator to generate a predictive model, evaluates the performance of the first predictive model generated from the important features of the related task using labeled data in the related dataset, and performs training of the predictive model generator so as to increase the evaluation value.

[0011] According to the present invention, it is possible to accurately learn a prediction model even for a problem in which no data on the task to be solved is given.

[0012] Fig. 1 is a diagram illustrating processing by a prediction device according to an embodiment. Fig. 2 is a diagram schematically illustrating an example of the configuration of a prediction device according to an embodiment. Fig. 3 is a flowchart illustrating a processing procedure of a learning process. Fig. 4 is a flowchart illustrating a processing procedure of a prediction process. Fig. 5 is a diagram illustrating processing by a learning unit. Fig. 6 is a diagram illustrating an example of a computer that realizes the prediction device by executing a program.

[0013] An embodiment of the present invention will be described in detail below with reference to the drawings. Note that the present invention is not limited to this embodiment. In addition, in the drawings, identical parts are denoted by the same reference numerals. Note that, hereinafter, when "￣A" is written for A, which is a vector, matrix, or scalar, it is equivalent to "a symbol with "￣" written immediately above "A". When "~A" is written, it is equivalent to "a symbol with "~" written immediately above "A".

[0014] [Embodiment] In the embodiment, attention is paid to the fact that even if it is difficult to collect the learning data itself, it may be possible to obtain "prior knowledge" of the task to be solved.

[0015] In the embodiment, prior knowledge about a task, called "important features," is used. Important features refer to features that will contribute to the prediction of the task.

[0016] For example, consider a document classification task related to soccer and baseball. In this case, each word corresponds to a feature. Words such as "dribble," "hat trick," "home run," and "pitcher" are considered to be words related to soccer and baseball, and therefore could be important features for this document classification task.

[0017] In this way, it is often possible to identify such important features even if the data itself is not available.

[0018] For example, in a document classification task, it is easy to associate relevant important features (words) from category names (baseball, soccer, etc.). Also, when classifying new malware in security applications, even if it is difficult to collect the data on new malware itself, the characteristics of that malware can be obtained from security reports, etc. In medical applications, sharing patient data itself is difficult due to privacy concerns, but it may be possible to share a small number of important features (because it is difficult to reconstruct patient data from that information).

[0019] On the other hand, even if there is no data for the task you want to solve, you may be able to obtain training data for related tasks. In the example above, it may be possible to obtain documents from other categories, data on existing malware, or patient data from your own organization.

[0020] In the embodiment, we propose a method for accurately learning a predictive model from only a small amount of important feature information of a target task by utilizing a dataset of tasks related to the target task (related dataset).

[0021] [Overview of Learning Process] Fig. 1 is a diagram illustrating the processing of a prediction device according to an embodiment. The prediction device learns a highly accurate prediction model by utilizing information from labeled datasets of multiple related tasks in addition to information on a small number of important features of related tasks. Related tasks are tasks related to a target task.

[0022] As shown in Figure 1, in the learning phase, learning is performed using only data from datasets of multiple related tasks. The dataset for each related task contains data with accurate labels assigned to features (e.g., data D1 for task 1). Furthermore, it is assumed that important features for each task (e.g., important feature F1 for task 1) have been extracted in advance from the labeled data for the related tasks.

[0023] Then, in the learning phase, a prediction model is generated from the important features using a prediction model generator 151 for generating a prediction model ((1) in FIG. 1).

[0024] Next, in the training phase, the predictive model generator is trained to directly maximize the generalization performance (predictive performance for untrained data) of the generated predictive model. That is, the performance of the predictive model generated from a small number of important features of a related task is evaluated using labeled data of the related task, and the predictive model generator 151 for generating a predictive model is trained to increase the evaluation value ((2) in FIG. 1). Here, the generalization performance can be calculated directly from labeled data of the related task (e.g., data D1).

[0025] In the above learning, any feature selection method can be used to extract important features. If important features are known, they may be used. A specific method for generating a prediction model from important features will be described in the examples.

[0026] In the test phase, the trained predictive model generator is used to generate a predictive model suited to the target data by inputting the important features of the target task. In the test phase, the generated predictive model is used to make predictions on the test data.

