Data processing method, device, electronic device, medium and program product

Abstract

The embodiments of the present application disclose a data processing method, apparatus, electronic device, medium, and program product for use in the field of machine learning technology. The method comprises: obtaining N types of object attribute information of a target object, performing feature extraction on the target object based on the N types of object attribute information, obtaining object features of the target object under each type of object attribute information, generating prediction indicator distributions for the target object under target indicators based on the object features corresponding to each type of object attribute information, and determining a target prediction indicator and a target accuracy index of the target prediction indicator for the target object based on the prediction indicator distributions corresponding to each type of object attribute information. By using the embodiments of the present application, the prediction efficiency and accuracy of the uncertainty of indicators and indicator results can be improved.

Description

Technical Field

[0001] The present application relates to the field of machine learning technology, and in particular to a data processing method, device, electronic device, medium, and program product. Background Art

[0002] Currently, the prediction of object indicators can be achieved through machine learning related technologies. A model for indicator prediction can be trained through machine learning, such as a trained model that can predict the stability of a substance. Among them, determining the uncertainty of the prediction result when outputting the prediction result is an important task. Existing methods for determining uncertainty usually use a model that randomly discards neurons multiple times during the prediction process to determine the uncertainty based on the multiple prediction results obtained. Since this method requires discarding some model parameters each time during prediction, this additional operation increases working time and will cause a certain degree of uncertainty, which can easily lead to low prediction efficiency and accuracy. Therefore, how to improve the prediction efficiency and accuracy of the uncertainty of indicators and indicator results has become a problem that needs to be solved urgently. Summary of the Invention

[0003] The embodiments of the present application provide a data processing method, device, electronic device, medium and program product, which can improve the prediction efficiency and accuracy of the uncertainty of indicators and indicator results.

[0004] In one aspect, an embodiment of the present application provides a data processing method, the method comprising:

[0005] Get N types of object attribute information of the target object, where N is an integer greater than 1;

[0006] Extract features of the target object based on N types of object attribute information to obtain the object features of the target object under each type of object attribute information;

[0007] Based on the object features corresponding to each object attribute information, the predicted indicator distribution of the target object under the target indicator is generated respectively; the predicted indicator distribution corresponding to any object attribute information is within the indicator range of the target object under the target indicator;

[0008] The target prediction index of the target object and the target accuracy index of the target prediction index are determined according to the prediction index distribution corresponding to each object attribute information.

[0009] In one aspect, an embodiment of the present application provides a data processing device, comprising:

[0010] An acquisition module is used to obtain N types of object attribute information of a target object, where N is an integer greater than 1;

[0011] A processing module is used to extract features of the target object based on N types of object attribute information to obtain object features of the target object under each type of object attribute information;

[0012] The processing module is further configured to generate a prediction indicator distribution of the target object under the target indicator based on the object features corresponding to each type of object attribute information; the prediction indicator distribution corresponding to any type of object attribute information is within the indicator range of the target object under the target indicator;

[0013] The determination module is used to determine the target prediction index and the target accuracy index of the target prediction index of the target object according to the prediction index distribution corresponding to each object attribute information.

[0014] On the one hand, an embodiment of the present application provides an electronic device, which includes a processor and a memory, wherein the memory is used to store a computer program, the computer program includes program instructions, and the processor is configured to call the program instructions to execute some or all of the steps in the above method.

[0015] On the one hand, an embodiment of the present application provides a computer-readable storage medium, which stores a computer program, and the computer program includes program instructions, which, when executed by a processor, are used to execute some or all of the steps in the above method.

[0016] Accordingly, according to one aspect of the present application, a computer program product or computer program is provided, the computer program product or computer program including program instructions stored in a computer-readable storage medium. A processor of a computer device reads the program instructions from the computer-readable storage medium and executes the program instructions, causing the computer device to perform the data processing method provided above.

[0017] In an embodiment of the present application, N types of object attribute information of the target object can be obtained, and features of the target object can be extracted based on the N types of object attribute information to obtain object features of the target object under each type of object attribute information; the object features are obtained by the joint action of the N types of attribute information, so that the object features can contain more feature information; based on the object features corresponding to each type of object attribute information, the prediction index distribution of the target object under the target index is generated respectively; the same resources (such as parameters involved in the prediction process) are used when predicting each prediction index distribution, thereby avoiding the prediction deviation caused by the use of different resources; the target prediction index and the target accuracy index are determined according to the prediction index distribution corresponding to each type of object attribute information; the target prediction index and the target accuracy index can be determined together through the prediction index distribution, and the uncertainty is expressed by the target accuracy index, thereby improving the prediction efficiency and accuracy of the uncertainty of the indicator and the corresponding indicator result. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following is a brief introduction to the drawings required for use in the description of the embodiments. Obviously, the drawings described below are some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0019] Figure 1 A schematic diagram of an application architecture provided in an embodiment of the present application;

[0020] Figure 2 A flowchart of a data processing method provided in an embodiment of the present application;

[0021] Figure 3 A flowchart of a data processing method provided in an embodiment of the present application;

[0022] Figure 4a A schematic diagram of a feature extraction scenario provided in an embodiment of the present application;

[0023] Figure 4b A schematic diagram of a feature extraction scenario provided in an embodiment of the present application;

[0024] Figure 4c A schematic diagram of a feature extraction scenario provided in an embodiment of the present application;

[0025] Figure 4d A schematic diagram of a feature extraction scenario provided in an embodiment of the present application;

[0026] Figure 5 A schematic diagram of a scenario for predicting a target indicator of a target object provided in an embodiment of the present application;

[0027] Figure 6 A schematic diagram of the structure of a data processing device provided in an embodiment of the present application;

[0028] Figure 7 A schematic diagram of the structure of an electronic device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0029] The technical solutions in the embodiments of the present application will be described clearly and completely below in conjunction with the drawings in the embodiments of the present application.

[0030] The data processing method proposed in the embodiment of the present application is implemented in an electronic device, which can be a server or a terminal. Among them, the server can be an independent physical server, or a server cluster or distributed system composed of multiple physical servers, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, CDN (Content Delivery Network), and big data and artificial intelligence platforms. The terminal can be a smart phone, tablet computer, laptop computer, desktop computer, smart speaker, smart watch intelligent voice interaction device, smart home appliance, car terminal, aircraft, etc., but is not limited to this.

