Computer-readable recording medium, machine learning method, and information processing apparatus

Abstract

To enable learning utilizing evidence information.SOLUTION: In learning the machine learning model 102, the information processing apparatus 101 acquires first feature information 112 representing features of the first data 110 to which the correct answer 111 is given. The information processing apparatus 101 calculates the information amount of the acquired first feature information 112. In a case where the calculated information amount of the first feature information 112 is equal to or less than the threshold value, the information processing apparatus 101 creates the learning data 130 in which the second feature information 113 obtained by adding the first evidence information 120 to the first feature information 112 and the correct answer 111 are associated with each other. On the other hand, in a case where the information amount of the calculated first feature information 112 is larger than the threshold value, the information processing apparatus 101 creates the learning data 140 in which the first feature information 112 and the correct answer 111 are associated with each other without adding the first evidence information 120.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] The present invention relates to a machine learning program, a machine learning method, and an information processing device. [Background technology]

[0002] In fields where open data is developing, human curation is common, and information that can be used as evidence can be obtained. In machine learning, evidence information can be used as a feature that is almost equivalent to the correct answer. For example, when learning about the pathogenicity of fusion genes, the number of papers claiming that a certain fusion gene is pathogenic can be used as evidence information.

[0003] Prior art techniques include generating a teacher pattern for generating training cases to be used in learning using a first training case for learning a model and clue information indicating the basis for using the first training case for learning, generating a second training case using the teacher pattern, and outputting the second training case and the clue information used to generate the teacher pattern that generated the second training case. [Prior art documents] [Patent documents]

[0004] [Patent Document 1] Japanese Patent Application Publication No. 2018-81569 Summary of the Invention [Problem to be solved by the invention]

[0005] However, in conventional technology, when training a machine learning model, if evidence information that can be almost equivalent to the correct answer is included in the features during training, there is a problem that this leads to a decrease in the estimation accuracy of the machine learning model.

[0006] In one aspect, the present invention aims to enable learning that utilizes evidence information. [Means for solving the problem]

[0007] In one embodiment, a machine learning program is provided that, when training a machine learning model that uses feature information representing the characteristics of data as input and outputs an inference result for the data, calculates the amount of information of first feature information representing the characteristics of first data to which a correct answer has been assigned, and if the calculated amount of information is equal to or less than a threshold, creates training data that matches the correct answer with second feature information in which first evidence information that can be the basis for the correct answer obtained for the first data has been added to the first feature information, and if the amount of information is greater than the threshold, creates training data that matches the first feature information with the correct answer. [Effects of the Invention]

[0008] Advantageous Effects of Invention According to one aspect of the present invention, it is possible to achieve an effect of enabling learning that utilizes evidence information. [Brief explanation of the drawings]

[0009] [Figure 1] FIG. 1 is an explanatory diagram illustrating an example of a machine learning method according to an embodiment. [Figure 2] FIG. 2 is an explanatory diagram showing an example of the system configuration of the information processing system 200. As shown in FIG. [Figure 3] FIG. 3 is a block diagram showing an example of the hardware configuration of the learning device 201. [Figure 4] FIG. 4 is an explanatory diagram showing an example of the contents stored in the knowledge database 220. As shown in FIG. [Figure 5] FIG. 5 is a block diagram showing an example of the functional configuration of the learning device 201. [Figure 6] FIG. 6 is an explanatory diagram (part 1) showing an example of the operation of the learning model M during learning. [Figure 7] FIG. 7 is an explanatory diagram (part 2) showing an example of the operation of the learning model M during learning. [Figure 8] FIG. 8 is an explanatory diagram (part 3) showing an example of the operation of the learning model M during learning. [Figure 9] FIG. 9 is an explanatory diagram (part 4) showing an example of the operation of the learning model M during learning. [Figure 10] FIG. 10 is an explanatory diagram (part 5) showing an example of the operation of the learning model M during learning. [Figure 11] FIG. 11 is an explanatory diagram (part 6) showing an example of the operation of the learning model M during learning. [Figure 12] FIG. 12 is an explanatory diagram (part 7) showing an example of the operation of the learning model M during learning. [Figure 13] FIG. 13 is an explanatory diagram (part 8) showing an example of the operation of the learning model M during learning. [Figure 14] FIG. 14 is an explanatory diagram (part 9) showing an example of the operation of the learning model M during learning. [Figure 15] FIG. 15 is an explanatory diagram (part 10) showing an example of the operation of the learning model M during learning. [Figure 16] FIG. 16 is an explanatory diagram (part 1) showing an example of operation during estimation using the learning model M. [Figure 17] FIG. 17 is an explanatory diagram (part 2) showing an example of operation during estimation using the learning model M. [Figure 18] FIG. 18 is a flowchart showing an example of the learning process procedure of the learning device 201. [Figure 19] FIG. 19 is a flowchart showing an example of the estimation processing procedure of the learning device 201. DETAILED DESCRIPTION OF THE INVENTION

[0010] DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS Hereinafter, with reference to the accompanying drawings, embodiments of a machine learning program, a machine learning method, and an information processing device according to the present invention will be described in detail.

[0011] (Embodiment) Fig. 1 is an explanatory diagram showing an example of a machine learning method according to an embodiment. In Fig. 1, an information processing device 101 is a computer that creates training data used for training a machine learning model 102. The machine learning model 102 receives feature information of data as input and outputs an estimation result for the data.

[0012] The data feature information is information that represents the characteristics of the data, and corresponds to, for example, the feature quantity of the data. The data feature information is expressed, for example, by a tree-structured graph. The machine learning model 102 can be used, for example, to predict whether a certain fusion gene is pathogenic, whether a certain chemical substance is harmful, or the name of a disease in a patient with certain symptoms.

[0013] In fields where open activities are active, evidence information can be used through curation. Curation is the process of collecting, selecting, and editing data according to a specific perspective or theme, adding new value to the information and making it easier to use. Evidence information is information that can serve as the basis for deriving some kind of result.

[0014] In machine learning, evidence information is useful as knowledge and can be used as a feature that is almost equivalent to the correct answer, but there is a risk of data leakage. For example, if evidence information is uniformly included in features during learning, the weight of the feature becomes too large, and it may not be possible to properly estimate a solution for data that does not have evidence.

[0015] Therefore, in this embodiment, a machine learning method that enables learning using evidence information will be described. Here, an example of processing by the information processing device 101 will be described. When training the machine learning model 102, the information processing device 101 performs, for example, the following processes (1) to (5).

[0016] (1) The information processing device 101 acquires first feature information 112 representing the features of first data 110 to which a correct answer 111 has been assigned. Here, the first data 110 is information for which the answer to the problem is known. The correct answer 111 is information indicating the correct answer of the first data 110. The first feature information 112 is expressed, for example, by a graph (directed graph or undirected graph) in which each piece of information representing the feature of the first data 110 is a node.

[0017] Specifically, for example, the information processing device 101 may refer to a knowledge database (for example, a knowledge database 220 shown in FIG. 2, which will be described later) to acquire the first feature information 112. Alternatively, the information processing device 101 may acquire the first feature information 112 through an operation input by a user.

[0018] (2) The information processing device 101 calculates the amount of information of the acquired first feature information 112. Here, the amount of information of the first feature information 112 indicates the amount of information. For example, the amount of information of the first feature information 112 increases as the first feature information 112 represents more information, and decreases as the first feature information 112 represents less information.

[0019] For example, it is assumed that the first feature information 112 is expressed by a graph in which each piece of information representing the feature of the first data 110 is a node. In this case, the information processing device 101 may calculate the number of nodes in the graph representing the feature of the first data 110 as the amount of information of the first feature information 112.

[0020] (3) The information processing device 101 determines whether the calculated amount of information in the first feature information 112 is equal to or less than a threshold. Here, the threshold can be set arbitrarily. For example, the threshold may be set from the perspective of evaluating the amount of information and determining whether effective learning can be performed using only the first feature information 112.