[0027] [Prediction Device] Fig. 2 is a diagram schematically illustrating an example of the configuration of a prediction device according to an embodiment. The prediction device 1 according to the embodiment is realized, for example, by loading a predetermined program into a computer or the like including a ROM (Read Only Memory), a RAM (Random Access Memory), a CPU (Central Processing Unit), etc., and having the CPU execute the predetermined program. The prediction device 1 also has a communication interface for transmitting and receiving various information to and from other devices connected via a network, etc.

[0028] The prediction device 1 (learning device) is realized by a general-purpose computer such as a workstation, a personal computer, etc. As shown in Fig. 2, the prediction device 1 has a learning unit 10 that performs learning processing and a prediction unit 20 that performs prediction processing.

[0029] The learning unit 10 uses a plurality of related data sets to perform learning of a prediction model generator 151 (described later) for generating a prediction model.

[0030] When a small number of important features of a target task are given, the prediction unit 20 uses a prediction model generator 151 to generate a prediction model suitable for the target task. The prediction unit 20 uses the obtained prediction model to make predictions on test data. The prediction unit 20 may be implemented on the same hardware as the learning unit 10, or on different hardware.

[0031] [Learning Unit] The learning unit 10 has a learning data input unit 11 (first acquisition unit), a feature extraction unit 12, an important feature extraction unit 13 (second acquisition unit), a prediction model generator learning unit 14 (learning unit), and a storage unit 15.

[0032] The training data input unit 11 is implemented using input devices such as a keyboard and a mouse, and accepts input of various instruction information in response to input operations by an operator. The training data input unit 11 functions as an acquisition unit and receives labeled datasets of multiple related tasks (hereinafter referred to as related datasets) as input. As described above, the related datasets are data in which accurate labels are assigned to feature quantities. The related datasets may be input to the training unit 10 from an external server device or the like via a communication control unit (not shown) implemented by a network interface card (NIC) or the like.

[0033] The feature extraction unit 12 converts each data item in the input related dataset into a feature vector. The feature extraction unit 12 outputs the converted feature vector to the prediction model generator learning unit 14. Here, a feature vector is a representation of the features of the required data as an n-dimensional numeric vector. A method commonly used in machine learning is used for conversion into a feature vector. For example, if the data is text, methods such as morphological analysis, n-gram, and delimiter methods can be applied.

[0034] The important feature extraction unit 13 extracts important features for each task using labeled data of the task to be processed. The important feature extraction unit 13 extracts important features of the related task from the labeled data of the related task. Any feature selection method (e.g., L1 logistic regression, Lasso, etc.) can be used as a method for extracting important features of the related task. Note that the important feature extraction unit 13 may extract important features of the related task from the labeled data of the related task, or may acquire important features of the related task by receiving input of important features extracted in advance from the labeled data of the related task.

[0035] The prediction model generator training unit 14 trains the prediction model generator 151 to generate a prediction model. The prediction model generator 151 generates a first prediction model from a small number of important features of the related task. The method for generating this prediction model will be described later.

[0036] The prediction model generator training unit 14 then trains the prediction model generator 151 so as to directly maximize the generalization performance of the obtained first prediction model. The prediction model generator training unit 14 evaluates the performance of the first prediction model generated from a small number of important features of the related task using labeled data in the related dataset, and trains the prediction model generator 151 so as to increase the evaluation value.

[0037] In the test phase, when a small number of important features of the target task and test data are given, the predictive model generator 151 generates a second predictive model from the small number of important features of the target task. When test data of the target task is input, this second predictive model is a predictive model that makes predictions based on the input test data.

[0038] The storage unit 15 is realized by a semiconductor memory element such as a RAM or a flash memory, or a storage device such as a hard disk or an optical disk. The storage unit 15 stores a trained prediction model generator 151.

[0039] [Prediction Unit] The prediction unit 20 includes a data input unit 21, a feature extraction unit 22, a learning and prediction unit 23, and a prediction result output unit 24.

