Information Processing System, Information Processing Method, and Information Processing Program

Abstract

An information processing system according to an embodiment acquires numerical representations and compounding ratios for respective multiple component objects, and performs machine learning and application of the multiple compounding ratios based on the multiple numerical representations and the multiple compounding ratios corresponding to the multiple component objects, thereby calculating a composite feature vector representing features of a composite object obtained by compounding the multiple component objects, and outputs the composite feature vector.

Description

Technical Field

[0001] One aspect of the present invention relates to an information processing system, an information processing method, and an information processing program. Background Art

[0002] A method of analyzing a composite object obtained by combining multiple component objects using machine learning is used. For example, Patent Document 1 describes a method of predicting the bondability between the three-dimensional structure of a biopolymer and the three-dimensional structure of a compound. The method includes the following steps: generating a predicted three-dimensional structure of a complex of a biopolymer and a compound based on the three-dimensional structure of the biopolymer and the three-dimensional structure of the compound; transforming the predicted three-dimensional structure into a predicted three-dimensional structure vector representing the result of checking the interaction pattern; and predicting the bondability between the three-dimensional structure of the biopolymer and the three-dimensional structure of the compound by using a machine learning algorithm to judge the predicted three-dimensional structure vector.

[0003] Prior Art Documents

[0004] Patent Documents

[0005] Patent Document 1: Japanese Unexamined Patent Application Publication No. 2019-28879 Summary of the Invention

[0006] Technical Problem to be Solved by the Invention

[0007] In the case where the component objects are diverse or there are multiple of them, it is not possible to prepare a sufficient amount of data for these component objects. As a result, the analysis accuracy of the composite object may not reach the expected level. Therefore, a structure is desired that can improve the analysis accuracy of the composite object even when a sufficient amount of data cannot be prepared for the component objects.

[0008] Means for Solving the Technical Problem

[0009] The information processing system according to one aspect of the present invention includes at least one processor. The at least one processor performs the following processing: acquiring a numerical representation and a compounding ratio for each of a plurality of component objects; performing machine learning and applying the plurality of compounding ratios based on the plurality of numerical representations and the plurality of compounding ratios corresponding to the plurality of component objects, thereby calculating a composite feature vector representing the characteristics of a composite object obtained by combining the plurality of component objects; and outputting the composite feature vector.

[0010] An information processing method according to an aspect of the present invention is executed by an information processing system including at least one processor. The information processing method includes the following steps: obtaining numerical representations and compound ratios for respective multiple component objects; performing machine learning and applying the multiple compound ratios based on the multiple numerical representations and the multiple compound ratios corresponding to the multiple component objects, thereby calculating a composite feature vector representing the characteristics of a composite object obtained by compounding the multiple component objects; and outputting the composite feature vector.

[0011] An information processing program according to an aspect of the present invention causes a computer to execute the following steps: obtaining numerical representations and compound ratios for respective multiple component objects; performing machine learning and applying the multiple compound ratios based on the multiple numerical representations and the multiple compound ratios corresponding to the multiple component objects, thereby calculating a composite feature vector representing the characteristics of a composite object obtained by compounding the multiple component objects; and outputting the composite feature vector.

[0012] In such a manner, since machine learning and application of the compound ratio are performed for each component object, the analysis accuracy of the composite object can be improved even when it is not possible to prepare a sufficient amount of data for the component object.

[0013] Advantageous Effects of the Invention

[0014] According to an aspect of the present invention, the analysis accuracy of the composite object can be improved even when it is not possible to prepare a sufficient amount of data for the component object. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 FIG. is an example of a hardware configuration of a computer constituting the information processing system according to the embodiment.

[0016] Figure 2 FIG. is an example of a functional configuration of the information processing system according to the embodiment.

[0017] Figure 3 FIG. is a flowchart of an example of the operation of the information processing system according to the embodiment.

[0018] Figure 4 FIG. is an example of a diagram showing the order of calculating the composite feature vector.

[0019] Figure 5 FIG. is an example of a diagram showing the application of the compound ratio in the middle of machine learning.

[0020] Figure 6 FIG. is a diagram showing a specific example of the order of calculating the composite feature vector.

[0021] Figure 7 FIG. is another example of a diagram showing the order of calculating the composite feature vector. Detailed Embodiments

[0022] Hereinafter, with reference to the drawings, embodiments of the present invention will be described in detail. In the description of the drawings, the same or equivalent elements are denoted by the same reference numerals, and redundant descriptions are omitted.

