Information processing apparatus, information processing method, and information processing system
By optimizing the autoencoder model and interpolation values, the problem of difficulty in obtaining the overall trend caused by missing data was solved, and the accuracy and yield of data recovery were improved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-07-22
- Publication Date
- 2026-06-26
AI Technical Summary
When missing values exist in the data, existing technologies struggle to properly capture the overall trend of the data, affecting the stability of the yield rate.
By using an autoencoder model for training and optimizing the imputed values, and repeatedly updating the imputed values to recover missing data, combined with Gaussian layers and other machine learning methods, data optimization and tendency inference are achieved.
Even in the case of missing data, it is possible to appropriately infer the overall trend of the data, thereby improving the stability of the yield and the accuracy of data recovery.
Smart Images
Figure CN122286289A_ABST
Abstract
Description
[0001] This application enjoys priority based on Japanese Patent Application No. 2024-229457 (filed on December 25, 2024). This application incorporates the entire contents of the basic application by reference to that basic application. Technical Field
[0002] The embodiments of the present invention relate to information processing apparatus, information processing method, and information processing system. Background Technology
[0003] In manufacturing and selling products, maintaining a high and stable yield rate is crucial. Observing the product as a whole and appropriately obtaining information about the locations and areas where defects occur often leads to a more stable yield rate. For example, when manufacturing multiple components in a concentrated manner, determining under what conditions each component was appropriately manufactured when viewed as a whole can help maintain a high final yield rate.
[0004] In cases where a whole comprises multiple components, the overall trend is sometimes obtained during inspection using sensors or images. However, due to the performance of the sensors, gaps may sometimes appear in the images. Interpolating these missing values may affect the overall trend, so it is preferable to appropriately interpolate these missing values. Summary of the Invention
[0005] The present invention provides an information processing apparatus, information processing method and information processing system that can appropriately obtain the overall tendency even in the presence of deficiencies.
[0006] The information processing apparatus of the embodiment includes a processing circuit. The processing circuit performs the following processing: interpolating missing values of data that cause the first data to be missing using a first interpolation value to obtain second data; training an autoencoder model that outputs the first data when the second data is input; interpolating missing values of data that cause the third data to be missing using multiple second interpolation values to obtain multiple fourth data; calculating the error between the third data and the output data obtained by inputting the multiple fourth data into the autoencoder model; updating the first interpolation value based on the error using interpolation values extracted from the multiple second interpolation values; and repeatedly performing the processing from the acquisition of the second data using the updated first interpolation value to optimize the first interpolation value and the autoencoder model. Attached Figure Description
[0007] Figure 1 This is a diagram illustrating an example of the processing flow of one implementation method.
[0008] Figure 2This is a diagram illustrating an example of missing data generation in one implementation method.
[0009] Figure 3 This is a diagram illustrating an example of interpolation data generation in one implementation method.
[0010] Figure 4 This is a conceptual diagram illustrating an example of an autoencoder model according to one implementation method.
[0011] Figure 5 This is a conceptual diagram representing an example of learning a model of a presumed state of an implementation.
[0012] Figure 6 This is a conceptual diagram representing an example of learning a model of a presumed state of an implementation.
[0013] Figure 7 This is a conceptual diagram representing an example of learning a model of a presumed state of an implementation.
[0014] Figure 8 This is a conceptual diagram representing an example of learning a model of a presumed state of an implementation.
[0015] Figure 9 This is a conceptual diagram representing an example of learning a model of a presumed state of an implementation. Detailed Implementation
[0016] Hereinafter, embodiments will be described with reference to the accompanying drawings. The information processing apparatus, information processing method, and information processing system of the present invention perform the following information processing: when data exhibits a certain tendency when viewed as a whole, even if missing values exist in the data, the tendency is appropriately inferred based on the data containing the missing values.
[0017] The data can be, for example, image data, sound data, other binary data, or text data. As an example without limitation, the methods involved in semiconductor devices are sometimes used, but it should be noted that the methods of the present invention can also be applied to various other fields.