[0040] The data input unit 21 is realized using input devices such as a keyboard and a mouse, and accepts input of various instruction information in response to input operations by an operator. The data input unit 21 accepts input of a small number of important features of a target task and test data. The data input unit 21 outputs the small number of important features of the target task and the test data that it has accepted to the feature extraction unit 22. Once the prediction unit 20 has generated a prediction model from the important features, it can execute prediction by simply inputting the test data of the task.

[0041] The small number of important features of the target task and the test data may be input to the prediction unit 20 from an external server device or the like via a communication control unit (not shown) realized by a NIC or the like. The data input unit 21 may be the same hardware as the training data input unit 11.

[0042] Similar to the feature extraction unit 12 of the learning unit 10 , the feature extraction unit 22 converts the data to be predicted into a feature vector and outputs it to the learning / prediction unit 23 .

[0043] When a small number of important features of a target task are given, the learning and prediction unit 23 generates a second prediction model from the small number of important features of the target task using the prediction model generator 151 trained by the prediction model generator training unit 14. The obtained second prediction model is saved, and thereafter, when test data of the target task is input, the learning and prediction unit 23 uses the generated second prediction model to predict the test data of the target task.

[0044] The prediction result output unit 24 is realized by a display device such as a liquid crystal display, a printing device such as a printer, an information communication device, etc., and outputs the results of the prediction process to the operator.

[0045] 3 and 4, a processing procedure of the prediction device 1 will be described. The prediction processing of the prediction device 1 includes a learning process by the learning unit 10 and a prediction process by the prediction unit 20.

[0046] [Learning Process] Fig. 3 is a flowchart showing the processing procedure of the learning process, which is started, for example, when a user performs an operation input to instruct the start of the learning process.

[0047] 3, the training data input unit 11 receives a plurality of related data sets as input (step S1). The feature extraction unit 12 converts each data item in the received related data sets into a feature vector (step S2).

[0048] The important feature extraction unit 13 extracts important features of each task using the labeled data of each task (step S3).

[0049] The prediction model generator training unit 14 performs training of the prediction model generator 151 that generates a first prediction model from the small number of important features extracted in step S3 (step S4). The prediction model generator training unit 14 evaluates the performance of the first prediction model generated from the small number of important features using labeled data, and performs training of the prediction model generator 151 so as to increase the evaluation value.

[0050] The prediction model generator learning unit 14 stores the learned prediction model generator 151 in the storage unit 15 .

[0051] [Prediction Process] Fig. 4 is a flowchart showing the processing procedure of the prediction process. The flowchart in Fig. 4 starts, for example, when a user performs an operation input to instruct the start of the prediction process.

[0052] The data input unit 21 receives as input a small number of important features of the target task and test data (step S11). The feature extraction unit 22 converts each of the input data into a feature vector (step S12).

[0053] The learning / prediction unit 23 uses the prediction model generator 151 to generate a second prediction model from a small number of important features of the target task, and then uses the generated second prediction model to predict test data for the target task (step S13). From then on, the prediction unit 20 can output prediction results by inputting only the test data for the target task. The prediction result output unit 24 then outputs the prediction results based on the second prediction model (step S14).

[0054] [Effects of the embodiment] In this way, in the embodiment, by using related tasks to train a model so as to directly maximize the generalization performance of a predictive model generated from a small number of important features, it becomes possible to generate a highly accurate predictive model from a small number of important features of a target task.

[0055] In the embodiment, since the important features of a related task can be extracted from the labeled data of the related task, the method can be applied even when the important features are not given to the related task as domain knowledge.

[0056] Therefore, according to the embodiment, even if only a small number of important features are available for the task to be solved, highly accurate prediction can be performed.

[0057] Example 1 An example of a specific method for realizing the embodiment will be described.

[0058] First, we will explain the problem setting.

[0059] In this example, for simplicity of explanation, a binary classification task will be described as an example, but application to multi-class classification and regression tasks is also possible.

[0060] In the training phase, labeled data D (Equation (1)) for T source tasks are given.

[0061]

[0062] In formula (1), x tn is the D-dimensional feature vector of the nth data of the tth task. tn ∈{c t0 , c t1} is a label. t ={z t0 , z t1} is a vector representing the important features of the t-th task. tk ∈{0, 1} D is a D-dimensional vector representing the important features for the kth class of task t, and the dth component z tkd When is 1, the dth feature is important to class k, and when is 0, it is unknown.