[0023] [System Overview]

[0024] The information processing system 10 according to the embodiment is a computer system that performs analysis related to a composite object obtained by combining multiple component objects at a given composite ratio. A component object refers to a tangible or intangible object used to generate a composite object. The composite object can be a tangible or intangible object. Examples of tangible objects include any substance or object. Examples of intangible objects include data and information. "Combining multiple component objects" means treating multiple component objects as one object, that is, the composite object. The method of combination is not limited. For example, it can be blending, mixing, synthesizing, combining, fusing, merging, or other methods. Analysis related to the composite object refers to a process for obtaining data representing certain characteristics of the composite object.

[0025] The multiple component objects can be any variety of materials. In this case, the composite object is a multi-component substance generated from these materials. A material refers to any constituent element used to generate a multi-component substance. For example, the multiple materials can be any variety of molecules, atoms, molecular structures, crystal structures, or amino acid sequences. In this case, the composite object is a multi-component substance obtained by combining these molecules, atoms, molecular structures, crystal structures, or amino acid sequences by any method. For example, the material can be a polymer, while the multi-component substance can be a polymer alloy. The material can be a monomer, while the multi-component substance can be a polymer. The material can be a drug, that is, a chemical substance having a pharmacological effect, while the multi-component substance can be a pharmaceutical agent.

[0026] The information processing system 10 performs machine learning for analysis related to the composite object. Machine learning refers to a method of autonomously finding laws or rules by learning based on given information. The specific method of machine learning is not limited. For example, the information processing system 10 can perform machine learning using a machine learning model that is a computational model including a neural network. A neural network is an information processing model that simulates the structure of the human brain nervous system. As a more specific example, the information processing system 10 can also perform machine learning using at least one of a graph neural network (GNN), a convolutional neural network (CNN), a recurrent neural network (RNN), an attention RNN, and multi-head attention.

[0027] [Structure of the System]

[0028] The information processing system 10 is composed of one or more computers. When multiple computers are used, these computers are connected via a communication network such as the Internet or an intranet, thereby logically constructing an information processing system 10.

[0029] Figure 1 FIG. is an example showing the general hardware structure of the computer 100 that constitutes the information processing system 10. For example, the computer 100 includes a processor (such as a CPU) 101 that executes an operating system, application programs, etc., a main storage unit 102 composed of a ROM and a RAM, an auxiliary storage unit 103 composed of a hard disk, a flash memory, etc., a communication control unit 104 composed of a network card or a wireless communication module, an input device 105 such as a keyboard and a mouse, and an output device 106 such as a monitor.

[0030] Each functional component of the information processing system 10 is realized by reading a preset program into the processor 101 or the main storage unit 102 and causing the processor 101 to execute the program. According to the program, the processor 101 causes the communication control unit 104, the input device 105, or the output device 106 to operate, and reads and writes data in the main storage unit 102 or the auxiliary storage unit 103. The data or database required for processing is stored in the main storage unit 102 or the auxiliary storage unit 103.

[0031] Figure 2 FIG. is an example showing the functional structure of the information processing system 10. The information processing system 10 includes an acquisition unit 11, a calculation unit 12, and a prediction unit 13 as functional components.

[0032] The acquisition unit 11 is a functional component that acquires data related to a plurality of component objects. Specifically, the acquisition unit 11 acquires a numerical representation and a compound ratio for each of the plurality of component objects. The numerical representation of a component object refers to data that uses a plurality of numerical values to represent any attribute of the component object. The attribute of a component object refers to the property or characteristic possessed by the component object. The numerical representation can be visualized by various methods. For example, it can be visualized by methods such as numbers, English characters, text, molecular graphs, vectors, images, time-series data, etc., or can be visualized by a combination of any two or more of these methods. Each numerical value constituting the numerical representation can be represented in decimal, or can be represented in other notations such as binary or hexadecimal. The compound ratio of component objects refers to the ratio between a plurality of component objects. The specific types, units, and representation methods of the compound ratio are not limited and can be arbitrarily set according to the component object or the composite object. For example, the compound ratio can be represented by a ratio such as a percentage, or a histogram, or can be represented by the absolute amount of each component object.

[0033] The calculation unit 12 is a functional component that performs machine learning and the application of the plurality of composite ratios based on a plurality of numerical representations and a plurality of composite ratios corresponding to a plurality of component objects, thereby calculating a composite feature vector. The composite feature vector is a vector that represents the features of a composite object. The features of a composite object are any elements that make the composite object different from other objects. A vector is an n-dimensional quantity with n numerical values and can be represented as a one-dimensional arrangement. In one example, during the process of calculating the composite feature vector, the calculation unit 12 calculates the feature vectors of the respective plurality of component objects. The feature vector is a vector that represents the features of a component object. The features of a component object are any elements that make the component object different from other objects.