[0021] (4) When the amount of information in the first feature information 112 is below a threshold, the information processing device 101 creates learning data 130 that associates second feature information 113, in which the first evidence information 120 is added to the first feature information 112, with the correct answer 111.

[0022] Here, the first evidence information 120 is information that can serve as a basis for the correct answer 111 for the first data 110. For example, the first evidence information 120 may indicate the number of documents that indicate that the first data 110 is the correct answer 111. For the correct answer 111 for which the amount of information in the first feature information 112 can be said to be small, the information processing device 101 creates the training data 130 by including the first evidence information 120 in the first feature information 112.

[0023] (5) When the amount of information in the first feature information 112 is greater than a threshold, the information processing device 101 creates learning data 140 that associates the first feature information 112 with the correct answer 111. For the correct answer 111 in which the amount of information in the first feature information 112 can be said to be large, the information processing device 101 creates learning data 140 without including the first evidence information 120 in the first feature information 112.

[0024] In this way, the information processing device 101 can enable learning that utilizes evidence information while preventing a decrease in the estimation accuracy of the machine learning model 102. For example, when the amount of information in the first feature information 112 is small, the information processing device 101 can enable learning that takes into account the first evidence information 120, by using second feature information 113 in which the first evidence information 120 is added to the first feature information 112. Furthermore, when the amount of information in the first feature information 112 is large, the information processing device 101 can enable learning that makes use of the first feature information 112 as is, without adding the first evidence information 120.

[0025] (System configuration example of information processing system 200) Next, a system configuration example of an information processing system 200 including the information processing device 101 shown in Fig. 1 will be described. Here, the case where the information processing device 101 shown in Fig. 1 is applied to a learning device 201 in the information processing system 200 will be described as an example.

[0026] Fig. 2 is an explanatory diagram showing an example of the system configuration of an information processing system 200. In Fig. 2, the information processing system 200 includes a learning device 201, a management terminal 202, and a user terminal 203. In the information processing system 200, the learning device 201, the management terminal 202, and the user terminal 203 are connected via a wired or wireless network 210. The network 210 is, for example, the Internet, a LAN (Local Area Network), or a WAN (Wide Area Network).

[0027] Here, the learning device 201 is a computer that has a knowledge database 220 and creates learning data used to learn a learning model M (see FIG. 5 described later). The learning device 201 may also use the created learning data to learn the learning model M. The machine learning model 102 shown in FIG. 1 corresponds to the learning model M, for example.

[0028] Furthermore, the learning device 201 may output an estimation result for estimation target data using the trained learning model M. The estimation target data is received from, for example, a user terminal 203. The learning device 201 is, for example, a server. The contents stored in the knowledge database 220 will be described later with reference to FIG. 4.

[0029] The management terminal 202 is a computer used by an administrator of the information processing system 200. The user terminal 203 is a computer used by a user of the information processing system 200. The management terminal 202 and the user terminal 203 are, for example, a PC (Personal Computer), a tablet PC, a smartphone, or the like.

[0030] Here, the learning device 201 is provided separately from the management terminal 202 and the user terminal 203, but this is not limiting. For example, the learning device 201 may be realized by the management terminal 202 or the user terminal 203. Furthermore, the information processing system 200 may include multiple management terminals 202 and user terminals 203.

[0031] (Example of hardware configuration of learning device 201) Next, an example of the hardware configuration of the learning device 201 will be described.

[0032] Fig. 3 is a block diagram showing an example of the hardware configuration of a learning device 201. In Fig. 3, the learning device 201 has a CPU (Central Processing Unit) 301, a memory 302, a disk drive 303, a disk 304, a communication I / F (Interface) 305, a portable recording medium I / F 306, and a portable recording medium 307. Furthermore, each component is connected to each other by a bus 300.

[0033] Here, CPU 301 is responsible for overall control of learning device 201. CPU 301 may have multiple cores. Memory 302 includes, for example, a read-only memory (ROM) and a random access memory (RAM). A program stored in memory 302 is loaded into CPU 301, causing CPU 301 to execute coded processing.

[0034] The disk drive 303 controls reading / writing of data from / to the disk 304 under the control of the CPU 301. The disk 304 stores data written under the control of the disk drive 303. The disk 304 is, for example, a magnetic disk, an optical disk, or the like.

[0035] The communication I / F 305 is connected to the network 210 via a communication line, and is connected to an external computer (for example, the management terminal 202 and the user terminal 203 shown in FIG. 2) via the network 210. The communication I / F 305 serves as an interface between the network 210 and the inside of the device, and controls the input and output of data from the external computer. The communication I / F 305 is, for example, a modem or a LAN adapter.

[0036] The portable recording medium I / F 306 controls reading and writing of data from and to the portable recording medium 307 under the control of the CPU 301. The portable recording medium 307 stores data written under the control of the portable recording medium I / F 306. The portable recording medium 307 is, for example, a CD (Compact Disc)-ROM, a DVD (Digital Versatile Disk), or a USB (Universal Serial Bus) memory.

[0037] In addition to the above-described components, the learning device 201 may also have, for example, a graphics processing unit (GPU), a neural processing unit (NPU), an input device, a display, etc. Furthermore, the learning device 201 may not have, for example, the portable recording medium I / F 306 and the portable recording medium 307 among the above-described components. Furthermore, the management terminal 202 and the user terminal 203 shown in FIG. 2 can also be realized with the same hardware configuration as the learning device 201. However, the management terminal 202 and the user terminal 203 have, for example, an input device, a display, etc. in addition to the above-described components.

[0038] (Storage contents of knowledge database 220) Next, the contents stored in knowledge database 220 of learning device 201 will be described with reference to Fig. 4. Knowledge database 220 is realized by storage devices such as memory 302 and disk 304, for example.

[0039] Fig. 4 is an explanatory diagram showing an example of the contents stored in the knowledge database 220. In Fig. 4, the knowledge database 220 stores a knowledge graph (for example, knowledge graph 400). A knowledge graph is a graph structure in which various pieces of knowledge such as open data are systematically linked together.

[0040] For example, knowledge graph 400 includes nodes 401 to 415 and edges (e.g., edges 421 to 425) connecting the nodes. Each of nodes 401 to 415 represents knowledge. Each edge represents a relationship between nodes (between knowledge). An attribute is set for each edge. The attribute indicates the relationship between nodes (between knowledge), and indicates, for example, what type of information a downstream node has relative to an upstream node.

[0041] Here, node 401 represents a fusion gene. A fusion gene is a gene formed by fusing two different genes. Fusion genes can cause cancer. A fusion gene is represented, for example, by separating the names of the two genes that make up the fusion gene with a double colon.

[0042] Nodes 402 and 403 represent genes. An attribute "5'gene" is set to edge 421 connecting node 401 and node 402. The attribute "5'gene" indicates that it is the leading gene that forms the fusion gene. An attribute "3'gene" is set to edge 422 connecting node 401 and node 403. The attribute "3'gene" indicates that it is the trailing gene that forms the fusion gene.

[0043] Nodes 404 to 406 represent related literature. The edge (for example, edge 423) connecting node 401 with nodes 404 to 406 has the attribute "evidence" set. The attribute "evidence" indicates information that can serve as evidence. For example, the literature "Paper A" represented by node 404 describes that the fusion gene "BCR::ABL1" represented by node 401 is pathogenic.

[0044] Nodes 407 to 415 represent domains. The attribute "domain" is set to edges (for example, edge 424) connecting node 402 with nodes 407 to 410. The attribute "domain" is also set to edges (for example, edge 425) connecting node 403 with nodes 411 to 415. The attribute "domain" indicates the domain of a gene.

[0045] (Example of functional configuration of learning device 201) Next, an example of the functional configuration of the learning device 201 will be described.