[0063] Specifically, in the case of a text classification task between soccer (class = 0) and baseball (class = ), the words "dribble" and "hat trick" are related to soccer, so the component z corresponding to those words is 0d is 1, and the word "home run" is related to baseball, so the component z 0d´ becomes 1. Note that the subscript t is omitted in this explanation.

[0064] This z tcan be constructed by using a feature selection method such as logistic regression with L1 regularization on the labeled data of the t-th task. For simplicity, in this example, we assume that the same feature is not important in both classes at the same time. That is, z t0 ・z t1 = 0. In the prediction phase, only the important feature vector z = (z0, z1) of the target task is given.

[0065] Here, we assume that in real problems, it is rare to have a large amount of prior knowledge, so the number of important features given is small. In other words, we assume that the number of components in z that are 1 is small. The goal is to learn a highly accurate binary classifier from z.

[0066] First, we explain the predictive model generator for generating a classifier from z (test phase), and then explain how to train the generator (training phase).

[0067] As a model of the classifier, consider the probability model of equation (2).

[0068]

[0069] In equation (2), σ is a sigmoid function. h is a neural network with parameter Θ. w is a linear weight. In this model, if the content of σ() is 0 or greater, x is classified as y=1, otherwise y=0. w is a task-specific parameter, and the objective of the test phase is to find w from z. Θ is a task-wide parameter, and the objective of the training phase is to learn Θ using a large amount of related data.

[0070] [Procedure for Calculating w from z] Next, a specific procedure for calculating w from z will be described. In this Example 1, the following is assumed.

[0071] Important features of class k (z kd =1) increases, the model's output / probability of class k also increases. kd = 1) decreases, the model's output / probability of class k also decreases.

[0072] For example, in document classification, the more frequently the word "dribble" appears in a document, the higher the probability that the document belongs to the "soccer" class. In the medical field, the higher the blood pressure (feature) value, the higher the probability of myocardial infarction (class). Since it is believed that prior knowledge that humans can find is often simple, the above simple assumptions were used in this example. It should be noted that there may be important features that decrease the probability of class 0 when the important feature d for class 0 increases. In this case, note that the feature can be rephrased as important for class 1. This is because a decrease in the probability of class 0 means an increase in the probability of class 1.

[0073] Based on this assumption, in the first embodiment, w is calculated to minimize the objective function of equation (3).

[0074]

[0075] where ~x n is artificial data generated from a probability distribution common to all tasks (the distribution can be a Gaussian distribution or a uniform distribution, etc.). n (d) is ~x n The dth component of nd is the vector obtained by adding I n inc and I n dec is defined by equation (4).

[0076]

[0077] In other words, I n inc For the feature d contained in n (d) When you input the original vector ~x n Compared to , the desired feature is that the output wh(x) of class 1 increases. n dec The feature d is the transformed vector ~x n (d) When you enter the original vector ~x nCompared to the set, we want the output wh(x) of class 1 to decrease (the output of class 0 to increase).

[0078] These requirements are met by minimizing this loss (Equation (3)). Therefore, by minimizing this loss, the information of the important feature z can be reflected in w to train the model. Minimization can be performed using the gradient method.

[0079] [Learning Procedure] Here, a method for learning the parameter Θ using a related data set will be described. The objective function is given by equation (5).

[0080]

[0081] In equation (5), Q is the labeled data for evaluation extracted from the t-th task, and S is the artificial data generated from the distribution common to all tasks (using the same symbols as S in the test phase). ̂z is the important feature z of task t. t represents a vector sampled from * represents the class-specific weights of the classifier trained with ̂z and S. The training method is the same as in the test phase.

[0082] The objective function in equation (5) is the trained classifier σ(w * This represents the negative log-likelihood when evaluating h(x) using labeled data Q. Since this classification (Equation (5)) is differentiable with respect to the parameter Θ, this optimization can be performed using standard stochastic gradient descent. By performing this training using multiple related data, we can expect to obtain a neural network h with high generalization performance.