[0031] Next, the technical terms involved in the technical fields to which the solutions of the embodiments of the present application may be applied are introduced:

[0032] 1. Artificial Intelligence

[0033] The embodiments of the present application relate to the field of artificial intelligence (AI) technology, specifically the field of machine learning (ML) technology in artificial intelligence. Machine learning (ML) is a multi-disciplinary interdisciplinary subject involving probability theory, statistics, approximation theory, convex analysis, algorithm complexity theory and other disciplines. It specializes in studying how computers simulate or implement human learning behavior to acquire new knowledge or skills, and reorganize existing knowledge structures to continuously improve their own performance. Machine learning is the core of artificial intelligence and the fundamental way to make computers intelligent. Its applications are spread across all areas of artificial intelligence. Machine learning and deep learning generally include artificial neural networks, belief networks, reinforcement learning, transfer learning, inductive learning, and formula-based learning. For example, the technical solution of this application can use machine learning to predict the target indicators of the target object and obtain the target prediction index and the target accuracy index.

[0034] In some embodiments, see Figure 1 , Figure 1 The following is a schematic diagram of an application architecture provided in an embodiment of the present application, through which the data processing method proposed in the present application can be executed. Figure 1As shown, an electronic device may be included; wherein the electronic device obtains N types of object attribute information of the target object, performs feature extraction based on the N types of object attribute information, obtains object features corresponding to each type of object attribute information, and the object features are obtained by the joint action of the N types of object attribute information, and generates a prediction index distribution corresponding to each type of object attribute information based on the object features corresponding to each type of object attribute information, so as to determine the final target prediction index of the target object and the target accuracy index of the target prediction index. Optionally, the above process can also be performed by a prediction model, that is, a prediction model can be deployed in the electronic device, which can be composed of a trained prediction network, and generates object features corresponding to each type of object attribute information and a prediction index distribution corresponding to each type of object attribute information in the trained prediction network.

[0035] It is understandable that Figure 1 The present invention only exemplifies the possible application architectures of the technical solution of the present application and does not limit the specific architecture of the technical solution of the present application. That is, the technical solution of the present application can also provide other forms of application architectures.

[0036] Optionally, in some embodiments, the electronic device may execute the data processing method according to actual business needs to improve the prediction efficiency and accuracy of the uncertainty of indicators and indicator results. The technical solution of the present application can be applied to scenarios where target indicators are predicted for any type of target object. For example, the target object may be a chemical molecule, and the target indicator may be the stability of the chemical molecule, that is, the electronic device may obtain N types of molecular attribute information of the chemical molecule, and obtain the molecular features corresponding to each type of molecular attribute information through feature extraction to generate a prediction indicator distribution corresponding to each type of molecular attribute information, and determine the predicted stability of the chemical molecule and the accuracy index of the predicted stability according to the predicted indicator distribution; for another example, the target object may also be a medical image of the person being diagnosed, and the target indicator may be the abnormality index of the person being diagnosed, that is, the electronic device may obtain N types of image attribute information of the medical image, and obtain the image features corresponding to each type of image attribute information through feature extraction to generate a prediction indicator distribution corresponding to each type of image attribute information, and determine the predicted abnormal value of the medical image and the accuracy index of the predicted abnormal value according to the predicted indicator distribution.

[0037] Optionally, the data involved in this application, such as object attribute information, target prediction indicators and accuracy index of the target object, etc., can be stored in a database, or can be stored in a blockchain, such as through a blockchain distributed system, which is not limited in this application.

[0038] It is understood that the above scenarios are merely examples and do not limit the application scenarios of the technical solutions provided in the embodiments of this application. The technical solutions of this application can also be applied to other scenarios. For example, those skilled in the art will appreciate that with the evolution of system architecture and the emergence of new business scenarios, the technical solutions provided in the embodiments of this application will also be applicable to similar technical problems.

[0039] The solutions provided in the embodiments of the present application involve technologies such as machine learning of artificial intelligence, which are specifically described through the following embodiments:

[0040] Based on the above description, the present application embodiment proposes a data processing method, which can be executed by the electronic device mentioned above. Figure 2 , Figure 2 This is a flow chart of a data processing method provided in an embodiment of the present application. Figure 2 As shown, the process of the data processing method of the embodiment of the present application may include the following:

[0041] S201: Obtain N types of object attribute information of a target object.

[0042] Wherein, N is an integer greater than 1. The target object is an object for which a target indicator needs to be predicted. The target indicator has an indicator range, which can be set by relevant business personnel according to actual application scenarios. For example, the indicator range is 0-1.

[0043] In some embodiments, the target object can be any type of object. For example, it can be a chemical molecule, specifically a small drug molecule, etc., and the target object in this case can be represented by the chemical structure corresponding to the chemical molecule and the chemical elements contained therein; for example, it can be a medical image, specifically a CT image (Computed Tomography Imaging). There is no restriction on the target object here. The target indicator can be any type of object indicator. For example, if the target object is a chemical molecule, the target indicator can be the molecular properties of the chemical molecule (such as stability), etc.; if the target object is a medical image, the target indicator can be the abnormality index of the medical image, etc. There is no restriction on the target indicator here. In addition, the present application does not limit the source of the target object. For example, it can be obtained from a relevant database or website in the field, such as from a chemical database, which stores a variety of chemical structures and chemical elements obtained by chemical technicians from various chemical molecules; or it can be input by the user, such as a medical image taken by the user and uploaded.

[0044] In some embodiments, the electronic device can obtain N object representation methods and execute these N object representation methods on the target object to obtain N types of object attribute information. This can be understood as digitally representing the target object in different ways to obtain a string that can be used to represent the target object, with this string serving as the object attribute information; or it can be understood as encoding the target object in different ways to obtain encoded data that can be used to represent the target object, with this encoded data serving as the object attribute information. The N types of object representation methods can be set by relevant business personnel based on the specific type of the target object.

[0045] Among them, different object representation methods obtain different object attribute information, and different object representation methods may have different focuses and shortcomings. When the target object is a chemical molecule, the object representation method may, for example, generate object attribute information based on chemical elements and combined with chemical structure, or based on chemical structure and combined with chemical elements. And some object representation methods, for example, for chemical molecules with relatively complex chemical structures but relatively few chemical elements, the object attribute information generated is relatively sparse, making the object attribute information unable to well represent the target object. The object attribute information at this time can also be called a molecular fingerprint, that is, molecular fingerprint extraction (Molecular Motif Extraction) is performed on the chemical molecules; thereby, molecular property prediction (Molecular Property Prediction) can be achieved by predicting target indicators. When the target object is a medical image, the object representation method may, for example, generate object attribute information based on the pixel value of each pixel of the medical image, or based on the difference between each pixel of the medical image and adjacent pixels. And some object representation methods are sensitive to image noise, so that the object attribute information generated for target objects with relatively high noise has a large deviation.