[0046] 5 is a block diagram showing an example of the functional configuration of the learning device 201. In FIG. 5, the learning device 201 includes an acquisition unit 501, a calculation unit 502, a determination unit 503, a creation unit 504, a learning unit 505, a reception unit 506, an estimation unit 507, an output unit 508, and a storage unit 510. The acquisition unit 501 to the output unit 508 are functions that constitute the control unit 500, and specifically, the functions are realized by causing the CPU 301 to execute a program stored in a storage device such as the memory 302, the disk 304, or the portable recording medium 307 shown in FIG. 3, or by the communication I / F 305. The processing results of each functional unit are stored in a storage device such as the memory 302 or the disk 304.

[0047] The storage unit 510 is realized by a storage device such as the memory 302 or the disk 304. Specifically, the storage unit 510 stores, for example, the knowledge database 220 and the learning model M shown in Fig. 4. The learning model M is a machine learning model that receives the feature quantities of data as input and outputs an estimation result for the data.

[0048] The data may be, for example, gene information indicating a fusion gene. In this case, the inference result for the data may indicate, for example, whether or not the fusion gene is pathogenic. The data may also be substance information indicating a chemical substance. In this case, the inference result for the data may indicate, for example, whether or not the chemical substance is harmful. The data may also be patient information indicating a patient with a certain symptom. In this case, the inference result for the data may indicate, for example, the name of the patient's disease.

[0049] Note that, although the following description takes as an example a case where the storage unit 510 is included in the learning device 201, the present invention is not limited to this. For example, the storage unit 510 may be included in a computer other than the learning device 201. In this case, the learning device 201 can access the other computer to refer to the contents stored in the storage unit 510.

[0050] <When learning model M> First, the functional units used when learning the learning model M will be described.

[0051] The acquiring unit 501 acquires correct answer data. Here, the correct answer data is data (first data) to which a correct answer label has been assigned. The correct answer label is information indicating the correct answer to a question given about the data. For example, the correct answer data may include a correct answer label and genetic information. The genetic information corresponds to data to which a correct answer label has been assigned. In this case, if the question given about the genetic information is "pathogenicity of the fusion gene," the correct answer label indicates whether or not the fusion gene indicated by the genetic information is pathogenic.

[0052] The correct answer data may also include a correct answer label and substance information. The substance information corresponds to data to which a correct answer label has been assigned. In this case, if the question given for the substance information is "harmfulness of chemical substance," the correct answer label indicates whether or not the chemical substance indicated by the substance information is harmful. The correct answer data may also include a correct answer label and patient information. The patient information corresponds to data to which a correct answer label has been assigned. In this case, if the question given for the patient information is "name of disease," the correct answer label indicates the name of the disease of the patient indicated by the patient information.

[0053] Specifically, for example, the acquiring unit 501 may acquire the correct answer data by receiving it from the management terminal 202 shown in Fig. 2. The acquiring unit 501 may also acquire the correct answer data by an operation input by a user (for example, an administrator of the information processing system 200) using an input device (not shown) of the acquiring unit 501.

[0054] In the following description, "genetic information indicating a fusion gene" may be used as an example of data to which a correct answer label is assigned. The correct answer 111 shown in FIG. 1 corresponds to the correct answer label. The first data 110 to which the correct answer 111 shown in FIG. 1 is assigned corresponds to the correct answer data.

[0055] The acquiring unit 501 acquires the feature amount of the supervised data. Here, the feature amount of the supervised data is feature information (first feature information) that represents the feature of the data to which the supervised label is assigned. What information is acquired as the feature amount of the supervised data is predetermined depending on, for example, what the supervised data indicates.

[0056] For example, in the case of genetic information, the feature of the acquired correct answer data includes domain information of each gene that forms the fusion gene indicated by the genetic information. However, the feature of the acquired correct answer data does not include evidence information. Furthermore, it is preferable that the feature of the acquired correct answer data does not include information that is substantially equivalent to the correct answer, other than evidence information. The first feature information 112 shown in FIG. 1 corresponds to the feature of the correct answer data.

[0057] Specifically, for example, the acquiring unit 501 may acquire the feature quantities of the supervised data by referring to the storage unit 510 that stores a knowledge graph. For example, the knowledge graph includes nodes that represent each of a plurality of pieces of knowledge and edges that represent the relationships between the pieces of knowledge included in the plurality of pieces of knowledge. The feature quantities of the supervised data are represented, for example, by a tree-structured graph (directed graph or undirected graph).

[0058] More specifically, for example, the acquiring unit 501 refers to the genetic information included in the correct answer data to identify two genes that form the fusion gene indicated by the genetic information. Next, the acquiring unit 501 refers to the knowledge database 220 shown in Fig. 4 to extract domain information for each of the identified genes. The domain information is information about the genes that form the fusion gene, and corresponds to information that represents the characteristics of the fusion gene.

[0059] The acquiring unit 501 then generates a feature graph based on the extracted domain information. The feature graph corresponds to the feature amounts of the supervised data, and is, for example, a tree-structured graph with the extracted domain information as nodes. An example of generating a feature graph representing the feature amounts of the supervised data will be described later with reference to, for example, Figs. 6 and 9.

[0060] The acquiring unit 501 may acquire the feature amount (for example, a feature graph) of the correct answer data by receiving it from the management terminal 202. The acquiring unit 501 may also acquire the feature amount (for example, a feature graph) of the correct answer data by a user's operation input using an input device (not shown) of the acquiring unit 501.

[0061] The acquiring unit 501 acquires evidence information of the correct answer data. Here, the evidence information of the correct answer data is information (first evidence information) that can be a basis for the correct answer of the correct answer data. What information is acquired as the evidence information of the correct answer data is predetermined depending on, for example, what the correct answer data indicates.

[0062] For example, in the case of genetic information, the evidence information of the acquired correct answer data may include information indicating the number of literature on the pathogenicity of the fusion gene indicated by the genetic information. The literature on pathogenicity is, for example, a paper or article that asserts that the fusion gene indicated by the genetic information is the correct answer indicated by the correct answer label ("pathogenic" or "non-pathogenic").

[0063] Furthermore, evidence information may not necessarily be literature, but may be the number of pieces of information in which researchers or others have declared (labeled) a fusion gene indicated by genetic information as "pathogenic" or "non-pathogenic." The first evidence information 120 shown in FIG. 1 corresponds to the evidence information of the correct answer data. Specifically, for example, the acquiring unit 501 may acquire the evidence information of the correct answer data by referring to the storage unit 510.

[0064] More specifically, for example, the acquiring unit 501 refers to the knowledge database 220 to identify a node representing a fusion gene indicated by the genetic information. Next, the acquiring unit 501 refers to the knowledge database 220 to count the number of nodes representing related literature that are connected to the identified node by edges of the attribute "evidence." The acquiring unit 501 may then acquire the counted number of nodes as evidence information. The related literature is, for example, literature on the pathogenicity of the fusion gene represented by the identified node. An example of acquiring evidence information for supervised data will be described later with reference to, for example, FIG. 10.

[0065] The calculation unit 502 calculates the information amount of the feature amount of the acquired correct answer data. The information amount of the feature amount of the correct answer data may be expressed, for example, by the number of pieces of information (such as domain information) that represent the features of the correct answer data. Specifically, for example, the calculation unit 502 may calculate the information amount of the feature amount of the correct answer data based on the number of nodes in the generated feature graph. To explain in more detail, for example, the calculation unit 502 calculates the number of nodes that represent domain information in the feature graph as the information amount of the feature amount of the correct answer data.

[0066] The determination unit 503 determines whether the calculated amount of information is equal to or less than a threshold value α. Here, the threshold value α can be set arbitrarily. For example, the threshold value α may be set from the perspective of evaluating the amount of information and determining whether effective learning can be performed using only the feature quantities of the acquired correct answer data. For example, the threshold value α may be set to a value that determines that when the amount of information is equal to or less than the threshold value α, the amount of information is small and therefore adding information to the feature quantities of the correct answer data would enable more effective learning.