[0083] [Pseudo Code of Learning Method] Fig. 5 is a diagram for explaining the processing of the learning unit 10. Fig. 5 illustrates pseudo code of the processing of the learning unit 10.

[0084] First, the learning unit 10 receives labeled data D of a source task, artificially generated unlabeled data S (size N), and labeled data (query set) Q (size N Q ), the number of gradient steps I, the step size α, and the important feature vector z of the relevant task. t(relevant features wp) (J pieces), the task is to obtain a predefined distribution p of artificial data and obtain the parameters θ of the neural network.

[0085] The learning unit 10 randomly samples a task t from {1, ..., T} and calculates z t Randomly extract important features ~z for the relevant task from D. Then, randomly sample artificially generated unlabeled data S of size N from a predefined distribution p. t From size N Q Randomly extract labeled data Q (lines 2-5 of Algorithm 1).

[0086] The learning unit 10 initializes the linear weight w to 0, and then calculates w from z by updating the task-specific linear weight w for 1 to I using equation (3) as described above (lines 6-9 of Algorithm 1).

[0087] The learning unit 10 calculates the loss of w using equation (5) (line 10 of Algorithm 1), i.e., calculates the objective function.

[0088] Then, the learning unit 10 updates the parameter θ using the gradient of the loss (line 11 of Algorithm 1). That is, the model is updated based on the calculation results.

[0089] Example 2 A method of applying the specific example of Example 1 to a regression task will be described as Example 2. First, Equation (6) is used as a model equivalent to Equation (2).

[0090]

[0091] As in the binary classification case, in equation (6), w is a task-specific parameter and h is a task-common neural network.

[0092] The optimization problem in the test phase uses the same loss (Equation (3)) as for binary classification.

[0093] In the optimization problem of the learning phase, the negative log-likelihood was used as the objective function for binary classification, but mean squared error or the like is used for regression tasks.

[0094] Third Embodiment In a third embodiment, an extension of the specific example of the first embodiment will be described in which unlabeled data can be used in addition to the important features of the target task.

[0095] When unlabeled data is available, the objective function is Equation (7) instead of Equation (3).

[0096]

[0097] Here, the second term of Equation (7) is based on the objective function (Equation (3)) defined for the artificial data in Example 1, and the first term of Equation (7) is the loss for unlabeled data. Here, ~y represents a label artificially assigned to unlabeled data x using important feature information. The label assignment method is to assign an important feature vector z for each class to t0 , z t1 One possible method is to assign a label to the class that is closer to x.

[0098] According to this embodiment, by utilizing a dataset of tasks related to the target task (related dataset), a predictive model can be accurately learned from only a small amount of important feature information, so that a predictive model can be accurately learned even for a problem for which no data on the task to be solved is provided. Furthermore, according to this embodiment, as described in Examples 1 to 3, it can also be applied to general supervised learning problems, and represents an improvement in the field of prediction technology.

[0099] [System Configuration of the Embodiment] Each component of the prediction device 1 is a functional concept and does not necessarily have to be physically configured as shown in the figure. In other words, the specific form of distribution and integration of the functions of the prediction device 1 is not limited to that shown in the figure, and all or part of the functions can be functionally or physically distributed or integrated in any unit depending on various loads, usage conditions, etc.

[0100] Furthermore, all or any part of the processes performed in the prediction device 1 may be realized by a CPU (Central Processing Unit), a GPU (Graphics Processing Unit), and a program analyzed and executed by the CPU and the GPU. Furthermore, each process performed in the prediction device 1 may be realized as hardware using wired logic.

[0101] Furthermore, among the processes described in the embodiments, all or part of the processes described as being performed automatically can be performed manually. Alternatively, all or part of the processes described as being performed manually can be performed automatically using a known method. In addition, the processing procedures, control procedures, specific names, and information including various data and parameters described above and illustrated can be changed as appropriate unless otherwise specified.

[0102] 6 is a diagram showing an example of a computer in which a program is executed to realize the prediction device 1. The computer 1000 includes, for example, a memory 1010 and a CPU 1020. The computer 1000 also includes a hard disk drive interface 1030, a disk drive interface 1040, a serial port interface 1050, a video adapter 1060F, and a network interface 1070. These components are connected by a bus 1080.