[0046] In some embodiments, when the number of elements contained in the N types of object attribute information may be the same or different, an attribute alignment operation may be performed on the N types of object attribute information when the N types of object attribute information are obtained so that the lengths of the N types of object attribute information are the same. Optionally, after obtaining the N types of object attribute information, attribute processing may be performed on each type of object attribute information, and the following steps may be performed based on the processed N types of object attribute information. The attribute processing may be obtaining an embedding matrix, and determining the embedded attribute information corresponding to each type of object attribute information based on the embedding matrix. Specifically, the corresponding embedded attribute information may be obtained by multiplying the embedding matrix with each type of object attribute information. The embedding matrix corresponding to each type of object attribute information may be the same or different. The embedding matrix may be set by relevant business personnel or may be obtained based on model training.

[0047] S202 : Extract features of the target object based on N types of object attribute information to obtain object features of the target object under each type of object attribute information.

[0048] In some embodiments, the electronic device can perform feature extraction on N types of object attribute information to obtain object features corresponding to each type of object attribute information, thereby extracting features from the target object based on the N types of object attribute information. The object features corresponding to each type of object attribute information are obtained based on the corresponding object attribute information and the remaining object attribute information, so that the object features can cover more feature information.

[0049] In one possible implementation, assume that any one of N types of object attribute information is represented as the i-th type of object attribute information, where i is a positive integer less than or equal to N. The electronic device extracts features of the target object based on the N types of object attribute information. Obtaining the object features of the target object under each type of object attribute information can be accomplished by determining the degree of correlation between the N types of object attribute information and the i-th type of object attribute information, and then determining the object features corresponding to the i-th type of object attribute information based on the feature information extracted from each type of object attribute information. This allows the object features to include not only the feature information of the corresponding object attribute information, but also the feature information of the remaining object attribute information, thereby increasing the number of features covered. The feature extraction process is implemented using the concept of an attention mechanism.

[0050] In some embodiments, the feature extraction process may specifically include determining the distance between each type of object attribute information and the i-th type of object attribute information, normalizing the distance and using the normalized result as the degree of correlation between each type of object attribute information and the i-th type of object attribute information, multiplying the degree of correlation between each type of object attribute information and the i-th type of object attribute information by each type of object attribute information as the feature information extracted from each type of object attribute information, and determining the sum of the feature information extracted from each type of object attribute information as the object feature corresponding to the i-th type of object attribute information. Determining the distance may involve calculating the Euclidean distance or Chebyshev distance between each type of object attribute information and the i-th type of object attribute information, and the method for determining the distance is not limited herein.

[0051] In some embodiments, the feature extraction process described above may also specifically include obtaining an attention matrix to enhance the representational capability and complexity of object features. Based on the attention matrix, feature extraction may be performed on N types of object attribute information to obtain object features corresponding to each type of object attribute information. Specifically, the attention matrix may be used to determine transition attribute information for each type of object attribute information that includes parameter information of the attention matrix, determine the degree of correlation between the transition attribute information corresponding to the i-th type of object attribute information and the transition attribute information corresponding to each type of object attribute information, and determine feature information extracted from the transition attribute information corresponding to each type of object attribute information based on the correlation. Finally, the object features corresponding to the i-th type of object attribute information may be determined based on the sum of the feature information extracted from the transition attribute information corresponding to each type of object attribute information. Optionally, multiple feature extractions may be performed on the N types of object attribute information based on multiple sets of attention matrices, and the object features corresponding to each type of object attribute information may be determined based on the extraction results obtained from the multiple feature extractions. For the specific implementation process described above, reference may be made to the relevant description of the following embodiments. Among them, the attention matrix can be set by relevant business personnel or obtained through training as a model parameter of the prediction model. The prediction model can be composed of a trained prediction network, and the trained prediction network can include a feature extraction layer. The feature extraction layer can include a feature extractor, and the attention matrix is ​​used as a parameter in the feature extractor.

[0052] S203 : Based on the object features corresponding to each object attribute information, generate prediction indicator distributions of the target object under the target indicator.

[0053] The target indicator has an indicator range, so the predicted indicator distribution corresponding to any type of object attribute information can be a distribution within the indicator range of the target object under the target indicator.

[0054] In one possible embodiment, the electronic device can obtain the above-mentioned prediction model, which includes an indicator distribution prediction layer, and the indicator distribution prediction layer can be a fully connected layer. The electronic device can predict the object features corresponding to each object attribute information based on the indicator distribution prediction layer, and obtain the predicted indicator distribution corresponding to the target object under each object attribute information, that is, the predicted indicator distribution corresponding to each object attribute information can be understood as a plurality of initial indicator distributions within the indicator range of the target object under the target indicator. The predicted indicator distribution can represent the probability distribution of the target indicator within the indicator range. Wherein, when predicting the object features corresponding to each object attribute information, each object feature shares all parameter resources and structural resources in the indicator distribution prediction layer. Compared with discarding neurons each time prediction so that the model resources used for each prediction are different, the present application can avoid the prediction impact caused by different resource usage.

[0055] In some embodiments, the type of the prediction indicator distribution can be a Gaussian distribution. The key parameters of the Gaussian distribution include mean and variance. Therefore, when generating the prediction indicator distribution, the object features corresponding to each type of object attribute information can be predicted through the indicator distribution prediction layer, and the predicted mean and predicted variance corresponding to each type of object attribute information can be output. The prediction indicator distribution corresponding to each type of object attribute information is determined based on the predicted mean and predicted variance corresponding to each type of object attribute information. At this time, the prediction indicator distribution is the probability distribution of the target indicator in the entire horizontal axis domain, that is, the probability distribution from negative infinity to positive infinity. The electronic device can intercept the probability distribution within the indicator range to obtain the final prediction indicator distribution corresponding to the target object under each type of object attribute information.