[0034] The calculation unit 12 includes an embedding unit 121, an interaction unit 122, an aggregation unit 123, and a ratio application unit 124. The embedding unit 121 is a functional component that uses machine learning to generate a set of other vectors of the same number from a set of vectors. In one example, the embedding unit 121 generates feature vectors from unstructured data. Unstructured data refers to data that cannot be represented by a fixed-length vector. The interaction unit 122 is a functional component that uses machine learning or other methods to generate a set of other vectors of the same number from a set of vectors. In one example, the interaction unit 122 can accept the input of feature vectors that have been obtained through machine learning. The aggregation unit 123 is a functional component that uses machine learning or other methods to aggregate a set of vectors (a plurality of vectors) into one vector. The ratio application unit 124 is a functional component that applies the composite ratio.

[0035] The prediction unit 13 is a functional component that predicts the characteristics of a composite object and outputs its predicted value. The characteristics of a composite object refer to the unique properties that the composite object has.

[0036] In one example, each of at least one machine learning model used in this embodiment is a learned model that expects the highest inference accuracy, and thus can be called the "optimal machine learning model". However, it should be noted that this learned model is not limited to being "the best in reality". The learned model is generated by a given computer processing a plurality of combinations of teacher data including input vectors and labels. The given computer calculates an output vector by inputting the input vector into the machine learning model, and obtains the error between the predicted value obtained from the calculated output vector and the label represented by the teacher data (i.e., the difference between the inference result and the correct answer). And the computer updates the given parameters in the machine learning model based on this error. The computer generates the learned model by repeating such learning. The computer that generates the learned model is not limited. For example, it can be the information processing system 10 or other computer systems. The process of generating the learned model can be called the learning stage, and the process of using this learned model can be called the application stage.

[0037] In one example, at least a part of the machine learning model used in the present embodiment can be described by a function that does not depend on the input order. With this structure, the influence of the arrangement order of multiple vectors in machine learning can be excluded.

[0038] [Data]

[0039] As described above, each component object can be a material, and the composite object can be a multi-component substance. In this case, the numerical representation of the component object (material) may include a numerical value representing the chemical structure of the material, or may include a numerical value representing the constituent repeating unit (CRU) of the chemical structure of the material. The composite ratio can be a mixing ratio or a blending ratio. The predicted value of the property of the composite object (multi-component substance) can represent at least one of the glass transition temperature (Tg) and the elastic modulus of the multi-component substance.

[0040] [Operation of the System]

[0041] Reference Figures 3 to 6 , the operation of the information processing system 10 will be described, and the information processing method according to the present embodiment will be described. Figure 3 FIG. is a flowchart showing an example of the operation of the information processing system 10 as a processing flow S1. Figure 4 FIG. is a diagram showing an example of the order of calculating the composite feature vector. Figure 5 FIG. is a diagram showing an example of applying the composite ratio in the middle of machine learning. Figure 6 FIG. is a diagram showing a specific example of the order of calculating the composite feature vector, corresponding to Figure 4 corresponds.

[0042] As Figure 3 shown, in step S11, the acquisition unit 11 acquires the numerical representation and the composite ratio for each of the plurality of component objects. As an example, if information related to two component objects Ea and Eb is input, the acquisition unit 11 acquires, for example, the numerical representation {1, 1, 2, 3, 4, 3, 3, 5, 6, 7, 5, 4} of the component object Ea, the numerical representation {1, 1, 5, 6, 4, 3, 3, 5, 1, 7, 0, 0} of the component object Eb, and the composite ratio {0.7, 0.3} of the component objects Ea and Eb. In this example, each numerical representation is represented by a vector. The composite ratio {0.7, 0.3} means that the component objects Ea and Eb are used in a ratio of 7:3 to obtain the composite object.

[0043] The acquisition unit 11 can acquire the data of each of the plurality of component objects by any method. For example, the acquisition unit 11 can read the data by accessing a given database, can receive the data from another computer or computer system, or can accept the data input by the user of the information processing system 10. Alternatively, the acquisition unit 11 can also acquire the data by any two or more of these methods.

[0044] In step S12, the calculation unit 12 calculates a composite feature vector based on a plurality of numerical representations and a plurality of composite ratios corresponding to a plurality of component objects. In this calculation, the calculation unit 12 performs machine learning and application of the composite ratio at least once each. The order of calculating the composite feature vector is not limited, and various methods can be adopted.

[0045] Reference Figure 4 , an example of the details of step S12 will be described. In this example, the calculation unit 12 calculates a composite feature vector a based on a plurality of numerical representations X and a plurality of composite ratios R corresponding to a plurality of component objects.