[0067] The creation unit 504 creates learning data used for learning the learning model M. Specifically, for example, when the amount of information is equal to or less than a threshold value α, the creation unit 504 adds the acquired evidence information to the feature of the acquired correct answer data. Then, the creation unit 504 creates learning data in which the correct answer label is associated with the feature to which the evidence information is added. The second feature information 113 shown in FIG. 1 corresponds to the feature to which the evidence information is added.

[0068] On the other hand, when the amount of information is greater than the threshold α, the creation unit 504 creates learning data that associates the correct label with the feature of the acquired correct data. In this way, when the amount of information is small, for example, the creation unit 504 adds evidence information to the feature of the correct data, enabling learning based on the evidence information. On the other hand, when the amount of information is large, the creation unit 504 does not add evidence information, enabling learning that directly utilizes the feature of the correct data.

[0069] The learning unit 505 uses the created learning data to learn the learning model M. Specifically, for example, the learning unit 505 uses the created learning data by machine learning such as deep learning to perform supervised learning to learn the learning model M.

[0070] More specifically, for example, the learning unit 505 provides the created learning data to a learning device to learn the learning model M. The learning device is a learning system in machine learning, and for example, compares a result estimated based on the provided data with a correct result and adjusts various parameters to learn the learning model M.

[0071] <When estimating using learning model M> Next, the functional units used during estimation using the learning model M will be described.

[0072] The receiving unit 506 receives estimation target data. Here, the estimation target data is information that identifies a target for which a solution to a problem is to be estimated. For example, if the problem is "pathogenicity of fusion genes," the solution is whether or not the fusion gene is pathogenic. In this case, the estimation target data is, for example, genetic information indicating a fusion gene whose pathogenicity has not been fully elucidated.

[0073] Specifically, for example, the acquiring unit 501 may acquire the estimation target data by receiving it from the user terminal 203 shown in Fig. 2. Furthermore, the acquiring unit 501 may acquire the estimation target data by an operation input by a user (for example, a user of the information processing system 200) using an input device (not shown) of the acquiring unit 501.

[0074] The estimation unit 507 acquires an estimation result for the received estimation target data using the trained learning model M. Specifically, for example, the estimation unit 507 acquires a feature amount (third feature information) of the received estimation target data. Here, the feature amount of the estimation target data is feature information that represents the feature of the estimation target data. What information is acquired as the feature amount of the estimation target data is predetermined depending on, for example, what the estimation target data indicates.

[0075] For example, in the case of genetic information, the feature quantities of the inference target data to be acquired include domain information of each gene that forms the fusion gene indicated by the genetic information. However, the feature quantities of the inference target data acquired here do not include evidence information. The feature quantities of the inference target data are represented in the same data format as the feature quantities during training, for example, by a tree-structured graph.

[0076] More specifically, for example, the estimation unit 507 identifies two genes that form a fusion gene indicated by the estimation target data (gene information). Next, the estimation unit 507 refers to the knowledge database 220 and extracts domain information for each of the identified genes. Then, the estimation unit 507 generates a feature graph based on the extracted domain information.

[0077] The feature graph corresponds to the feature amounts of the estimation target data, and is, for example, a tree-structured graph (directed graph or undirected graph) with extracted domain information as nodes. An example of generating a feature graph representing the feature amounts of the estimation target data will be described later with reference to FIG.

[0078] The estimation unit 507 may acquire the feature amount (for example, a feature graph) of the estimation target data by receiving it from the user terminal 203. The estimation unit 507 may also acquire the feature amount (for example, a feature graph) of the estimation target data by a user's operation input using an input device (not shown) of the estimation unit 507.

[0079] Furthermore, the estimation unit 507 acquires evidence information (second evidence information) of the estimation target data. Here, the evidence information of the estimation target data is information that can be a basis for the correct answer for the estimation target data. What information is acquired as the evidence information of the estimation target data is determined in advance depending on, for example, what the estimation target data indicates.

[0080] For example, in the case of genetic information, the evidence information of the estimation target data to be acquired may include information indicating the number of literatures regarding the pathogenicity of the fusion gene indicated by the genetic information. Literature regarding pathogenicity is, for example, a paper or article asserting that the fusion gene indicated by the estimation target data is pathogenic or not pathogenic. Specifically, for example, the estimation unit 507 may acquire the evidence information of the estimation target data (genetic information) by referring to the storage unit 510.

[0081] More specifically, for example, the estimation unit 507 refers to the knowledge database 220 to identify a node representing a fusion gene indicated by the estimation target data (gene information). Then, the estimation unit 507 counts the number of nodes representing related literature that are connected to the identified node by edges with the attribute "evidence." The estimation unit 507 may then acquire the counted number of nodes as evidence information.

[0082] For example, if the number of nodes representing related documents is "3", the evidence information of the inference target data (gene information) will be "3". Also, if the number of nodes representing related documents is "0", the evidence information of the inference target data (gene information) will be "0". An example of obtaining evidence information of the inference target data will be described later with reference to FIG. 17, for example.

[0083] Next, the estimation unit 507 adds the acquired evidence information to the feature quantities of the acquired estimation target data. More specifically, for example, the estimation unit 507 adds a node representing the evidence information to the generated feature graph. Then, the estimation unit 507 inputs the feature quantities with the evidence information added (fourth feature information (e.g., feature graph)) into the trained learning model M, thereby acquiring an estimation result for the estimation target data.

[0084] The output unit 508 outputs the acquired estimation result in association with the estimation target data. The output format of the output unit 508 may be, for example, storage in a storage device such as the memory 302 or the disk 304, transmission to another computer via the communication I / F 305, display on a display (not shown), or printout on a printer (not shown).

[0085] For example, suppose that estimation target data (genetic information) is accepted by receiving it from the user terminal 203. In this case, the output unit 508 transmits an estimation result for the estimation target data (genetic information) to the user terminal 203. The estimation result for the estimation target data (genetic information) indicates whether or not the fusion gene identified by the genetic information is pathogenic.

[0086] Furthermore, the output unit 508 may output the created learning data to another computer. The other computer may be, for example, the management terminal 202 or the user terminal 203. In this case, the learning model M may be trained in the other computer using the created learning data.

[0087] Furthermore, the output unit 508 may output the trained learning model M to another computer. In this case, the other computer may receive estimation target data and use the trained learning model M to obtain an estimation result for the received estimation target data.

[0088] In the above description, the feature amount of each data (correct answer data, estimation target data) is represented by a tree-structured graph (directed graph or undirected graph), but this is not limiting. For example, the feature amount of each data may be represented by fixed-length data. The fixed-length data is, for example, a feature vector. The elements of the feature vector represent, for example, domain information, evidence information, etc.

[0089] Furthermore, the functional units (acquisition unit 501 to output unit 508) of the learning device 201 may be realized by multiple computers (e.g., the learning device 201, the management terminal 202, and the user terminal 203) in the information processing system 200. In this case, communication between the functional units of different computers is performed, for example, by transmission and reception between the functional units via the network 210.

[0090] (Example of operation during learning of learning model M) Next, an example of the operation of the learning device 201 when learning the learning model M will be described with reference to FIGS.

[0091] 6 to 15 are explanatory diagrams showing an example of the operation during learning of the learning model M. Here, a case where a correct answer data set 600 is acquired when learning the learning model M will be described as an example. The correct answer data set 600 includes one or more correct answer data (data to which a correct answer label has been assigned). In addition, a case where the data to which a correct answer label has been assigned is regarded as "gene information indicating a fusion gene" and the learning model M for estimating the pathogenicity of the fusion gene will be described as an example.

[0092] 6, the learning device 201 selects unselected (unprocessed) supervised data from the supervised data set 600. Here, it is assumed that supervised data 610 is selected. The supervised data 610 is genetic information 612 to which a supervised label 611 is assigned. The supervised label 611 indicates the correct answer "pathogenic." The genetic information 612 indicates the fusion gene "BCR::ABL1."