[0056] In some embodiments, the type of the prediction index distribution can be a discrete distribution, and multiple discrete points can be set according to the index range. When generating the prediction index distribution, the object features corresponding to each type of object attribute information can be predicted through the index distribution prediction layer, and the probability of each type of object attribute information at each discrete point can be output. The prediction index distribution corresponding to each type of object attribute information can be determined based on the probability of each type of object attribute information at each discrete point. For example, the index range is 0-1, and multiple discrete points are set as [0, 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, 1]. The object features corresponding to the object attribute information are predicted through the index distribution prediction layer to obtain the probability of the object attribute information at each discrete point. Subsequently, the discrete distribution constructed based on the probability at each discrete point can be used as the prediction index distribution, or an approximate continuous distribution can be constructed based on the probability at each discrete point, and the continuous distribution within the index range can be used as the prediction index distribution.

[0057] S204: Determine the target prediction index and the target accuracy index of the target prediction index of the target object according to the prediction index distribution corresponding to each object attribute information.

[0058] In one possible embodiment, the electronic device may determine the target indicator distribution of the target object based on the prediction indicator distribution (initial indicator distribution) corresponding to each type of object attribute information, and determine the target prediction indicator of the target object for the target indicator and the target accuracy index of the target prediction indicator based on the target indicator distribution; or the electronic device may first determine the initial prediction indicator corresponding to each type of object attribute information and the initial accuracy index of the initial prediction indicator corresponding to each type of object attribute information based on the prediction indicator distribution corresponding to each type of object attribute information, and determine the target prediction indicator based on the initial prediction indicator corresponding to each type of object attribute information, and determine the target accuracy index based on the initial accuracy index of the initial prediction indicator corresponding to each type of object attribute information. The prediction indicator distribution corresponding to each type of object attribute information jointly acts to determine the target prediction indicator, so that the feature information involved in the prediction process is relatively complete, and thus the prediction result (target prediction indicator and target accuracy index) is relatively accurate.

[0059] In some embodiments, the electronic device may determine the target indicator distribution of the target object by weighted summing up the prediction indicator distribution corresponding to each object attribute information. The weighting coefficient may be set by relevant business personnel based on an empirical value, or may be used as a model parameter in the above-mentioned prediction model and predicted together with the prediction indicator distribution. The target indicator distribution is a distribution of the indicator range. Determining the target prediction indicator and the target accuracy index based on the target indicator distribution may be calculating the mean of the target indicator distribution and using it as the target prediction indicator, and calculating the variance of the target indicator distribution and determining the target accuracy index based on the variance, such as taking the difference between 1 and the variance as the target accuracy index (i.e., 1-variance). The variance can represent the uncertainty of the target prediction indicator to achieve uncertainty estimation (Uncertainty Quantification) for the prediction result.

[0060] In some embodiments, the electronic device determines the initial prediction index corresponding to each type of object attribute information and the initial accuracy index of the initial prediction index corresponding to each type of object attribute information by respectively calculating the mean of the prediction index distribution corresponding to each type of object attribute information and using it as the corresponding initial prediction index, and calculating the variance of the prediction index distribution corresponding to each type of object attribute information and determining the corresponding initial accuracy index based on the variance; the method of determining the initial accuracy index based on the variance can be the same as the method of determining the target accuracy index described above. And determining the target prediction index based on all the initial prediction indicators can be a weighted sum of all the initial prediction indicators to obtain the target prediction index; determining the target accuracy index based on the initial accuracy index of all the initial prediction indicators can be a weighted sum of all the initial accuracy indices to obtain the target accuracy index. The aforementioned weighting coefficients used can be set by relevant business personnel based on empirical values, or can be used as model parameters in the above-mentioned prediction model and predicted together with the prediction indicator distribution.

[0061] In an embodiment of the present application, a target object for which a target indicator needs to be predicted can be obtained, and N types of object attribute information of the target object can be obtained. Feature extraction is performed on the target object based on the N types of object attribute information to obtain the object feature of the target object under each type of object attribute information. The object feature is obtained by the joint action of the N types of attribute information, so that the object feature can contain more feature information. The predicted indicator distribution of the target object within the indicator range is generated based on the object feature corresponding to each type of object attribute information. The same resources are used when predicting each predicted indicator distribution, thereby avoiding the prediction deviation caused by the use of different resources. The target prediction indicator of the target object and the target accuracy index of the target prediction indicator are determined according to the predicted indicator distribution corresponding to each type of object attribute information. The target prediction indicator and the target accuracy index can be determined together through the predicted indicator distribution, and the uncertainty can be expressed by the target accuracy index, thereby improving the prediction efficiency and accuracy of the uncertainty of the indicator and the corresponding indicator result.

[0062] See Figure 3 , Figure 3 This is a flow chart of a data processing method provided in an embodiment of the present application, which can be executed by the electronic device mentioned above. Figure 3 As shown, the process of the data processing method in the embodiment of the present application may include the following:

[0063] S301: Acquire N types of object attribute information of a target object. The specific implementation of step S301 can refer to the relevant description of the above embodiment.

[0064] S302 : Extract features of the target object based on N types of object attribute information to obtain object features of the target object under each type of object attribute information.

[0065] In one possible implementation, the electronic device may call a prediction model to perform feature extraction. The prediction model is composed of a trained prediction network. The trained prediction network includes a feature extraction layer. The feature extraction layer may include a feature extractor. The electronic device may perform feature extraction on N types of object attribute information based on the feature extractor in the feature extraction layer to obtain object features corresponding to each type of object attribute information.

[0066] In some embodiments, any one of N types of object attribute information is represented as the i-th type of object attribute information; the electronic device performs feature extraction on the N types of object attribute information based on a feature extractor to obtain the object features corresponding to each type of object attribute information. Specifically, the electronic device may determine the transition attribute information of each type of object attribute information in the feature extraction layer, determine the degree of correlation between the transition attribute information corresponding to the i-th type of object attribute information and the transition attribute information corresponding to each type of object attribute information, and determine the feature information extracted from the transition attribute information corresponding to each type of object attribute information based on the correlation degree, so as to determine the object features corresponding to the i-th type of object attribute information. The obtained object features include the relationship features between different object attribute information and the feature information representing the deeper level of the corresponding object attribute information. When performing feature extraction, each type of object attribute information shares all parameter resources and structure resources in the feature extraction layer.

[0067] In some embodiments, the feature extractor may be a Transformer structure, that is, its parameters may include three attention matrices (let it be represented as the first attention matrix W Q , the second attention matrix W K , the third attention matrix W V ), through the three attention matrices, the complexity of the prediction model can be increased, the expressive ability of the model can be enhanced, and the object features can contain more feature information.