[0046] In step S121, the embedding unit 121 performs machine learning using an embedding function for calculating vector features, and calculates a feature vector Z for a plurality of component objects based on the numerical representation X. In the embedding function, the input vector (in this example, the numerical representation X) and the output vector (in this example, the feature vector Z) are in a one-to-one relationship. The embedding unit 121 calculates the feature vector Z of each of the plurality of component objects by inputting the plurality of numerical representations X corresponding to the plurality of component objects into the machine learning model using the embedding function. In one example, for each of the plurality of component objects, the embedding unit 121 inputs the numerical representation X corresponding to the component object into the machine learning model using the embedding function, thereby calculating the feature vector Z of the component object. The feature vector Z is a vector representing the features of the component object. In one example, the machine learning model using the embedding function can generate the feature vector Z, which is a fixed-length vector, from the numerical representation X, which is unstructured data. The machine learning model is not limited, and can be determined according to any policy considering factors such as the types of component objects and composite objects. For example, the embedding unit 121 can use a graph neural network (GNN), a convolutional neural network (CNN), or a recurrent neural network (RNN) to perform the machine learning using the embedding function.

[0047] In step S122, the ratio application unit 124 performs the application of the composite ratio R in association with the embedding function (more specifically, the machine learning model using the embedding function). The timing of applying the composite ratio R is not limited. For example, the ratio application unit 124 can apply the composite ratio R to the numerical representation X, so step S122 can be performed before step S121. Alternatively, the ratio application unit 124 can apply the composite ratio R to the feature vector Z, so step S122 can be performed after step S121. Alternatively, the ratio application unit 124 can also apply the composite ratio R to the output data of a certain intermediate layer (i.e., the intermediate result in machine learning) during the machine learning using the embedding function, so step S122 can also be a part of step S121.

[0048] In the present invention, "applying a composite ratio in association with a certain function" means applying the composite ratio to at least one of the input data of the function, the output data of the function, or the intermediate result (intermediate data) of the function. "Applying a composite ratio in association with a certain machine learning model" means applying the composite ratio to at least one of the input data (input vector) to the machine learning model, the output data (output vector) from the machine learning model, or the intermediate result (output data in a certain intermediate layer) in the machine learning model. In the present invention, "applying a composite ratio to (a certain object data)" means processing to change the object data according to the composite ratio. The method of applying the composite ratio is not limited. For example, the ratio application unit 124 can perform its application by connecting the composite ratio as an additional component to the object data. Alternatively, the ratio application unit 124 can also perform its application by multiplying or adding the composite ratio to each component of the object data. The ratio application unit 124 can perform the application of the composite ratio through these simple operations.

[0049] Reference Figure 5 , an example of the process of applying a composite ratio in the middle of machine learning having a multi-layer structure will be described. In this example, the ratio application unit 124 applies a ratio to each output value from the Nth layer of the machine learning model. The output value to which the ratio has been applied is processed as the input value of the (N + 1)th layer. For each of the Nth layer and the (N + 1)th layer, each node represents a corresponding component object. Assume that the output values of the nodes in the Nth layer are x1, x2, x3, x4, ……, and the composite ratio R of multiple component objects is expressed as "r1:r2:r3:r4 ……". The composite ratios r1, r2, r3, r4 correspond to the output values x1, x2, x3, x4 respectively. The ratio application unit 124 obtains the output value x1′ by applying the composite ratio r1 to the output value x1, obtains the output value x2′ by applying the composite ratio r2 to the output value x2, obtains the output value x3′ by applying the composite ratio r3 to the output value x3, and obtains the output value x4′ by applying the composite ratio r4 to the output value x4. These output values x1′, x2′, x3′, x4′ are processed as the input values of the (N + 1)th layer.

[0050] As Figure 5As shown in the example, by applying multiple composite ratios to the output data of the intermediate layer of a machine learning model, the composite ratios can be appropriately applied regardless of whether the data input to the machine learning model is unstructured or structured. As an example, assume a machine learning model that includes an embedding function for transforming unstructured data into a vector of a fixed length. If an attempt is made to apply a composite ratio to unstructured data before processing by this machine learning model, the correspondence between each value of the unstructured data and each value of the composite ratio is not obvious, making it difficult to apply the composite ratio. In addition, the embedding function may not be able to exhibit its original performance for this unstructured data. In one example, by applying multiple composite ratios to the output data of the intermediate layer of a machine learning model that includes an embedding function, it is also possible to expect the machine learning model to exhibit the desired performance.

[0051] In step S123, the interaction unit 122 machine-learns, using an interaction function for interacting multiple vectors, other feature vectors M for multiple component objects based on the feature vectors Z. In the interaction function, the input vector (in this example, the feature vector Z) and the output vector (in this example, the other feature vector M) are in a one-to-one relationship. In one example, the interaction unit 122 machine-learns other feature vectors M for each of the multiple component objects by inputting a set of multiple feature vectors Z corresponding to the multiple component objects into the machine learning model for the interaction function. The machine learning model is not limited, and can be determined according to any policy considering factors such as the types of component objects and composite objects. For example, the interaction unit 122 can use a convolutional neural network (CNN) or a recurrent neural network (RNN) to perform the machine learning for the interaction function. In other examples, the interaction unit 122 can calculate the feature vector M using an interaction function that does not use machine learning.