[0018] One of the information processing devices in this invention operates as an estimation device that infers the overall tendency involved in data with missing values. Another of the information processing devices in this invention operates as a learning device that optimizes the model used for estimation within the aforementioned information processing device.
[0019] Furthermore, one of the information processing devices in this invention can also be an information processing device that implements the actions of both the learning device and the estimation device described above. Additionally, the aforementioned information processing can also be implemented not through a single information processing device, but through an information processing system in which multiple information processing devices operate in a timed cooperative manner at least in a certain part.
[0020] (First Implementation)
[0021] First, one embodiment of the information processing device as a training device will be described. The information processing device includes, for example, a storage circuit and a processing circuit. The processing circuit is configured to perform training. The storage circuit stores the necessary data.
[0022] The processing circuit can be a dedicated digital or analog circuit such as an ASIC (Application Specific Integrated Circuit) or a DSP (Digital Signal Processor), or a general-purpose processor such as a CPU (Central Processing Unit) or a GPU (Graphics Processing Unit), or a programmable circuit such as an FPGA (Field Programmable Gate Array).
[0023] The storage circuit can be located internally in the information processing device or externally, capable of accessing data through a suitable interface. The storage circuit can also store data needed during training, data during training, and data after training is complete. When the information processing device uses processing circuitry as hardware resources to specifically implement software-based information processing, the storage circuit can also store programs, executable files, or similar data used to control the software-based information processing.
[0024] As described above, they can also be configured as an information processing system consisting of several interconnected storage circuits and processing circuits. The processes performed by the information processing device can be appropriately implemented using these processing and storage circuits.
[0025] Figure 1 This is a diagram illustrating the processing flow of one embodiment. In this embodiment, the information processing device generates a model that appropriately generates data without missing data based on the missing data.
[0026] The information processing device first prepares training data sets and validation data sets (S100). These data are stored in an external storage circuit, such as a memory, and are prepared for use by the information processing device only when needed. The training data is used to learn the autoencoder model AE. The validation data is used to validate the autoencoder model AE and is also used for optimizing interpolation values.
[0027] The information processing device extracts first data from the training data for training the autoencoder model AE (S102). Subsequently, the information processing device removes the first data to generate missing data (S104).
[0028] For example, when the input data (training and validation data) is image data, the information processing device can generate data by missing a portion of the image. The information processing device can similarly generate missing data from audio data, other binary data, and text data.
[0029] The information processing device can generate missing data based on the characteristics of the sensor that acquires the input data. For example, when the input data is image data, the information processing device can generate missing data by considering the potential defects that may occur in the sensor that acquired the image, thus causing the image data to be missing.
[0030] Figure 2 This diagram illustrates an example of generating missing data for image data according to one embodiment. The information processing device, for certain images shown on the left (e.g., first data), causes values to be missing in one or more regions, thereby generating the missing data shown on the right. In the diagram on the right, the shaded areas represent the missing regions.
[0031] Return to Figure 1 The information processing device uses the first interpolation value to interpolate the missing parts of the generated missing data, generating second data (S106). In the initial iteration, a predetermined initial value can also be set for the first interpolation value. By interpolating the missing data in this way, the information processing device generates data that can be input into the autoencoder model AE.
[0032] Figure 3 This diagram illustrates an example of interpolation data generation for image data according to one embodiment. The information processing device uses a first interpolation value to interpolate missing portions of the missing data shown on the left. The first interpolation value is set, for example, between the minimum and maximum brightness values. The diagram on the right shows the data after interpolation (second data), with the shaded areas indicating the portions after interpolation using the first interpolation value.
[0033] Return to Figure 1The information processing device inputs the second data into the autoencoder model AE and obtains the output data corresponding to the second data (S108). The autoencoder model AE has an encoder ENC and a decoder DEC. It is a model that transforms the image input by the encoder ENC into a latent variable Z (latest vector), and then transforms the latent variable Z into output data through the decoder DEC.
[0034] The information processing device compares the output data of the second data input to the autoencoder model AE with the first data, and learns in a way that the autoencoder model AE becomes a model that outputs the first data when the second data is input (S112). This learning can also be performed based on general autoencoder methods.