[0068] In some embodiments, let the i-th object attribute information be represented as H i , then the transition attribute information of the i-th object attribute information (q i 、k i 、v i ) can be obtained by the following formula:

[0069] q i =H i *W Q

[0070] k i =H i *W K

[0071] v i =H i *WV

[0072] The transition attribute information of the above-mentioned i-th object attribute information is the result obtained by multiplying the i-th object attribute information with different attention matrices respectively, which represents the vector representation obtained by mapping the object attribute information based on different attention matrices; each object attribute information obtains the corresponding transition attribute information in the above-mentioned manner.

[0073] In some embodiments, the degree of correlation can be determined by regularization, that is, determining the transition attribute information corresponding to the i-th object attribute information and the transition attribute information corresponding to the j-th object attribute information (q j 、k j 、v j )'s correlation level I ij It can be obtained by the following formula:

[0074]

[0075] Among them, q i .k j Indicates q i With k j dot product; represents k j The square root of the number of dimensions; j is an integer from 1 to N, and j can be equal to i; thereby obtaining the degree of correlation between the transition attribute information corresponding to the i-th object attribute information and the transition attribute information corresponding to each type of object attribute information.

[0076] In some embodiments, according to the degree of correlation I ij Determine the feature information L extracted from the transition attribute information corresponding to the j-th object attribute information ij It can be obtained by the following formula:

[0077] L ij =I ij *v j

[0078] In some embodiments, the electronic device may determine the sum of the feature information extracted from the transition attribute information corresponding to each type of object attribute information as the object feature corresponding to the i-th type of object attribute information; or it may be that the sum of the feature information is input into the fully connected layer connected to the feature extractor, and the output through the fully connected layer is determined as the final object feature.

[0079] In one possible embodiment, the feature extraction layer includes R feature extractors, where R is an integer greater than 1, each feature extractor includes a set of attention matrices, and the attention matrices between each feature extractor are independent of each other. Multiple feature extractions can also be performed on N types of object attribute information based on the multiple sets of attention matrices, and the object features corresponding to each type of object attribute information can be determined based on the extraction results obtained from the multiple feature extractions.

[0080] In some embodiments, the R feature extractors described above can be serial in the feature extraction layer, that is, the feature extraction layer can include R sublayers, each of which can include a feature extractor, and each feature extractor is executed according to the above process. Therefore, multiple feature extractions can be performed by using the output of the previous feature extractor as the input of the next feature extractor, and the output of the last feature extractor as the object feature corresponding to each type of object attribute information. Alternatively, each sublayer can be connected to a fully connected layer, and the output of the previous feature extractor is input to the feature extractor via the fully connected layer.

[0081] In some embodiments, the R feature extractors can be parallel in the feature extraction layer. In this case, it can be understood that the R feature extractors serve as a sublayer, and each feature extractor executes according to the above process. That is, based on N types of object attribute information, the target object is extracted, and the object features of the target object under each type of object attribute information are obtained. In the feature extraction layer, the N types of object attribute information are input into each feature extractor respectively, and the R feature extractors are called to extract the i-th type of object attribute information based on the N types of object attribute information, and R initial extraction features of the i-th type of object attribute information are obtained. According to the R initial extraction features, the object features of the target object under the i-th type of object attribute information are generated to realize multiple feature extractions. At this time, when multiple feature extractors are processed in parallel, it can be understood that a multi-head attention mechanism is adopted.

[0082] Among them, the R feature extractors can be connected to a fully connected layer; therefore, generating the object features of the target object under the i-th object attribute information based on the R initial extracted features can be: feature splicing of the R initial extracted features to obtain the target extracted features of the i-th object attribute information, and feature dimensionality reduction of the target extracted features through the fully connected layer to obtain the object features of the target object under the i-th object attribute information.

[0083] In some embodiments, the R feature extractors in the feature extraction layer may be partially parallel and partially serial, that is, the feature extraction layer may include X sublayers, each of which may include one or more feature extractors, each sublayer is connected to a fully connected layer, and the output of each sublayer is input to the next sublayer via the fully connected layer, so that the dimension of the output of each sublayer remains consistent. The execution process in each sublayer and between each sublayer can refer to the relevant description above.

[0084] For example, Figure 4a-4d As shown, Figure 4a-4d A schematic diagram of a feature extraction scenario provided in an embodiment of the present application; wherein:

[0085] like Figure 4a As shown, the feature extraction layer includes a feature extractor (W Q 、W K 、W V ), N kinds of object attribute information are first attribute information H1 and second attribute information H2; based on the feature extractor, the transition attribute information (q1, k1, v1) of H1 and the transition attribute information (q2, k2, v2) of H2 are determined; the correlation degree I between the transition attribute information of H1 and the transition attribute information of H1 is determined 11 and the degree of correlation with the transition attribute information of H2 12 , and determine the correlation degree I between the transition attribute information of H2 and the transition attribute information of H1 21 and the degree of correlation with the transition attribute information of H2 22 According to I 11 Determine the feature information L extracted from the transition attribute information corresponding to H1 under H1 11 and according to I 12 Determine the feature information L extracted from the transition attribute information corresponding to H2 12 , according to I 21 Determine the feature information L extracted from the transition attribute information corresponding to H1 under H2 21 and according to I 22 Determine the feature information L extracted from the transition attribute information corresponding to H2 22 ; L 11 and L 12 The sum of the two and the output of the fully connected layer is determined as the object feature Z1 of H1, and L 21 and L 22 The sum is determined as the object feature Z2 of H2 through the output of the fully connected layer;

[0086] For example, Figure 4b As shown, the feature extraction layer includes two feature extractors (extractor 1 (W1 Q 、W1 K 、W1V ), Extractor 2 (W2 Q 、W2 K 、W2 V ), each feature extractor is serially connected to a fully connected layer; N kinds of object attribute information are the first attribute information and the second attribute information Will and Input to Extractor 1, which follows the above process and outputs through the fully connected layer Corresponding and Corresponding Will and Input to Extractor 2, which follows the above process and outputs through the fully connected layer Corresponding and Corresponding Will Determined The object feature Z1, and the Determined The object feature Z2;