[0052] In step S124, the ratio application unit 124 performs the application of the composite ratio R in association with the interaction function (more specifically, the machine learning model for the interaction function). The timing of applying the composite ratio R is not limited. For example, the ratio application unit 124 can apply the composite ratio R to the feature vector Z, so step S124 can be performed before step S123. Or, the ratio application unit 124 can apply the composite ratio R to the other feature vector M, so step S124 can be performed after step S123. Or, the ratio application unit 124 can also apply the composite ratio R to the output data of a certain intermediate layer (i.e., the intermediate result in machine learning) during the machine learning of the interaction function, so step S124 can also be part of step S123. As described above, the method of applying the composite ratio is not limited.

[0053] In step S125, the aggregation unit 123 aggregates multiple vectors into one vector. In one example, the aggregation unit 123 uses machine learning with an aggregation function for aggregating multiple vectors into one vector to calculate a composite feature vector a based on multiple feature vectors M. In the aggregation function, the input vectors (feature vectors M in this example) and the output vector (composite feature vector a) are in an N-to-1 relationship. In one example, the aggregation unit 123 calculates the composite feature vector a by inputting a set of multiple feature vectors M corresponding to multiple component objects into the aggregation function using a machine learning model. The machine learning model is not limited and can be determined according to any policy considering factors such as the types of component objects and composite objects. For example, the aggregation unit 123 can use a convolutional neural network (CNN) or a recurrent neural network (RNN) to perform the machine learning for the aggregation function. In other examples, the aggregation unit 123 can calculate the composite feature vector a through an aggregation function without using machine learning, or can calculate the composite feature vector a by adding multiple feature vectors M, for example.

[0054] In step S126, the ratio application unit 124 performs the application of the composite ratio R in association with the aggregation function. The timing of applying the composite ratio R is not limited. For example, the ratio application unit 124 can apply the composite ratio R to the feature vector M, so step S126 can be performed before step S125. Or, the ratio application unit 124 can also apply the composite ratio R to the output data of a certain intermediate layer (i.e., the intermediate result in the machine learning model) in the middle of the machine learning for the aggregation function, so step S126 can also be a part of step S125. As described above, the method of applying the composite ratio is not limited.

[0055] In Figure 4 the example, the feature vector Z is an example of the first feature vector. The feature vector M is an example of the second feature vector and is also an example of the second feature vector reflecting multiple composite ratios. The machine learning model for the embedding function is an example of the first machine learning model, and the machine learning model for the interaction function is an example of the second machine learning model.

[0056] Refer to Figure 6 , a specific example of step S12 will be described. In this example, as the component objects, three materials (polymers), namely polystyrene, polyacrylic acid, and polybutyl methacrylate, are represented. For each of these materials, a numerical representation X in any form is prepared. The composite ratio in this example is 0.28 for polystyrene, 0.01 for polyacrylic acid, and 0.71 for polybutyl methacrylate. The calculation unit 12 performs step S12 (more specifically, steps S121 to S126) based on these data, thereby being able to calculate the composite feature vector a representing the characteristics of the multi-component substance (polymer alloy) obtained from these three materials.

[0057] Return to Figure 3 In step S13, the calculation unit 12 outputs a composite feature vector. In the present embodiment, the calculation unit 12 outputs the composite feature vector to the prediction unit 13 for subsequent processing in the information processing system 10. However, the output method of the composite feature vector is not limited thereto and can be designed according to any policy. For example, the calculation unit 12 can store the composite feature vector in a given database, send it to other computers or computer systems, or display it on a display device.

[0058] In step S14, the prediction unit 13 calculates a predicted value of the characteristics of the composite object based on the composite feature vector. The prediction method is not limited and can be designed according to any policy. For example, the prediction unit 13 can calculate the predicted value through machine learning based on the composite feature vector. Specifically, the prediction unit 13 calculates the predicted value by inputting the composite feature vector into a given machine learning model. The machine learning model used to obtain the predicted value is not limited and can be determined according to any policy considering factors such as the type of the composite object. For example, the prediction unit 13 can also perform machine learning using any neural network for solving regression problems or classification problems. Typically, the predicted value of a regression problem is represented by a numerical value, and the predicted value of a classification problem represents a category. The prediction unit 13 can also use methods other than machine learning to calculate the predicted value.

[0059] In step S15, the prediction unit 13 outputs the predicted value. The output method of the predicted value is not limited. For example, the prediction unit 13 can store the predicted value in a given database, send it to other computers or computer systems, or display it on a display device. Alternatively, the prediction unit 13 can also output the predicted value to other functional components for subsequent processing in the information processing system 10.