[0035] The information processing device continues learning the autoencoder model AE until optimization is complete. During learning, the information processing device can extract multiple first data points from the training data as needed and perform optimization. Furthermore, during learning, the information processing device can also generate multiple missing data points based on the same first data point and use first imputation values to generate second data points for these missing data points.
[0036] The learning termination condition is set to the appropriate execution of the autoencoder model AE. Termination conditions can be set as follows: a specified number of iterations, a specified amount of computation time, or the evaluation function falling below a specified threshold.
[0037] Prior to this learning, as another method, the information processing device may first train the autoencoder model AE to learn by outputting the first data input to the first data without using missing data, and then use the learned parameters to perform learning using the second data. That is, the information processing device can use an autoencoder model AE that has been pre-learned to output the first data when the first data is input. In this case, methods such as network distillation and dropout (random deactivation) of the autoencoder model AE can also be used to perform the processing in S112.
[0038] After the learning process using the training data in the autoencoder model AE is temporarily completed, the information processing unit moves on to processing using the validation data. The information processing unit extracts the third data from the validation data (S112). Next, the information processing unit generates the missing data for the fourth data in the same way as the first data (S114).
[0039] Next, the information processing device generates multiple fourth data using multiple second interpolation values for the missing data in the acquired third data (S116). The second interpolation values may or may not include the first interpolation value, but since it is necessary to compare other second interpolation values with the first interpolation value, it is preferable to include the first interpolation value.
[0040] The information processing device, for example, can randomly extract multiple values from the permissible range of interpolated values and set them as multiple second interpolated values. Alternatively, the information processing device can, for example, extract multiple values at equal intervals within the range of interpolated values and set them as multiple second interpolated values.
[0041] The information processing device inputs multiple fourth data into the auto encoder model AE to obtain output data (S118). Then, the information processing device updates the first interpolation value by comparing the multiple output data with the third data (S120).
[0042] The information processing device may also acquire the error (residual) between multiple output data and the third data, extract the second interpolation value corresponding to the output data with the smallest error, and update the first interpolation value using the extracted second interpolation value.
[0043] After updating the first interpolation value, the information processing device can repeatedly execute the processing from S102 to S120 until the termination condition (S122) is met. That is, the information processing device uses the new first interpolation value to learn the autoencoder model AE, and updates the first interpolation value using the second interpolation value for the learned autoencoder model AE.
[0044] In this way, the information processing device alternately and repeatedly updates the autoencoder model AE and the first interpolated value to fill in the missing values. Ultimately, the information processing device can, for example, repeatedly perform these processes until the error between the output data after inputting the fourth data into the autoencoder model AE and the third data (e.g., it can be either the minimum value or the average value) is lower than a predetermined threshold, until the first interpolated value is updated a predetermined number of times, and until conditions suitable for completing the learning are met.
[0045] Furthermore, the selection of the second interpolation value in the processing of S116 by the information processing device can be done not only randomly and uniformly, but also using other selection methods. For example, it can also be selected according to a normal distribution centered on the current first interpolation value.
[0046] This selection can also be changed based on the number of times S120 is executed. For example, the information processing device can also control parameters such as the standard deviation of a normal distribution according to the number of times S120 is executed.
[0047] As another example, the information processing device can change the method for selecting the second interpolation value in the next iteration based on the error between the output data and the third data. For example, as seen in methods such as simulated annealing, the information processing device can select the next second interpolation value based on the error to avoid falling into a local optimum.
[0048] In the processing of S120, the information processing device can extract the second interpolation value with the smallest error between the output data and the third data, and set the second interpolation value as the first interpolation value for the next iteration. Alternatively, the information processing device can calculate the first interpolation value for the next iteration based on the error between the output data and the third data, using the second interpolation value (and the first interpolation value).
[0049] The information processing device may, for example, update the first interpolation value based on various statistical measures such as the error between the output data after the fourth data is input into the autoencoder model AE and the third data, and the average value obtained by weighting each second interpolation value.