[0093] In this case, the learning device 201 generates, for example, a feature graph as the feature amount of the genetic information 612. Specifically, for example, the learning device 201 identifies two genes that form the fusion gene "BCR::ABL1" indicated by the genetic information 612. Here, the gene "BCR" and the gene "ABL1" are identified.

[0094] The learning device 201 extracts domain information for each of the identified genes "BCR" and "ABL1" by referring to the knowledge database 220. More specifically, for example, the learning device 201 refers to the knowledge database 220 to identify the knowledge graph 400 corresponding to the fusion gene "BCR::ABL1."

[0095] Next, the learning device 201 extracts domain information of the gene "BCR" from the identified knowledge graph 400. For example, the learning device 201 refers to the knowledge graph 400, traces the edge 421 of the attribute "5'gene" from the node 401 representing the fusion gene "BCR::ABL1", and identifies the node 402 representing the gene "BCR".

[0096] The learning device 201 traces the edges of the attribute "domain" (e.g., edge 424) from the identified node 402 to identify nodes 407-411 that represent domains. For example, node 407 represents the domain "Bcr-Abl_Oligo." The learning device 201 then extracts the domains represented by the identified nodes 407-411 as domain information for the gene "BCR."

[0097] Similarly, the learning device 201 extracts domain information of the gene "ABL1" from the knowledge graph 400. Here, the domains represented by nodes 412 to 415 are extracted as the domain information of the gene "ABL1." Then, the learning device 201 generates a feature graph G1 based on the extracted domain information of each of the genes "BCR" and "ABL1."

[0098] The feature graph G1 is a directed graph with a tree structure including nodes N10 to N19. Node N10 is the node representing the starting point. Nodes N11 to N19 represent domain information of the genes "BCR" and "ABL1" and correspond to those representing the features of the gene information 612 (fusion gene "BCR::ABL1"). Each of nodes N11 to N19 is connected to node N10 by an edge of the attribute "domain", for example.

[0099] 7, the learning device 201 calculates the information amount of the generated feature graph G1 (feature amount of the genetic information 612). Here, the learning device 201 calculates the number of nodes N11 to N19 representing domain information, "9," among the nodes N10 to N19 in the feature graph G1, as the information amount of the feature graph G1 (feature amount of the genetic information 612).

[0100] Next, the learning device 201 determines whether the calculated information amount "9" is equal to or less than the threshold value α. Here, the threshold value α is set to "α=5." In this case, the learning device 201 determines that the calculated information amount "9" is greater than the threshold value α. Then, the learning device 201 creates learning data D1 as shown in FIG. 8 without adding evidence information.

[0101] 8, the learning data D1 is information that associates the correct answer label 611 with the feature graph G1 (the feature amount of the genetic information 612). As a result, when the amount of information is large, the learning device 201 enables learning that makes use of the feature amount of the genetic information 612 as it is, without adding evidence information.

[0102] 9, the learning device 201 selects unselected (unprocessed) supervised data from the supervised data set 600. Here, it is assumed that supervised data 620 is selected. The supervised data 620 is genetic information 622 to which a supervised label 621 is assigned. The supervised label 621 indicates the correct answer "pathogenic." The genetic information 622 indicates the fusion gene "JAZF1::SUZ12."

[0103] In this case, the learning device 201 generates, for example, a feature graph as the feature amount of the genetic information 622. Specifically, for example, the learning device 201 identifies two genes that form the fusion gene "JAZF1::SUZ12" indicated by the genetic information 622. Here, the gene "JAZF1" and the gene "SUZ12" are identified.

[0104] The learning device 201 extracts domain information for each of the identified genes "JAZF1" and "SUZ12" by referring to the knowledge database 220. Then, the learning device 201 generates a feature graph G2 based on the extracted domain information for each of the genes "JAZF1" and "SUZ12".

[0105] The feature graph G2 is a directed tree graph including nodes N20 and N21. Node N20 is the node representing the starting point. Node N21 represents the domain information of either gene "JAZF1" or "SUZ12" and corresponds to the characteristic of gene information 622 (fusion gene "JAZF1::SUZ12"). Node N21 is connected to node N20 by, for example, an edge of the attribute "domain."

[0106] Next, the learning device 201 calculates the information amount of the generated feature graph G2 (feature amount of the genetic information 622). Here, the learning device 201 calculates the number of nodes N21 representing domain information, "1," among the nodes N20 and N21 in the feature graph G2, as the information amount of the feature graph G2 (feature amount of the genetic information 622).

[0107] 10, the learning device 201 determines whether the calculated information amount "1" is equal to or less than the threshold value α. Here, the threshold value α is "α=5." Therefore, the learning device 201 determines that the calculated information amount "1" is equal to or less than the threshold value α. In this case, the learning device 201 refers to the knowledge database 220 to acquire evidence information about the genetic information 622.

[0108] Here, it is assumed that the number of documents claiming that the fusion gene "JAZF1::SUZ12" indicated by the genetic information 622 (see FIG. 9) is "pathogenic" as indicated by the correct answer label 621, "5", is acquired as evidence information. In this case, the learning device 201 adds the acquired evidence information "5" to the feature of the genetic information 622.

[0109] Specifically, for example, the learning device 201 adds a node N22 representing evidence information "5" to the feature graph G2 (the feature amount of the genetic information 622). More specifically, for example, the learning device 201 adds the node N22 to the feature graph G2 by connecting the node N22 to the node N20 representing the starting point. The node N22 is connected to the node N20 by an edge of the attribute "evidences", for example.

[0110] Then, the learning device 201 creates learning data D2. Learning data D2 is information that associates the correct answer label 621 with a feature graph G2 (feature amount of genetic information 622) to which a node N22 representing evidence information "5" has been added. As a result, when the amount of information is small, the learning device 201 supplements the amount of information by adding the evidence information "5" to the feature amount of the genetic information 622, thereby enabling learning that takes the evidence information "5" into consideration.

[0111] 11, the learning device 201 selects unselected (unprocessed) supervised data from the supervised data set 600. Here, it is assumed that supervised data 630 is selected. The supervised data 630 is genetic information 632 to which a supervised label 631 is assigned. The supervised label 631 indicates the correct answer "non-pathogenic." The genetic information 632 indicates the fusion gene "COL6A2::STAT6."

[0112] In this case, the learning device 201 generates, for example, a feature graph as the feature amount of the genetic information 632. Specifically, for example, the learning device 201 identifies two genes that form the fusion gene "COL6A2::STAT6" indicated by the genetic information 632. Here, the gene "COL6A2" and the gene "STAT6" are identified.

[0113] The learning device 201 extracts domain information for each of the identified genes "COL6A2" and "STAT6" by referring to the knowledge database 220. Then, the learning device 201 generates a feature graph G3 based on the extracted domain information for each of the genes "COL6A2" and "STAT6".

[0114] The feature graph G3 is a directed graph with a tree structure including nodes N30 to N38. Node N30 is the node representing the starting point. Nodes N31 to N38 represent domain information of the genes "COL6A2" and "STAT6" and correspond to those representing the features of the gene information 632 (fusion gene "COL6A2::STAT6"). Each of nodes N31 to N38 is connected to node N30 by an edge of the attribute "domain," for example.

[0115] 12, the learning device 201 calculates the information amount of the generated feature graph G3 (the feature amount of the genetic information 632). Here, the learning device 201 calculates the number of nodes N31 to N38 representing domain information, "8," among the nodes N30 to N38 in the feature graph G3, as the information amount of the feature graph G3 (the feature amount of the genetic information 632).

[0116] Next, the learning device 201 determines whether the calculated information amount "8" is equal to or less than the threshold value α. Here, the threshold value α is "α=5." Therefore, the learning device 201 determines that the calculated information amount "8" is greater than the threshold value α. In this case, the learning device 201 creates learning data D3 as shown in FIG. 13 without adding evidence information.

[0117] 13, learning data D3 is information that associates correct answer label 631 with feature graph G3 (feature amount of genetic information 632). This allows the learning device 201 to perform learning that makes use of the feature amount of genetic information 632 as it is, without adding evidence information, when the amount of information is large.