[0060] As described above, the order of calculating the composite feature vector is not limited. Refer to Figure 7 for another example of its calculation order. Figure 7 is a diagram showing another example related to the details of step S12.

[0061] As Figure 7As shown in Example (a), the calculation unit 12 can calculate the composite feature vector using machine learning with an embedding function and an aggregation function, without using an interaction function with machine learning. In one example, the calculation unit 12 calculates the composite feature vector by executing steps S121, S122, and S125. In step S121, the embedding unit 121 calculates the feature vector Z for multiple component objects based on the numerical representation X using machine learning with an embedding function. In step S122, the ratio application unit 124 performs the application of the composite ratio R in association with the machine learning model using the embedding function. As described above, the timing of applying the composite ratio R is not limited. In step S125, the aggregation unit 123 calculates the composite feature vector a based on the multiple feature vectors Z reflecting the multiple composite ratios R. In one example, the aggregation unit 123 calculates the composite feature vector a by inputting the set of the multiple feature vectors Z into the machine learning model using the aggregation function. Alternatively, the aggregation unit 123 can calculate the composite feature vector a by inputting the set of the multiple feature vectors Z into an aggregation function that does not use machine learning.

[0062] In Figure 7 Example (a), since it includes machine learning with an embedding function, even when the numerical representation X is non-standard data that is not represented by a fixed-length vector, it is possible to generate the feature vector Z as a fixed-length vector based on the numerical representation X. Such processing is called representation learning. Through this representation learning, it is possible to reduce the domain knowledge required to construct a machine learning model (learned model), and it is also possible to improve the prediction accuracy.

[0063] As Figure 7 shown in Example (b), the calculation unit 12 can calculate the composite feature vector using machine learning with an interaction function and an aggregation function, without using an embedding function with machine learning. In one example, the calculation unit 12 calculates the composite feature vector by executing steps S123, S124, and S125. In step S123, the interaction unit 122 calculates the feature vector M for multiple component objects based on the numerical representation X using machine learning with an interaction function. In step S124, the ratio application unit 124 performs the application of the composite ratio R in association with the machine learning model using the interaction function. As described above, the timing of applying the composite ratio R is not limited. In step S125, the aggregation unit 123 calculates the composite feature vector a based on the multiple feature vectors M reflecting the multiple composite ratios R. In one example, the aggregation unit 123 calculates the composite feature vector a by inputting the set of the multiple feature vectors M into the machine learning model using the aggregation function. Alternatively, the aggregation unit 123 can calculate the composite feature vector a by inputting the set of the multiple feature vectors M into an aggregation function that does not use machine learning.

[0064] In Figure 7In example (b), since the interaction function is included in machine learning, the non-linear response caused by the combined change of the numerical representation X can be learned with high precision.

[0065] As Figure 7 shown in example (c), the calculation unit 12 can calculate the composite feature vector by machine learning using an aggregation function, without using machine learning using an embedding function and machine learning using an interaction function. In one example, the calculation unit 12 calculates the composite feature vector by executing steps S125 and S126. In step S125, the aggregation unit 123 calculates the composite feature vector a by inputting a set of multiple numerical representations X corresponding to multiple component objects into the machine learning model using the aggregation function. In step S126, the ratio application unit 124 performs the application of the composite ratio R in association with the machine learning model using the aggregation function. As described above, the timing of applying the composite ratio R is not limited.

[0066] In Figure 7 example (c), the processing order for calculating the composite feature vector is simple, so the calculation load can be reduced.

[0067] Thus, regarding the order of obtaining the composite feature vector a from multiple numerical representations X, various methods can be considered. In any case, the calculation unit 12 calculates the composite feature vector by performing machine learning and the application of the composite ratio at least once each.

[0068] At least one of the embedding unit 121 and the interaction unit 122, the aggregation unit 123, and the ratio application unit 124 can be constructed by a single neural network. That is, the calculation unit 12 can be constructed by a single neural network. In other words, the embedding unit 121, the interaction unit 122, the aggregation unit 123, and the ratio application unit 124 are all part of this single neural network. In the case of using such a single neural network, as Figure 5 shown, the ratio application unit 124 applies the ratio in the intermediate layer.

[0069] [Program]

[0070] An information processing program for causing a computer or a computer system to function as an information processing system 10 includes program codes for causing the computer system to function as an acquisition unit 11, a calculation unit 12 (an embedding unit 121, an interaction unit 122, an aggregation unit 123, and a ratio application unit 124), and a prediction unit 13. The information processing program can be provided on the basis of being non-temporarily recorded on a tangible recording medium such as a CD-ROM, a DVD-ROM, or a semiconductor memory. Alternatively, the information processing program can also be provided via a communication network as a data signal superimposed on a carrier wave. The provided information processing program is stored, for example, in an auxiliary storage unit 103. The processor 101 reads and executes the information processing program from the auxiliary storage unit 103, thereby implementing the above-described respective functional elements.