[0050] As described above, according to this embodiment, by alternately and repeatedly updating the autoencoder model and the first interpolation value, optimization of the model for inferring the original data based on the missing data can be achieved. By using the recovered original data, the appropriate overall data tendency can be determined based on the missing data. The overall data tendency can be determined based on rules or using a learned model that outputs the overall data tendency.
[0051] (Second Implementation)
[0052] In the first embodiment described above, an example of using an autoencoder model to recover missing data was explained. In this embodiment, a method for improving the accuracy of the recovery based on this autoencoder model is further explained.
[0053] Figure 4 This is a diagram illustrating an example of an autoencoder model AE, representing a conceptual implementation. The autoencoder model AE, for example, includes a Gaussian layer GL in addition to the encoder ENC and decoder DEC.
[0054] The Gaussian layer (GL) is placed before the input layer of the encoder ENC, adding Gaussian noise to the data input to the encoder ENC. The information processing device then... Figure 1 When the second and fourth data are input into the auto encoder model AE, noise is first applied to the data through the Gaussian layer GL, and the noise-treated data is then input into the encoder ENC.
[0055] Instead of continuously applying a predetermined noise to the data, Gaussian layers (GL) can add noise to the data, for example, by generating noise according to the distribution of the parameters of the noise superposition.
[0056] Information processing devices can replace Figure 1 The autoencoder model AE is used Figure 4 The autoencoder model AE is used to perform the above-mentioned autoencoder model AE and the optimization of the first interpolation value.
[0057] As described above, in this embodiment, a Gaussian layer GL can be added to the encoder ENC data input to the autoencoder model AE to add noise. By having this Gaussian layer GL, it is possible to learn a more robust model that takes into account the biases caused by noise in the input data acquired by the sensor, individual differences in the sensors, etc.
[0058] (Third Implementation)
[0059] In the foregoing embodiments, models for recovering non-missing data from missing data have been described. By using these models, non-missing data can be obtained, and thus the overall tendency of the data can be obtained through various methods, such as generally known rule-based methods and methods using models learned through machine learning.
[0060] In this embodiment, an example that uses the overall tendency of the data from the autoencoder model AE described in the foregoing embodiments is described. Furthermore, the following description uses a Gaussian layer GL as described in the second embodiment, but of course, no other method will be used. Figure 1 The same processing can be achieved in the case of the Gaussian layer GL shown, and it can achieve the same effect as the first embodiment.
[0061] Figure 5 This diagram conceptually illustrates an example of the structure of a model for acquiring state Y (representing state) from data after data X (representing characteristics) has been missing, and the learning of this model. The autoencoder model AE can also be formed through the aforementioned embodiments.
[0062] The information processing device performs the same processing as in the aforementioned embodiment to cause data X to be missing, performs interpolation, and inputs the interpolated data into the automatic encoder model AE to obtain the estimated recovered data f(X) of data X (S200). As the interpolation value, the first interpolation value optimized by the above embodiment can be used.
[0063] The information processing device optimizes the first model so that it outputs state Y (S202) when the acquired estimated recovery data f(X) and data X are input. The state Y corresponding to data X can generate training data based on data such as experimental values and theoretical values.
[0064] The first model could also be a linear regression model, for example. In this case, the information processing device can optimize the model by setting the state Y as the target variable for the data X, which is the explanatory variable. The linear regression method can be any method.
[0065] Alternatively, as another example without limitation, the first model can also be other models such as MLP (Multi-Layer Perceptron), CNN (Convolutional Neural Network), GNN (Graph Neural Network). In these examples, the information processing device can perform model learning in any way suitable for each model.
[0066] The first model after this learning is optimized to obtain the state Y1 (first state) based on the result of inputting the missing data X into the autoencoder model AE.
[0067] As described above, according to this embodiment, by using the optimized interpolation value, the optimized autoencoder model AE, and the optimized first model, it is possible to infer the state representing the tendency of the data as a whole based on the missing data.
[0068] Furthermore, the information processing device is capable of performing the training of the autoencoder model AE and the optimization of the first interpolation value in parallel with the training of the first model.