[0118] In FIG. 14, the learning device 201 selects unselected (unprocessed) correct answer data from the correct answer dataset 600. Here, it is assumed that correct answer data 640 is selected. The correct answer data 640 is genetic information 642 to which a correct answer label 641 is assigned. The correct answer label 641 indicates the correct answer "non-pathogenic." The genetic information 642 indicates the fusion gene "CYP2C9::CYP2C19."

[0119] In this case, the learning device 201 generates, for example, a feature graph as the feature amount of the genetic information 642. Specifically, for example, the learning device 201 identifies two genes that form the fusion gene "CYP2C9::CYP2C19" indicated by the genetic information 642. Here, the gene "CYP2C9" and the gene "CYP2C19" are identified.

[0120] The learning device 201 extracts domain information for each of the identified genes "CYP2C9" and "CYP2C19" by referring to the knowledge database 220. Then, the learning device 201 generates a feature graph G4 based on the extracted domain information for each of the genes "CYP2C9" and "CYP2C19".

[0121] The feature graph G4 is a directed tree graph including nodes N40 and N41. Node N40 is the starting node. Node N41 represents the domain information of either the gene "CYP2C9" or "CYP2C19" and corresponds to the characteristic of the gene information 642 (fusion gene "CYP2C9::CYP2C19"). Node N41 is connected to node N40 by an edge of the attribute "domain," for example.

[0122] Next, the learning device 201 calculates the information amount of the generated feature graph G4 (feature amount of the genetic information 642). Here, the learning device 201 calculates the number of nodes N41 representing domain information ("1") among the nodes N40 and N41 in the feature graph G4 as the information amount of the feature graph G4 (feature amount of the genetic information 642).

[0123] 15, the learning device 201 determines whether the calculated information amount "1" is equal to or less than the threshold value α. Here, the threshold value α is "α=5." Therefore, the learning device 201 determines that the calculated information amount "1" is equal to or less than the threshold value α. In this case, the learning device 201 refers to the knowledge database 220 and acquires evidence information about the genetic information 642 (see FIG. 14).

[0124] Here, it is assumed that the number of documents "0" asserting that the fusion gene "CYP2C9::CYP2C19" indicated by the genetic information 642 is the correct answer "non-pathogenic" indicated by the correct answer label 641 is acquired as evidence information. The number of documents "0" corresponds to the absence of evidence that the correct answer is "non-pathogenic."

[0125] In this case, the learning device 201 adds the acquired evidence information "0" to the feature amount of the genetic information 642. Specifically, for example, the learning device 201 adds a node N42 representing the evidence information "0" to the feature graph G4 (the feature amount of the genetic information 642). To explain in more detail, for example, the learning device 201 adds node N42 to the feature graph G4 by connecting node N42 to node N40 representing the starting point. Node N42 is connected to node N40 by, for example, an edge of the attribute "evidences."

[0126] Then, the learning device 201 creates learning data D4. Learning data D4 is information that associates the correct answer label 641 with a feature graph G4 (feature amount of genetic information 642) to which a node N42 representing evidence information "0" has been added. As a result, when the amount of information is small, the learning device 201 supplements the amount of information by adding evidence information "0" to the feature amount of genetic information 642, enabling learning that takes evidence information "0" into consideration.

[0127] When there is no unselected (unprocessed) supervised data remaining from the supervised data set 600, the learning device 201 uses the created learning data (e.g., learning data D1 to D4) to learn the learning model M. This allows the learning device 201 to generate a highly accurate learning model M.

[0128] (Example of operation when estimating using learning model M) Next, an example of the operation of the learning device 201 during estimation using the learning model M will be described with reference to FIGS.

[0129] 16 and 17 are explanatory diagrams showing an example of the operation during estimation using the learning model M. Here, an example will be described in which genetic information 1601 is received as estimation target data. The genetic information 1601 indicates the fusion gene "RNF130::BRAF." This fusion gene "RNF130::BRAF" is, for example, a fusion gene whose pathogenicity is unknown to the user.

[0130] In this case, the learning device 201 generates, for example, a feature graph as the feature amount of the genetic information 1601. Specifically, for example, the learning device 201 identifies two genes that form the fusion gene "RNF130::BRAF" indicated by the genetic information 1601. Here, the gene "RNF130" and the gene "BRAF" are identified.

[0131] The learning device 201 extracts domain information for each of the identified genes, "RNF130" and "BRAF," by referring to the knowledge database 220. Then, the learning device 201 generates a feature graph 1602 based on the extracted domain information for each of the genes, "RNF130" and "BRAF."

[0132] The feature graph 1602 is a directed graph with a tree structure including nodes 1610 to 1617. Node 1610 is the node representing the starting point. Nodes 1611 to 1617 represent domain information of either the gene "RNF130" or "BRAF," and correspond to those representing the features of the gene information 1601 (fusion gene "RNF130::BRAF"). Each of the nodes 1611 to 1617 is connected to the node 1610 by an edge of the attribute "domain," for example.

[0133] 17, the learning device 201 refers to the knowledge database 220 to acquire evidence information about the genetic information 1601. Here, it is assumed that the number of documents that could be the basis for the correct answer (for example, documents claiming that the gene is pathogenic) for the fusion gene "RNF130::BRAF" indicated by the genetic information 1601, "1", is acquired as evidence information.

[0134] In this case, the learning device 201 adds the acquired evidence information "1" to the feature of the genetic information 1601. Specifically, for example, the learning device 201 adds node 1618 representing the evidence information "1" to the feature graph 1602 (the feature of the genetic information 1601). To explain in more detail, for example, the learning device 201 adds node 1618 to the feature graph 1602 by connecting node 1618 to node 1610 representing the starting point. Node 1618 is connected to node 1610 by, for example, an edge of the attribute "evidences."

[0135] Then, the learning device 201 inputs the feature graph 1602 (feature of the genetic information 1601) to which the node 1618 representing the evidence information "1" has been added into the learning model M, thereby obtaining an inference result 1603 for the genetic information 1601. Here, the inference result 1603 indicates that the fusion gene "RNF130::BRAF" indicated by the genetic information 1601 is "pathogenic."

[0136] (Various processing procedures of the learning device 201) Next, we will explain various processing procedures of the learning device 201. First, we will explain the learning processing procedures of the learning device 201 with reference to FIG.

[0137] 18 is a flowchart showing an example of the learning process procedure of the learning device 201. In the flowchart of FIG. 18, first, the learning device 201 acquires a correct answer dataset (step S1801). The correct answer dataset includes one or more pieces of correct answer data. The correct answer data is data to which a correct answer label has been assigned.

[0138] Here, the data to which the correct label is assigned (correct data) is defined as genetic information. The genetic information indicates a fusion gene. The correct label indicates whether the fusion gene is pathogenic or not.

[0139] Next, the learning device 201 selects unselected supervised data from the acquired supervised data set (step S1802).Then, the learning device 201 refers to the knowledge database 220 and extracts domain information of each gene that forms the fusion gene indicated by the genetic information included in the selected supervised data (step S1803).

[0140] Next, the learning device 201 generates a feature graph representing the features of the gene information (fusion gene) based on the extracted domain information of each gene (step S1804).The learning device 201 then calculates the amount of information in the feature graph by counting the number of nodes representing domain information among the nodes in the generated feature graph (step S1805).

[0141] Next, the learning device 201 determines whether the calculated amount of information is equal to or less than a threshold value α (step S1806). If the amount of information is equal to or less than the threshold value α (step S1806: Yes), the learning device 201 refers to the knowledge database 220 and acquires evidence information about the genetic information included in the selected supervised data (step S1807).

[0142] The learning device 201 then adds a node representing the obtained evidence information to the generated feature graph (step S1808). Next, the learning device 201 creates learning data that associates the correct answer label with the feature graph to which the node representing the evidence information has been added (step S1809), and proceeds to step S1811.