[0071] [Effect]

[0072] As described above, an information processing system according to an aspect of the present invention includes at least one processor. The at least one processor performs the following processing: acquiring numerical representations and compound ratios for respective ones of a plurality of component objects; performing machine learning and application of the plurality of compound ratios based on the plurality of numerical representations and the plurality of compound ratios corresponding to the plurality of component objects, thereby calculating a composite feature vector representing features of a composite object obtained by compounding the plurality of component objects; and outputting the composite feature vector.

[0073] An information processing method according to an aspect of the present invention is executed by an information processing system including at least one processor. The information processing method includes the following steps: acquiring numerical representations and compound ratios for respective ones of a plurality of component objects; performing machine learning and application of the plurality of compound ratios based on the plurality of numerical representations and the plurality of compound ratios corresponding to the plurality of component objects, thereby calculating a composite feature vector representing features of a composite object obtained by compounding the plurality of component objects; and outputting the composite feature vector.

[0074] An information processing program according to an aspect of the present invention causes a computer to execute the following steps: acquiring numerical representations and compound ratios for respective ones of a plurality of component objects; performing machine learning and application of the plurality of compound ratios based on the plurality of numerical representations and the plurality of compound ratios corresponding to the plurality of component objects, thereby calculating a composite feature vector representing features of a composite object obtained by compounding the plurality of component objects; and outputting the composite feature vector.

[0075] In such an aspect, since machine learning and application of the compound ratio are performed for each component object, it is possible to improve the analysis accuracy of the composite object even when it is not possible to prepare a sufficient amount of data for the component object.

[0076] In an information processing system according to another aspect, at least one processor can calculate the feature vector of each of a plurality of component objects by inputting a plurality of numerical representations into a machine learning model, perform the application of a plurality of composite ratios in association with the machine learning model, and calculate a composite feature vector by inputting the plurality of feature vectors reflecting the plurality of composite ratios into an aggregation function. Through this series of operations, even when it is not possible to prepare a sufficient amount of data for the component objects, it is possible to improve the analysis accuracy of the composite object.

[0077] In an information processing system according to another aspect, at least one processor can calculate the first feature vector of each of a plurality of component objects by inputting a plurality of numerical representations into a first machine learning model, calculate the second feature vector of each of the plurality of component objects by inputting the plurality of first feature vectors into a second machine learning model, perform the application of a plurality of composite ratios in association with at least one machine learning model selected from the first machine learning model and the second machine learning model, and calculate a composite feature vector by inputting the plurality of second feature vectors reflecting the plurality of composite ratios into an aggregation function. By performing machine learning in two stages in this way, even when it is not possible to prepare a sufficient amount of data for the component objects, it is possible to further improve the analysis accuracy of the composite object.

[0078] In an information processing system according to another aspect, the first machine learning model can be a machine learning model that generates a first feature vector, which is a fixed-length vector, from a numerical representation that is non-fixed-form data. By using this first machine learning model, it is possible to obtain a composite feature vector from a numerical representation that cannot be represented by a fixed-length vector.

[0079] In an information processing system according to another aspect, the application of a plurality of composite ratios in association with the machine learning model can also be to apply the plurality of composite ratios to the output data of an intermediate layer of the machine learning model. By setting the timing of applying the composite ratios in this way, it is possible to appropriately apply the composite ratios regardless of whether the data input to the machine learning model is non-fixed-form or fixed-form.

[0080] In an information processing system according to another aspect, at least one processor can calculate a predicted value of the characteristics of the composite object by inputting the composite feature vector into another machine learning model, and output the predicted value. Through this process, it is possible to calculate the characteristics of the composite object with high accuracy.

[0081] In an information processing system according to another aspect, the component object can be a material, and the composite object can be a multi-component substance. In this case, even when it is not possible to prepare a sufficient amount of data for the material, it is possible to improve the analysis accuracy of the multi-component substance.

[0082] In an information processing system according to another approach, the material may be a polymer, and the multi-component substance may be a polymer alloy. In this case, even when it is not possible to prepare a sufficient amount of data for the polymer, the analysis accuracy of the polymer alloy can be improved. Polymer alloys are very diverse, and correspondingly, the types of polymers are also numerous. For such polymers and polymer alloys, generally only a part of the desirable combinations can be experimented, so in most cases, a sufficient amount of data cannot be obtained. According to this approach, even in such a situation where the data is insufficient, the polymer alloy can be analyzed with high precision.

[0083] [Modification Example]

[0084] As described above, the present invention has been described in detail according to its embodiments. However, the present invention is not limited to the above-described embodiments. The present invention can be variously modified without departing from its gist.