[0069] (Fourth Implementation)
[0070] In the third embodiment, the recovered data output from the autoencoder model AE is used to predict the state, but in this embodiment, a model that uses the latent variable Z output from the encoder ENC of the autoencoder model AE to predict the state is described.
[0071] Figure 6 This is a conceptual illustration of the structure of a model for acquiring state Y from data where data X is missing, and an example of how this model learns. The autoencoder model AE can also be formed through the aforementioned embodiments.
[0072] The information processing device, similar to the aforementioned embodiments, renders data X missing, interpolates it using the optimized first interpolation value, and inputs it into the optimized autoencoder model AE. The information processing device then acquires the latent variable Z (S300) from the encoder ENC output of the autoencoder model AE.
[0073] The information processing device trains a second model to obtain the state Y from the latent variable Z (S302). The second model can also be formed, for example, by a decision tree. In the case of a decision tree, the information processing device can train the second model using machine learning methods such as XGBoost to infer the state Y based on the latent variable Z.
[0074] The data, i.e., the latent variable Z, which has undergone dimensionality compression by the encoder ENC, is restored to complete data by the decoder DEC. This complete data is data with a form from which states can be extracted. That is, the latent variable Z is a vector with low dimensionality that has state features corresponding to the input data X.
[0075] The second model is formed to extract the feature quantities of state Y from the latent variable Z and output the state Y. The information processing device is optimized to extract state Y from the latent variable Z which has the features of state Y.
[0076] The information processing device performs optimization of the second model, for example, by performing machine learning (e.g., training based on XGBoost), such that when the input is the imputed value optimized based on the missing data X and the latent variable Z obtained via the autoencoder model AE, the output is the state Y as teacher data.
[0077] The optimized second model is formed as a model that outputs state Y2 (second state) when the input latent variable Z.
[0078] As described above, according to this embodiment, by using the optimized interpolation value, the optimized autoencoder model AE, and the optimized second model, it is possible to infer the state representing the tendency of the data as a whole based on the missing data.
[0079] Furthermore, similar to the third embodiment, the information processing device can perform the training of the autoencoder model AE and the optimization of the first interpolation value in parallel with the training of the second model as a multi-task learning.
[0080] (Fifth Embodiment) In the third embodiment, the estimation of the state based on the data output from the autoencoder model AE through the first model was described. In the fourth embodiment, the estimation of the state based on the latent variables extracted from the autoencoder model AE through the second model was described. These estimations can be performed separately as described above, but it is also possible to learn to estimate the state Y using the results from both.
[0081] Figure 7 This is a conceptual diagram illustrating, as an example, the configuration of a model for acquiring state Y from data where data X is missing, and the learning process of that model. The autoencoder model AE, the first model, and the second model have the same configuration as in the aforementioned embodiments.
[0082] The information processing device uses the optimization results of the first model and the second model to perform ensemble learning and train these models (S400). As an unrestricted example, the information processing device can perform ensemble learning after optimizing the first interpolation value and the autoencoder model AE. Alternatively, as an unrestricted example, the information processing device can also train the first interpolation value, the autoencoder model AE, the first model, and the second model through ensemble learning.
[0083] This training can also be set up as a multi-task learning program, as mentioned above.
[0084] More specifically, the information processing device, for example, uses a first interpolation value learned from the learning data and an autoencoder model AE to train a first model and a second model according to the processing described in the third and fourth embodiments. The information processing device is capable of acquiring data (X, Y), i.e., input data X and state data Y, from the same or different learning data for these models, and realizing learning based on the data (X, Y).
[0085] By performing ensemble learning based on state Y1 output from the first model and state Y2 output from the second model, the information processing device is able to achieve learning that suppresses overlearning in both the first and second models.
[0086] As described above, according to this embodiment, the accuracy of the optimized interpolation value, the optimized autoencoder model AE, the optimized first model, and the optimized second model can be further improved. As a result, it is possible to estimate the state representing the tendency of the data as a whole with higher accuracy based on data with missing data.