[0143] If the calculated information amount is greater than the threshold value α (step S1806: No), the learning device 201 creates learning data that associates the correct answer label with the generated feature graph (step S1810).Then, the learning device 201 determines whether there is any unselected correct answer data that has not been selected from the correct answer dataset (step S1811).

[0144] If there is unselected supervised data (step S1811: Yes), the learning device 201 returns to step S1802. On the other hand, if there is no unselected supervised data (step S1811: No), the learning model M is trained using the created learning data (step S1812), and the series of processes according to this flowchart ends.

[0145] This allows the learning device 201 to generate a highly accurate learning model M by utilizing evidence information.

[0146] Next, the estimation process procedure of the learning device 201 will be described with reference to FIG.

[0147] 19 is a flowchart showing an example of an estimation processing procedure of the learning device 201. In the flowchart of FIG. 19, first, the learning device 201 determines whether estimation target data has been received (step S1901). The estimation target data is information that identifies a target for which a solution to a problem is to be estimated.

[0148] Here, the problem is defined as "pathogenicity of the fusion gene," the solution is defined as "presence or absence of pathogenicity," and the data to be estimated is defined as "gene information indicating the fusion gene."

[0149] The learning device 201 waits to receive inference target data (step S1901: No). If the learning device 201 receives inference target data (step S1901: Yes), it refers to the knowledge database 220 and extracts domain information of each gene that forms the fusion gene indicated by the inference target data (step S1902).

[0150] Next, the learning device 201 generates a feature graph representing the features of the inference target data (fusion gene) based on the extracted domain information of each gene (step S1903). Next, the learning device 201 refers to the knowledge database 220 to acquire evidence information about the inference target data (fusion gene) (step S1904).

[0151] Then, the learning device 201 adds a node representing the acquired evidence information to the generated feature graph (step S1905). Next, the learning device 201 inputs the feature graph to which the node representing the evidence information has been added to the trained learning model M, thereby acquiring an inference result for the inference target data (fusion gene) (step S1906).

[0152] Then, the learning device 201 outputs the acquired estimation result in association with the estimation target data (fusion gene) (step S1907), and ends the series of processes according to this flowchart.

[0153] This allows the learning device 201 to accurately estimate the presence or absence of pathogenicity for the fusion gene indicated by the estimation target data.

[0154] As described above, the learning device 201 according to the embodiment can calculate the information amount of the feature amount (first feature information) of the correct answer data (first data) when learning the learning model M. The learning model M is a machine learning model that receives the feature amount of data as input and outputs an estimation result for the data. The correct answer data is data to which a correct answer label is assigned. Then, when the calculated information amount is equal to or less than the threshold α, the learning device 201 can add evidence information (first evidence information) acquired about the correct answer data to the feature amount of the correct answer data, and generate learning data in which the correct answer label is associated with the feature amount of the correct answer data to which the evidence information is added (second feature information). The evidence information is information that can serve as a basis for the correct answer (correct answer label). Furthermore, when the calculated information amount is greater than the threshold α, the learning device 201 can generate learning data in which the correct answer label is associated with the feature amount of the correct answer data.

[0155] This allows the learning device 201 to enable learning that utilizes evidence information while preventing a decrease in the estimation accuracy of the learning model M. For example, when the amount of information is small, the learning device 201 can add evidence information to the features of the correct answer data to enable learning based on the evidence information. Furthermore, when the amount of information is large, the learning device 201 can enable learning that makes use of the features of the correct answer data as they are, without adding evidence information.

[0156] Furthermore, the learning device 201 can learn the learning model M using the created learning data.

[0157] This allows the learning device 201 to generate a highly accurate learning model M by utilizing evidence information that is almost equivalent to the correct answer.

[0158] Furthermore, learning device 201 can acquire the feature amount of the supervised data by referring to knowledge database 220. Furthermore, learning device 201 can acquire evidence information about the supervised data by referring to knowledge database 220 when the calculated amount of information is equal to or less than threshold value α.

[0159] This allows the learning device 201 to acquire the feature amounts and evidence information of the correct answer data by using a knowledge database (for example, knowledge database 220) in which open data and the like are accumulated.

[0160] Furthermore, according to the learning device 201, the features of the correct data can be represented by a feature graph in which each piece of information representing the feature of the correct data is a node, and the number of nodes in the feature graph (nodes representing each piece of information representing the feature of the correct data) can be calculated as the amount of information of the features of the correct data.

[0161] This allows the learning device 201 to quantitatively evaluate the information amount of the feature amount of the correct answer data.

[0162] Furthermore, according to the learning device 201, the correct answer data can be "gene information (fusion gene) to which a correct answer label indicating the presence or absence of pathogenicity has been assigned." Furthermore, according to the learning device 201, information indicating the number of literature regarding the pathogenicity of the fusion gene indicated by the correct answer data (gene information) can be used as evidence information.

[0163] This allows the learning device 201 to generate a learning model M that can accurately estimate the pathogenicity of a fusion gene.

[0164] Furthermore, according to the learning device 201, the feature amount of the correct answer data can be made to include domain information of each gene that forms the fusion gene indicated by the correct answer data.

[0165] This allows the learning device 201 to accurately represent the characteristics of the correct answer data.

[0166] Furthermore, the learning device 201 can receive inference target data and acquire the feature quantities (third feature information) of the received inference target data and evidence information (second evidence information) about the inference target data. The evidence information is information that can serve as a basis for a solution about the inference target data. The learning device 201 then adds the acquired evidence information to the feature quantities of the inference target data, and inputs the feature quantities of the inference target data with the added evidence information (fourth feature information) into the trained learning model M, thereby acquiring an inference result about the inference target data.

[0167] This allows the learning device 201 to accurately estimate a solution for the estimation target data. For example, the learning device 201 can add evidence information to supervised answer data with a small amount of information and perform learning based on the evidence information, thereby enabling highly accurate estimation of a solution for estimation target data for which evidence exists. Furthermore, the learning device 201 can perform learning that makes use of many feature quantities without adding evidence information to supervised answer data with a large amount of information, thereby enabling estimation of a solution for estimation target data for which evidence does not exist by utilizing the feature quantities of the estimation target data.

[0168] Furthermore, the learning device 201 can output the acquired estimation results in association with the estimation target data.

[0169] This allows the learning device 201 to provide an inference result for the inference target data. For example, the learning device 201 can provide an inference service in which it receives inference target data indicating a certain fusion gene and outputs an inference result indicating whether the fusion gene is pathogenic or not.

[0170] From these facts, the learning device 201 according to the embodiment can generate a highly accurate learning model M by utilizing evidence information obtained through curation in fields where open data is developing.

[0171] For example, the learning device 201 can allow data leaks while also allowing features other than leaked information (evidence information) to be sufficiently learned to generate the learning model M. This allows the learning device 201 to accurately estimate a solution even for data for which there is no evidence at the time of estimation, compared to when learning by uniformly including evidence information in features.

[0172] For example, the learning device 201 can utilize information (evidence information) that is nearly equivalent to the correct answer to accurately estimate the pathogenicity of a fusion gene. This enables the learning device 201 to accurately and quickly determine the pathogenicity of a fusion gene found in a patient's genome, which can be useful in determining optimal treatment strategies and contributing to the discovery of new drug targets, thereby contributing to the development of genomic medicine.

[0173] This learning method can be applied not only to estimating the pathogenicity of fusion genes, but also to various other cases where information nearly equivalent to the correct answer can be obtained. For example, when information such as papers on the harmfulness of chemical substances is available as evidence, this learning method can be applied to training a learning model M that estimates the harmfulness of those chemical substances based on their feature quantities. The feature quantities of chemical substances represent, for example, the characteristics of each component of the chemical substance.

[0174] Furthermore, when information such as papers on disease names corresponding to the patient's symptoms is available as evidence, this learning method can be applied to training a learning model M that predicts the name of a patient's disease based on the patient's features, such as the patient's symptoms, gender, age, and weight.