[0087] For example, Figure 4c As shown, the feature extraction layer includes two feature extractors (extractor a(W1 Q 、W1 K 、W1 V ), extractor b(W2 Q 、W2 K 、W2 V ), each feature extractor is connected in parallel to a fully connected layer; N kinds of object attribute information are the first attribute information H1 and the second attribute information H2; H1 and H2 are input into extractor a and extractor b respectively, and extractor a outputs the corresponding H1 according to the above process Corresponding to H2 Extractor b outputs the corresponding H1 according to the above process Corresponding to H2 Will and Splicing to get splicing feature 1, and and Splicing to obtain splicing feature 2, performing feature dimensionality reduction on splicing feature 1 through a fully connected layer to obtain object feature Z1 of H1, and performing feature dimensionality reduction on splicing feature 2 through a fully connected layer to obtain object feature Z2 of H2;

[0088] For example, Figure 4d As shown, the feature extraction layer includes four feature extractors (extractor a(W1 Q、W1 K 、W1 V ), extractor b(W2 Q 、W2 K 、W2 V ), extractor c(W3 Q 、W3 K 、W3 V ), extractor d(W4 Q 、W4 K 、W4 V ), extractor a and extractor b are a sublayer and connected to a fully connected layer, extractor c and extractor d are a sublayer and connected to a fully connected layer; N kinds of object attribute information are the first attribute information and the second attribute information Will and Input extractors a and b respectively, and extractor a outputs the same process as above. Corresponding and Corresponding The extractor b outputs the Corresponding and Corresponding Will and Splicing to get splicing feature 1, and and Splicing to obtain splicing feature 2, splicing feature 1 through the fully connected layer to obtain Corresponding The splicing feature 2 is obtained through the fully connected layer Corresponding Will and Input into extractor c and extractor d respectively, and extractor c outputs according to the above process Corresponding and Corresponding Output by extractor d according to the above process Corresponding and Corresponding Will and Splicing to get splicing feature 3, and and Splicing obtains splicing feature 4, and splicing feature 3 is obtained through the fully connected layer Corresponding The splicing feature 2 is obtained through the fully connected layer Corresponding Will Determined The object feature Z1, and the Determined The object feature Z2.

[0089] S303: Generate prediction index distribution of the target object under the target index based on the object features corresponding to each object attribute information. The specific implementation of step S303 can refer to the relevant description of the above embodiment.

[0090] S304: Determine the target prediction index and the target accuracy index of the target prediction index of the target object according to the prediction index distribution corresponding to each object attribute information.

[0091] In some embodiments, the electronic device may determine the initial prediction index of the target object under each type of object attribute information and the initial accuracy index of each initial prediction index based on the prediction index distribution corresponding to each type of object attribute information, determine the target prediction index based on the initial prediction index of the target object under each type of object attribute information, and determine the target accuracy index of the target prediction index based on the initial accuracy index of each initial prediction index. The target prediction index may be determined by weighted summing of each initial prediction index, and the target accuracy index may be determined by weighted summing of the initial accuracy indexes of each initial prediction index. The weighting coefficient may be obtained by model prediction.

[0092] In some embodiments, the initial prediction indicator can be determined by taking the mean of the corresponding prediction indicator distribution as the initial prediction indicator; or, if the prediction indicator distribution is a Gaussian distribution, the mean of the prediction indicator distribution under the entire horizontal axis domain can be used as the initial prediction indicator, that is, when generating the prediction indicator distribution, it is obtained by the prediction mean and prediction variance corresponding to each predicted object attribute information, so the prediction mean corresponding to each object attribute information can be used as the initial prediction indicator corresponding to each object attribute information.

[0093] In some embodiments, the initial accuracy index can be determined by taking the variance of the corresponding prediction indicator distribution as the initial prediction indicator; or, if the prediction indicator distribution is a Gaussian distribution, the variance of the prediction indicator distribution under the entire horizontal axis domain can be used as the initial prediction indicator, that is, when generating the prediction indicator distribution, it is obtained by the predicted mean and predicted variance corresponding to each predicted object attribute information, so the predicted variance corresponding to each object attribute information can be used as the predicted variance corresponding to each object attribute information.

[0094] In some embodiments, the prediction index distribution can measure the contribution of object attribute information to a target prediction index or target accuracy index. The weighted initial prediction index can represent the contribution to the target prediction index, and the weighted initial accuracy index can represent the contribution to the target accuracy index. This initial accuracy index or weighted initial accuracy index can be used to subsequently screen the N types of object attribute information. Compared to generating an overall prediction result for the input data, this approach can further determine the prediction quality of each type of object attribute information for the target object, refining the granularity of the prediction process.

[0095] In one possible embodiment, the above-mentioned prediction process can be implemented by a model, and the electronic device can obtain sample data and perform model training based on the sample data to obtain a prediction model. Specifically, the process can include obtaining a sample object and N types of sample attribute information of the sample object, wherein the sample object has an indicator label and an accuracy index label, calling a prediction network to extract features of the sample object based on the N types of sample attribute information, obtaining sample object features of the sample object under each type of sample attribute information, generating a sample indicator distribution of the sample object under the target indicator based on the sample object features corresponding to each type of sample attribute information, generating a sample indicator and a sample accuracy index corresponding to each type of sample attribute information based on the sample indicator distribution corresponding to each type of sample attribute information, generating a prediction deviation of the prediction network for the sample indicator distribution based on the sample indicator and sample accuracy index corresponding to each type of sample attribute information, the indicator label, and the accuracy index label, and correcting the network parameters of the prediction network based on the prediction deviation to obtain a trained prediction network.

[0096] In some embodiments, if the prediction indicator distribution is a Gaussian distribution, the model training may specifically include obtaining a sample object and obtaining N types of sample attribute information of the sample object, wherein the sample object has an indicator label, calling a prediction network to perform feature extraction on the sample object based on the N types of sample attribute information, obtaining sample object features of the sample object under each type of sample attribute information, generating a sample indicator distribution of the sample object under the target indicator based on the sample object features corresponding to each type of sample attribute information, generating a prediction deviation of the prediction network for the sample indicator distribution based on the sample indicator distribution and the indicator label corresponding to each type of sample attribute information, correcting the network parameters of the prediction network based on the prediction deviation, and obtaining a trained prediction network. The prediction model includes the trained prediction network. The trained prediction network may include a feature extraction layer and an indicator distribution prediction layer. In the prediction model, the feature extraction layer may perform feature extraction based on the N types of object attribute information to obtain object features, and the indicator distribution prediction layer may generate a predicted indicator distribution of the target object within the indicator range based on the object features.