[0103] The memory 1010 includes a ROM (Read Only Memory) 1011 and a RAM (Random Access Memory) 1012. The ROM 1011 stores, for example, a boot program such as a BIOS (Basic Input Output System). The hard disk drive interface 1030 is connected to a hard disk drive 1090. The disk drive interface 1040 is connected to a disk drive 1100. A removable storage medium such as a magnetic disk or optical disk is inserted into the disk drive 1100. The serial port interface 1050 is connected to, for example, a mouse 1110 and a keyboard 1120. The video adapter 1060 is connected to, for example, a display 1130.

[0104] The hard disk drive 1090 stores, for example, an OS (Operating System) 1091, an application program 1092, a program module 1093, and program data 1094. That is, a program that defines each process of the prediction device 1 is implemented as a program module 1093 in which code executable by the computer 1000 is written. The program module 1093 is stored, for example, in the hard disk drive 1090. For example, a program module 1093 for executing processes similar to those of the functional configuration of the prediction device 1 is stored in the hard disk drive 1090. Note that the hard disk drive 1090 may be replaced by an SSD (Solid State Drive).

[0105] Furthermore, setting data used in the processing of the above-described embodiment is stored as program data 1094, for example, in memory 1010 or hard disk drive 1090. Then, CPU 1020 reads out program module 1093 or program data 1094 stored in memory 1010 or hard disk drive 1090 into RAM 1012 as necessary and executes them.

[0106] The program module 1093 and program data 1094 may not necessarily be stored in the hard disk drive 1090, but may also be stored in, for example, a removable storage medium and read by the CPU 1020 via the disk drive 1100 or the like. Alternatively, the program module 1093 and program data 1094 may be stored in another computer connected via a network (such as a local area network (LAN) or a wide area network (WAN)). The program module 1093 and program data 1094 may then be read by the CPU 1020 from the other computer via the network interface 1070.

[0107] Although the present invention has been described above as an embodiment, the present invention is not limited to the descriptions and drawings that form part of the disclosure of the present invention. In other words, other embodiments, examples, and operational techniques that can be made by those skilled in the art based on the present invention are all included in the scope of the present invention.

[0108] REFERENCE SIGNS LIST 1 Prediction device 10 Learning unit 11 Learning data input unit 12, 22 Feature extraction unit 13 Important feature extraction unit 14 Prediction model generator learning unit 15 Storage unit 20 Prediction unit 21 Data input unit 23 Learning and prediction unit 24 Prediction result output unit 151 Prediction model generator

Claims

1. A learning device comprising: a first acquisition unit that acquires, as a related dataset, a labeled dataset of a related task that is related to a target task; a second acquisition unit that acquires important features of the related task that contribute to the prediction of the related task; and a learning unit that uses a predictive model generator to generate a predictive model, evaluates the performance of the first predictive model generated from the important features of the related task using labeled data from the related dataset, and performs training of the predictive model generator so as to increase the evaluation value.

2. The learning device according to claim 1, wherein the second acquisition unit extracts the important features of the related task from labeled data of the related task.

3. The learning device described in claim 1, further comprising a prediction unit that, when a small number of important features of a target task are given, generates a second prediction model corresponding to the target task using the prediction model generator trained by the learning unit, and predicts test data for the target task using the generated second prediction model.

4. A learning method executed by a learning device, comprising: a step of acquiring, as a related dataset, a labeled dataset of a related task related to a target task; a step of acquiring important features of the related task that contribute to the prediction of the related task; and a step of using a predictive model generator to generate a predictive model, evaluating the performance of a first predictive model generated from the important features of the related task using labeled data from the related dataset, and executing learning of the predictive model generator so as to increase the evaluation value.

5. A learning program for causing a computer to execute the steps of: acquiring, as a related dataset, a labeled dataset of a related task that is related to a target task; acquiring important features of the related task that contribute to the prediction of the related task; and using a predictive model generator to generate a predictive model, evaluating the performance of a first predictive model generated from the important features of the related task using labeled data from the related dataset, and performing training of the predictive model generator so as to increase the evaluation value.

Images