[0175] The machine learning method described in this embodiment can be realized by executing a prepared program on a computer such as a personal computer or a workstation. This machine learning program is recorded on a computer-readable recording medium such as a hard disk, flexible disk, CD-ROM, DVD, or USB memory, and is executed by being read from the recording medium by a computer. This machine learning program may also be distributed via a network such as the Internet.

[0176] In addition, the information processing device 101 (learning device 201) described in this embodiment can also be realized, for example, by an application-specific IC such as a standard cell or a structured ASIC (Application Specific Integrated Circuit) or a PLD (Programmable Logic Device) such as an FPGA.

[0177] The following additional notes are provided regarding the above-described embodiment.

[0178] (Appendix 1) When training a machine learning model that uses feature information representing the characteristics of data as input and outputs an inference result for the data, Calculating an amount of information of first feature information representing a feature of the first data to which the correct answer has been assigned; If the calculated amount of information is equal to or less than a threshold, creating training data in which second feature information is added to the first feature information, and the correct answer is associated with the first evidence information that may be a basis for the correct answer obtained for the first data; If the amount of information is greater than the threshold, training data is created in which the first feature information and the correct answer are associated with each other. A machine learning program that causes a computer to execute processing.

[0179] (Appendix 2) The machine learning model is trained using the created training data. 2. The machine learning program according to claim 1, characterized in that the program causes the computer to execute processing.

[0180] (Appendix 3) Accept the data to be estimated, acquiring third feature information representing features of the received data of the estimation target and second evidence information that may be a basis for a solution regarding the data of the estimation target; and inputting fourth feature information obtained by adding the acquired second evidence information to the acquired third feature information into the trained machine learning model, thereby obtaining an inference result for the data to be inferred. 3. The machine learning program according to claim 2, characterized in that the program causes the computer to execute processing.

[0181] (Appendix 4) The acquired estimation result is output in association with the data to be estimated. 4. The machine learning program according to claim 3, characterized in that the program causes the computer to execute processing.

[0182] (Appendix 5) The first data represents a fusion gene, The correct answer indicates whether the fusion gene is pathogenic or not. 5. The machine learning program according to any one of claims 1 to 4.

[0183] (Appendix 6) The machine learning program described in Appendix 5, characterized in that the first evidence information includes information indicating the number of publications regarding the pathogenicity of the fusion gene indicated by the first data.

[0184] (Appendix 7) The machine learning program described in Appendix 5 or 6, characterized in that the first feature information includes domain information of each gene that forms the fusion gene indicated by the first data.

[0185] (Supplementary Note 8) The first feature information is acquired by referring to a knowledge graph including nodes representing each of a plurality of pieces of knowledge and edges representing relationships between pieces of knowledge included in the plurality of pieces of knowledge; If the amount of information is equal to or less than a threshold, the knowledge graph is referenced to acquire the first evidence information. 8. The machine learning program according to any one of claims 1 to 7, which causes the computer to execute processing.

[0186] (Supplementary Note 9) The first feature information is expressed by a feature graph having each piece of information representing a feature of the first data as a node; The calculation process includes: calculating the number of nodes representing each piece of information in the feature graph as an information amount of the first feature information; 9. The machine learning program according to any one of appendices 1 to 8,

[0187] (Appendix 10) When training a machine learning model that inputs feature information representing the features of data and outputs an inference result for the data, Calculating an amount of information of first feature information representing a feature of the first data to which the correct answer has been assigned; If the calculated amount of information is equal to or less than a threshold, creating training data in which second feature information is added to the first feature information, and the correct answer is associated with the first evidence information that may be a basis for the correct answer obtained for the first data; If the amount of information is greater than the threshold, training data is created in which the first feature information and the correct answer are associated with each other. A machine learning method characterized in that the processing is executed by a computer.

[0188] (Appendix 11) When training a machine learning model that inputs feature information representing the features of data and outputs an inference result for the data, Calculating an amount of information of first feature information representing a feature of the first data to which the correct answer has been assigned; If the calculated amount of information is equal to or less than a threshold, creating training data in which second feature information is added to the first feature information, and the correct answer is associated with the first evidence information that may be a basis for the correct answer obtained for the first data; If the amount of information is greater than the threshold, training data is created in which the first feature information and the correct answer are associated with each other. An information processing device comprising a control unit. [Explanation of symbols]

[0189] 101 Information processing equipment 102 Machine Learning Models 110 First Data 111 Correct 112 First characteristic information 113 Second characteristic information 120 First Evidence Information 130,140,D1~D4 training data 200 Information Processing Systems 201 Learning Device 202 Management terminal 203 User terminal 210 Network 220 Knowledge Database 300 Bus 301 CPU 302 memory 303 Disk Drive 304 Disk 305 Communication I / F 306 Portable Recording Media I / F 307 Portable Recording Media 400 Knowledge Graph 401~415, 1610~1618, N10~N19, N20~N22, N30~N38, N40, N41 nodes 421,422,423,424,425 Edge 500 control section 501 Acquisition Department 502 Calculation Unit 503 Judgment section 504 Creation Department 505 Learning Department 506 Reception 507 Estimation Department 508 Output Section 510 Storage section 600 correct answer datasets 610,620,630,640 Correct data 611,621,631,641 correct labels 612,622,632,642,1601 Genetic information 1602,G1~G4 feature graph 1603 Estimation result

Claims

1. When training a machine learning model that uses feature information representing the features of data as input and outputs an estimation result for the data, Calculating an amount of information of first feature information representing a feature of the first data to which the correct answer has been assigned; If the calculated amount of information is equal to or less than a threshold, creating training data in which second feature information is added to the first feature information, and the correct answer is associated with the first evidence information that may be a basis for the correct answer obtained for the first data; If the amount of information is greater than the threshold, training data is created in which the first feature information and the correct answer are associated with each other. A machine learning program that causes a computer to execute processing.

2. training the machine learning model using the created training data; The machine learning program according to claim 1, wherein the program causes the computer to execute processing.

3. Accept the data to be estimated, acquiring third feature information representing features of the received data of the estimation target and second evidence information that may be a basis for a solution regarding the data of the estimation target; and inputting the acquired second evidence information and fourth feature information obtained by adding the acquired third feature information into the trained machine learning model to obtain an inference result for the data to be inferred.

3. The machine learning program according to claim 2, wherein the program causes the computer to execute processing.

4. outputting the acquired estimation result in association with the data to be estimated; 4. The machine learning program according to claim 3, wherein the program causes the computer to execute processing.

5. the first data indicates a fusion gene; The correct answer indicates whether the fusion gene is pathogenic or not.

5. The machine learning program according to claim 1, wherein the machine learning program is a program for generating a machine learning algorithm.

6. The machine learning program according to claim 5 , wherein the first evidence information includes information indicating the number of publications relating to the pathogenicity of the fusion gene indicated by the first data.

7. When training a machine learning model that uses feature information representing the features of data as input and outputs an estimation result for the data, Calculating an amount of information of first feature information representing a feature of the first data to which the correct answer has been assigned; If the calculated amount of information is equal to or less than a threshold, creating training data in which second feature information is added to the first feature information, and the correct answer is associated with the first evidence information that may be a basis for the correct answer obtained for the first data; If the amount of information is greater than the threshold, training data is created in which the first feature information and the correct answer are associated with each other. A machine learning method characterized in that the processing is executed by a computer.

8. When training a machine learning model that uses feature information representing the features of data as input and outputs an estimation result for the data, Calculating an amount of information of first feature information representing a feature of the first data to which the correct answer has been assigned; If the calculated amount of information is equal to or less than a threshold, creating training data in which second feature information is added to the first feature information, and the correct answer is associated with the first evidence information that may be a basis for the correct answer obtained for the first data; If the amount of information is greater than the threshold, training data is created in which the first feature information and the correct answer are associated with each other. An information processing device comprising a control unit.

Images