[0085] In the above embodiment, the information processing system 10 includes a prediction unit 13, but this functional component can be omitted. That is, the process of predicting the characteristics of the composite object can also be executed by a computer system different from the information processing system.

[0086] The processing order of the information processing method executed by at least one processor is not limited to the examples in the above embodiment. For example, a part of the above steps (processes) can be omitted, and the steps can be executed in other orders. Also, any two or more of the above steps can be combined, and a part of the steps can be modified or deleted. Or, other steps can be executed in addition to the above steps. For example, the processes of steps S14 and S15 can also be omitted. In Figure 4 In the step S12 shown, any one or two of the steps S122, S124, and S126 can also be omitted.

[0087] When comparing the magnitude relationship between two numerical values within the information processing system, either of the two criteria of "above" and "greater than" can be used, or either of the two criteria of "below" and "less than" can be used. The selection of such criteria does not change the technical meaning of the process of comparing the magnitude relationship between two numerical values.

[0088] In the present invention, the expression "at least one processor executes the first process, executes the second process,..., executes the nth process." or a corresponding expression represents the concept including the case where the execution subject (i.e., the processor) of the n processes from the first process to the nth process changes midway. That is, this expression represents the concept including both the case where all of the n processes are executed by the same processor and the case where the processor changes arbitrarily in the n processes.

[0089] Symbol Explanation

[0090] 10 - Information processing system, 11 - Acquisition unit, 12 - Calculation unit, 13 - Prediction unit, 121 - Embedding unit, 122 - Interaction unit, 123 - Aggregation unit, 124 - Ratio application unit.

Claims

1. An information processing system, comprising at least one processor, The at least one processor performs the following processing: Obtain a numerical representation and a mixing ratio for each of a plurality of polymers, where the numerical representation is unstructured data including the constituent repeating units of the chemical structure of the polymer; Input the plurality of numerical representations corresponding to the plurality of polymers into a first machine learning model, multiply the output data of the intermediate layer of the first machine learning model by the plurality of mixing ratios, and calculate, for each of the plurality of polymers, a first feature vector that is a fixed-length vector representing the characteristics of the polymer; Input the plurality of first feature vectors corresponding to the plurality of polymers into a second machine learning model, multiply the output data of the intermediate layer of the second machine learning model by the plurality of mixing ratios, and calculate, for each of the plurality of polymers, a second feature vector representing the characteristics of the polymer; Input the plurality of second feature vectors corresponding to the plurality of polymers into an aggregation function, and calculate a composite feature vector representing the characteristics of a polymer alloy obtained by compounding the plurality of polymers, Output Input the composite feature vector into another machine learning model and calculate a predicted value of the characteristics of the polymer alloy.

2. An information processing method, executed by an information processing system comprising at least one processor, the information processing method comprising the following steps: Obtain numerical representations and mixing ratios for multiple polymers respectively, where The numerical representation is unstructured data including the constituent repeating units of the chemical structure of the polymer; Input the plurality of numerical representations corresponding to the plurality of polymers into a first machine learning model, multiply the output data of the intermediate layer of the first machine learning model by the plurality of mixing ratios, and calculate, for each of the plurality of polymers, a first feature vector that is a fixed-length vector representing the characteristics of the polymer; Input the plurality of first feature vectors corresponding to the plurality of polymers into a second machine learning model, multiply the output data of the intermediate layer of the second machine learning model by the plurality of mixing ratios, and calculate, for each of the plurality of polymers, a second feature vector representing the characteristics of the polymer; and Input the plurality of second feature vectors corresponding to the plurality of polymers into an aggregation function, and calculate a composite feature vector representing the characteristics of a polymer alloy obtained by compounding the plurality of polymers, Output Input the composite feature vector into another machine learning model and calculate a predicted value of the characteristics of the polymer alloy.

3. An information processing program product, causing a computer to execute the following steps: Obtain the numerical representation and mixing ratio for each of multiple polymers, where The numerical representation is unstructured data including the constituent repeating units of the chemical structure of the polymer; Input the plurality of numerical representations corresponding to the plurality of polymers into a first machine learning model, multiply the output data of the intermediate layer of the first machine learning model by the plurality of mixing ratios, and calculate, for each of the plurality of polymers, a first feature vector that is a fixed-length vector representing the characteristics of the polymer; Input the multiple first feature vectors corresponding to the multiple polymers into a second machine learning model, multiply the output data of the intermediate layer of the second machine learning model by the multiple mixing ratios, and calculate, for each of the multiple polymers, a second feature vector representing the characteristics of the polymer; and Input the multiple second feature vectors corresponding to the multiple polymers into an aggregation function, and calculate a composite feature vector representing the characteristics of a polymer alloy obtained by compounding the multiple polymers, Output the prediction value of the characteristics of the polymer alloy by inputting the composite feature vector into another machine learning model.

Images