[0087] (Sixth Implementation Method)
[0088] In addition to using the autoencoder model AE described above, it is also possible to form a model that directly obtains the state Y from the missing data. This model can be formed, for example, by a model like XGBoost that infers the state from the missing data.
[0089] Figure 8 This is a diagram conceptually illustrating one implementation of a third model that directly infers state Y based on missing data. The information processing device acquires missing data from data X and trains the third model so that it outputs state Y when the missing data is input (S500). This training is achieved by performing learning on missing data, such as XGBoost.
[0090] The third model differs from the second model in that it directly applies a model like XGBoost to missing data. It is also possible to use this third model as a weak learning machine. Therefore, in this embodiment, a further learning method for using the third model as a weak learning machine will be described.
[0091] Figure 9 This is a diagram conceptually illustrating the structure of a model for obtaining state Y due to missing data X, and an example of how this model learns, according to one implementation. The autoencoder model AE, the first model, and the second model have the same structure as in the aforementioned implementations. The third model has the same... Figure 8 The model shown has the same structure.
[0092] The information processing device first generates missing data based on data X through the same processing as in the aforementioned embodiments. The information processing device then uses a first interpolation value to generate interpolated data for the missing data.
[0093] The information processing device inputs the stored data into the autoencoder model AE to obtain the output data f(X) and the latent variable Z. The information processing device then inputs the output data f(X) into the first model to obtain state Y1. Simultaneously, the information processing device inputs the latent variable Z into the second model to obtain state Y2. Furthermore, the information processing device inputs the missing data obtained by removing data X into the third model to obtain state Y3 (the third state).
[0094] The information processing device performs integrated learning using these states Y1, Y2, and Y3, thereby improving the accuracy of each model.
[0095] As an example without limitations, an information processing device can learn a first model, a second model, and a third model through ensemble learning.
[0096] As an example without limitations, the information processing device can also learn the first interpolation value, the autoencoder model AE, the first model, the second model, and the third model through ensemble learning.
[0097] By performing ensemble learning that references the output values of multiple paths that output the same state, the accuracy of each model can be improved. Furthermore, this learning can be configured as multi-task learning as needed.
[0098] As described above, according to this embodiment, not only can the optimization of the autoencoder model AE be achieved, but also the optimization of the model that acquires the overall tendency relative to the data can be achieved. That is, by using ensemble learning of at least two of the first state, second state, and third state acquired by different means, the information processing device can generate a highly accurate model.
[0099] The information processing device can use an optimized model to form a model for restoring missing data and / or a model for inferring the state of the data as a whole, representing the tendency of the data. Within the scope of this invention, the manner in which the information processing device uses the optimized model is also included.
[0100] In addition, an information processing system is capable of forming a state where a tendency to recover missing data and / or infer a whole from missing data can be achieved through multiple information processing devices.
[0101] As an example without limitation, the optimization and optimized model in this invention can be applied to the characteristics of semiconductor devices and the state of wafers. For example, data X (characteristic) can be set as an electrical characteristic. As another example without limitation, this electrical characteristic can also be a characteristic such as leakage current, withstand voltage, or on-resistance.
[0102] For example, state Y can also represent the electronic state of a device or the state of a semiconductor device predicted based on electrical characteristics. As an example without limitation, state Y can be used to estimate impurity concentration, the fabrication shape of a device (width, depth, taper angle, etc.), and also to estimate the precision of photolithography and fabrication accuracy, which affect the shape. Furthermore, state Y can also be used as an electrical characteristic to estimate other electrical characteristics.
[0103] Missing data can be, for example, data representing characteristics below a certain threshold. For instance, it could be missing information such as the adhesion of dust particles or dirt in an image, or a voltage withstand value. Furthermore, missing data can also include defects. Defects could include situations where, if a characteristic is judged as defective, subsequent characteristics are not measured. This allows for handling missing data in images, as well as missing values related to current or voltage.