[0097] The above generated prediction deviation L can be obtained by the following formula:

[0098]

[0099] Where i represents the attribute information of the i-th object; w i Represents the weighted coefficient of the prediction index distribution corresponding to the i-th object attribute information; N(x; μ i ; σ i ) represents the prediction index distribution (Gaussian distribution) corresponding to the attribute information of the i-th object, μ i represents the predicted mean corresponding to the attribute information of the i-th object, σ i It represents the prediction variance corresponding to the attribute information of the i-th object, x represents the value of the target indicator, and y represents the indicator label. The prediction deviation can also be obtained in other ways, which can be set by relevant business personnel when building the model.

[0100] In some embodiments, the model can be trained by predicting the deviation to minimize the prediction deviation until the model converges. At this time, the parameters in the model are the trained parameters. * The training process can be expressed by the following maximum likelihood estimation formula:

[0101] W * =argmax W [-L]

[0102] The above means that in the process of polling and correcting the model, W that makes -L the largest is obtained, and W at this time is used as the model parameter W of the trained prediction model. * .

[0103] Optionally, the above-mentioned prediction model includes the trained prediction network. The trained prediction network may include a feature extraction layer and an indicator distribution prediction layer. In the prediction model, the feature extraction layer may perform feature extraction based on N types of object attribute information to obtain object features, and the indicator distribution prediction layer may generate a prediction indicator distribution of the target object within the indicator range and a weighted coefficient for each prediction indicator distribution based on the object features. The weighted coefficient may be used to weightedly sum each prediction indicator distribution to obtain a target indicator distribution for the target object; or may be used to weightedly sum each initial prediction indicator to obtain a target prediction indicator and to weightedly sum the initial accuracy index of each initial prediction indicator to obtain a target accuracy index. The multiple prediction indicator distributions may be multiple Gaussian distributions, so the prediction model may be an architecture designed by combining a deformer and a Gaussian Mixture Model (GMM). In addition, by sharing resources in the prediction model during prediction, the number of parameters in the model and the degree of overfitting of the model may be reduced.

[0104] For example, Figure 5 As shown, Figure 5A schematic diagram of a scenario for predicting a target indicator of a target object provided by an embodiment of the present application; wherein, assuming that the target object is a molecular object, the prediction model performs prediction of the target indicator, the prediction model includes a feature extraction layer and an indicator distribution prediction layer, the feature extraction layer includes multiple sublayers, each sublayer includes multiple feature extractors and each sublayer is connected to a fully connected layer, and the indicator distribution prediction layer can be a fully connected layer; molecular fingerprints are extracted from the molecular object according to M molecular fingerprint extraction methods to obtain M object attribute information, and N object attribute information (H1, H2, ..., H2) is selected from the M object attribute information. N ) is input into the prediction model, and the feature extraction layer outputs the object features (Z1, Z2, ..., Z N ), and the indicator distribution prediction layer predicts the object features corresponding to each object attribute information, and outputs the output results corresponding to each object attribute information, which include weighted coefficients, predicted mean and predicted variance (w i ,μ i ,σ 2 i ), the prediction result of the target index of the molecular object can be obtained by weighting the predicted mean based on the weighted coefficient corresponding to each object attribute information (∑w i μ i ) and the weighted summation of the prediction variance based on the weighted coefficient corresponding to each object attribute information to obtain the uncertainty of the prediction result (∑w i μ i 2 ); The greater the uncertainty, the less reliable the prediction result; conversely, the smaller the uncertainty, the greater the reliability of the prediction result. Accordingly, the uncertainty can be converted into an accuracy index (1-∑w i μ i 2 ) to measure the reliability of the prediction results; the product of the weighted coefficient corresponding to the i-th object attribute information and the prediction variance can characterize the impact of the i-th object attribute information on the uncertainty of the final prediction result, which can be used for the subsequent screening of the output object attribute information.

[0105] S305: If the target accuracy index is less than the index threshold, reacquire N types of object attribute information, and re-determine the prediction index of the target object and the accuracy index of the prediction index based on the reacquired N types of object attribute information.

[0106] In one possible implementation, a target accuracy index represents the reliability of the target prediction index. If the target accuracy index is less than an index threshold, the target index can be re-predicted for the target object. The target object has M types of object attribute information, which are obtained by representing the target object based on M object representation methods, where M is an integer greater than N. The N types of object attribute information are extracted from the M types of object attribute information, and thus re-prediction can be performed based on the M types of object attribute information.

[0107] In some embodiments, the above may be to re-acquire N kinds of object attribute information from M kinds of object attribute information, and the re-acquired N kinds of object attribute information may be partially different or completely different from the N kinds of object attribute information used in the previous prediction, and the above process is performed on the re-acquired N kinds of object attribute information to re-determine the prediction index of the target object and the accuracy index of the prediction index, so that the determined target accuracy index is greater than or equal to the index threshold, indicating that the target prediction index obtained at this time is reliable, and the target prediction index can be applied in combination with the actual business scenario. If the target accuracy index obtained in all the re-predictions is less than the index threshold, the target prediction index corresponding to the highest target accuracy index can be used. In this way, the N kinds of object attribute information can be screened to obtain N kinds of object attribute information that can make the prediction results more reliable and accurate. The index threshold can be set by relevant business personnel based on experience, so that appropriate object attribute information can be dynamically selected for target indicator prediction during the prediction process.

[0108] In some embodiments, reacquiring N types of object attribute information from M types of object attribute information can specifically be: determining the object attribute information to be replaced from the N types of object attribute information based on the initial accuracy index corresponding to each type of object attribute information, selecting target object attribute information from the object attribute information other than the N types of object attribute information from the M types of object attribute information, and determining the object attribute information other than the object attribute information to be replaced and the target object attribute information in the N types of object attribute information as the reacquired N types of object attribute information.

[0109] In some embodiments, determining the object attribute information to be replaced from the N types of object attribute information based on the initial accuracy index corresponding to each type of object attribute information may include determining one or more types of object attribute information with the smallest initial accuracy index from the N types of object attribute information as the object attribute information to be replaced; or determining one or more types of object attribute information with the smallest weighted initial accuracy index from the N types of object attribute information as the object attribute information to be replaced. Selecting the target object attribute information from the object attribute information other than the N types of object attribute information may include randomly selecting, from the object attribute information other than the N types of object attribute information, the same number of object attribute information as the object attribute information to be replaced as the target object attribute information.