[0104] Several embodiments of the present invention have been described, but these embodiments are provided as examples and are not intended to limit the scope of the invention. These new embodiments can be implemented in various other ways, and various omissions, substitutions, and modifications can be made without departing from the spirit of the invention. These embodiments or variations thereof are included in the scope or spirit of the invention, and are included within the scope of the invention described in the patent claims and their equivalents.
Claims
1. An information processing device comprising a processing circuit, characterized in that, The processing circuit performs the following processing: The second data is obtained by imputing the missing values of the data that caused the first data to be missing using the first imputation value. An autoencoder model is trained, which outputs the first data when the second data is input. Multiple fourth data points are obtained by imputing the missing values of the data that cause the third data to be missing using multiple second imputation values. Calculate the error between the third data and the output data after inputting the plurality of fourth data into the autoencoder model. Based on the error, the first interpolated value is updated using the interpolated values extracted from the plurality of second interpolated values. The updated first interpolation value is used to repeatedly perform the processing from the acquisition of the second data to optimize the first interpolation value and the autoencoder model.
2. The information processing apparatus according to claim 1, wherein, The autoencoder model has a layer that superimposes Gaussian noise onto the input data.
3. The information processing apparatus according to claim 1, wherein, The processing circuit generates a first model, which performs linear regression on the state data corresponding to the input data with missing values and the output data after interpolating the input data using the first interpolation value and inputting it into the autoencoder model.
4. The information processing apparatus according to claim 3, wherein, The processing circuit also uses the first interpolation value to interpolate the input data with missing values and inputs it into the autoencoder model to obtain latent variables. Furthermore, a second model is trained, which obtains the state data from the latent variables based on the latent variables and the state data corresponding to the input data.
5. The information processing apparatus according to claim 4, wherein, The processing circuit interpolates the input data with missing values using the first interpolation value and inputs it into the autoencoder model and the first model to obtain a first state. Furthermore, the input data is interpolated using the first interpolation value and then input into the autoencoder model and the second model to obtain the second state. Furthermore, ensemble learning is performed based on the first state and the second state.
6. The information processing apparatus according to claim 4, wherein, The processing circuit also outputs a third model of state data when input data has missing values, and trains the third model based on the state data obtained from the input data and the state data corresponding to the input data.
7. The information processing apparatus according to claim 6, wherein, The processing circuit interpolates the input data with missing values using the first interpolation value and inputs it into the autoencoder model and the first model to obtain a first state. Furthermore, the input data is interpolated using the first interpolation value and then input into the autoencoder model and the second model to obtain the second state. Furthermore, the input data is fed into the third model to obtain the third state. Furthermore, ensemble learning is performed based on the first state, the second state, and the third state.
8. An information processing apparatus that uses a model trained by the information processing apparatus of any one of claims 1 to 7 to obtain a state corresponding to input data having missing values from input data.
9. An information processing method, characterized in that, The following processing is performed through the processing circuit: The second data is obtained by imputing the missing values of the data that caused the first data to be missing using the first imputation value. An autoencoder model is trained, which outputs the first data when the second data is input. Multiple fourth data points are obtained by imputing the missing values of the data that cause the third data to be missing using multiple second imputation values. Calculate the error between the third data and the output data after inputting the plurality of fourth data into the autoencoder model. Based on the error, the first interpolated value is updated using the interpolated values extracted from the plurality of second interpolated values. The updated first interpolation value is used to repeatedly perform the processing from the acquisition of the second data to optimize the first interpolation value and the autoencoder model.
10. An information processing system comprising: one or more storage circuits and one or more processing circuits, wherein at least one of the one or more processing circuits is used to perform the following processing: The second data is obtained by imputing the missing values of the data that caused the first data to be missing using the first imputation value. An autoencoder model is trained, which outputs the first data when the second data is input. Multiple fourth data points are obtained by imputing the missing values of the data that cause the third data to be missing using multiple second imputation values. Calculate the error between the third data and the output data after inputting the plurality of fourth data into the autoencoder model. Based on the error, the first interpolated value is updated using the interpolated values extracted from the plurality of second interpolated values. The updated first interpolation value is used to repeatedly perform the processing from the acquisition of the second data to optimize the first interpolation value and the autoencoder model.