[0110] In some embodiments, reacquiring N types of object attribute information from M types of object attribute information may also be randomly selecting N types of object attribute information from the M types of object attribute information, and the randomly selected object attribute information is a combination of object attribute information that has not been used in previous predictions.

[0111] In one possible embodiment, if the target accuracy index is greater than or equal to the index threshold, the target prediction indicator can be subsequently applied, and the application can be determined based on the actual business scenario. For example, if the target object is a molecular object representing a chemical molecule, and the target prediction indicator is used to indicate the stability of the molecular object, then the specific application can be to determine the stability level of the molecular object based on the target prediction indicator. When performing chemical molecule synthesis, technical personnel in the relevant field can use the stability level of the molecular object as a reference factor or select valuable chemical molecules. For another example, if the target object is a medical image, and the target prediction indicator is used to indicate the abnormality index of the medical image, then the specific application can be to determine the probability of disease indicated by the medical image based on the target prediction indicator. When performing diagnosis, technical personnel in the relevant field can use the probability of disease as a reference factor.

[0112] In an embodiment of the present application, a target object for which a target indicator needs to be predicted can be obtained, and N types of object attribute information of the target object can be obtained. Feature extraction is performed on the target object based on the N types of object attribute information to obtain an object feature of the target object under each type of object attribute information. The object feature is obtained by the joint action of the N types of attribute information, which can make the object feature contain more feature information. A prediction indicator distribution of the target object within the indicator range is generated based on the object feature corresponding to each type of object attribute information. The same resources are used when predicting each prediction indicator distribution, thereby avoiding prediction bias caused by different resource usage. A target prediction indicator and a target accuracy index of the target prediction indicator of the target object are determined based on the prediction indicator distribution corresponding to each type of object attribute information. The target prediction indicator and the target accuracy index can be determined together using the prediction indicator distribution, and uncertainty can be expressed using the target accuracy index, thereby improving the prediction efficiency and accuracy of the uncertainty of the indicator and the corresponding indicator result. If the target accuracy index is less than an index threshold, the N types of object attribute information are re-acquired, and the prediction indicator of the target object and the accuracy index of the prediction indicator are re-determined based on the re-acquired N types of object attribute information. The target accuracy index can be used to determine whether to re-predict, so as to ensure that the obtained target prediction indicator has higher reliability.

[0113] See Figure 6 , Figure 6 This is a structural diagram of a data processing device provided by this application. It should be noted that, Figure 6 The data processing device shown is used to execute the application Figure 2 and Figure 3 For the sake of convenience, only the part related to the embodiment of the present application is shown, and the specific technical details are not disclosed. Figure 2 and Figure 3 The data processing device 600 may include: an acquisition module 601, a processing module 602, and a determination module 603.

[0114] An acquisition module 601 is configured to acquire N types of object attribute information of a target object, where N is an integer greater than 1;

[0115] A processing module 602 is configured to extract features of a target object based on N types of object attribute information to obtain object features of the target object under each type of object attribute information;

[0116] The processing module 602 is further configured to generate a prediction indicator distribution of the target object under the target indicator based on the object features corresponding to each type of object attribute information; the prediction indicator distribution corresponding to any type of object attribute information is within the indicator range of the target object under the target indicator;

[0117] The determination module 603 is configured to determine a target prediction index and a target accuracy index of the target prediction index of the target object according to the prediction index distribution corresponding to each type of object attribute information.

[0118] In some embodiments, any one of the N types of object attribute information is represented as the i-th type of object attribute information, where i is a positive integer less than or equal to N;

[0119] When the processing module 602 is used to extract features of the target object based on N types of object attribute information and obtain the object features of the target object under each type of object attribute information, it is specifically used to:

[0120] Obtain a trained prediction network; the trained prediction network includes R feature extractors;

[0121] Call R feature extractors to extract features of the i-th object attribute information based on N types of object attribute information, and obtain R initial extracted features of the i-th object attribute information;

[0122] The object features of the target object under the i-th object attribute information are generated based on the R initial extracted features.

[0123] In some embodiments, when the processing module 602 is used to generate the object feature of the target object under the i-th object attribute information based on the R initial extracted features, it is specifically used to:

[0124] Perform feature splicing on the R initial extracted features to obtain the target extracted features of the i-th object attribute information;

[0125] Perform feature dimensionality reduction on the target extracted features to obtain the object features of the target object under the i-th object attribute information.

[0126] In some embodiments, the processing module 602 is further configured to:

[0127] Obtain a sample object and obtain N types of sample attribute information of the sample object; the sample object has an indicator label;

[0128] Call the prediction network to extract features of the sample object based on N types of sample attribute information, and obtain the sample object features under each type of sample attribute information;

[0129] Generate a sample indicator distribution of the sample object under the target indicator based on the sample object characteristics corresponding to each type of sample attribute information;

[0130] Generate the prediction network's prediction deviation for the sample indicator distribution based on the sample indicator distribution and indicator label corresponding to each sample attribute information;

[0131] The network parameters of the prediction network are corrected according to the prediction deviation to obtain a trained prediction network.

[0132] In some embodiments, when determining the target prediction index and the target accuracy index of the target prediction index of the target object according to the prediction index distribution corresponding to each type of object attribute information, the determination module 603 is specifically configured to:

[0133] According to the distribution of prediction indicators corresponding to each object attribute information, the initial prediction indicator of the target object under each object attribute information and the initial accuracy index of each initial prediction indicator are determined respectively;

[0134] Determine a target prediction index based on the initial prediction index of the target object under each object attribute information;

[0135] A target accuracy index of the target predictor is determined according to the initial accuracy index of each initial predictor.

[0136] In some embodiments, the target object has M types of object attribute information, and the N types of object attribute information are extracted from the M types of object attribute information, where M is an integer greater than N;

[0137] The processing module 602 is further configured to:

[0138] If the target accuracy index is less than the index threshold, the object attribute information to be replaced is determined from the N types of object attribute information according to the initial accuracy index corresponding to each type of object attribute information;

[0139] Selecting target object attribute information from the M types of object attribute information excluding the N types of object attribute information;

[0140] Based on the object attribute information other than the object attribute information to be replaced and the target object attribute information in the N types of object attribute information, a prediction index of the target object and an accuracy index of the prediction index are re-determined.

[0141] In some embodiments, the target object is a molecular object, and the target prediction indicator is used to indicate the stability of the molecular object;

[0142] The determination module 603 is further configured to:

[0143] If the target accuracy index is greater than or equal to the index threshold, the stability level of the molecular object is determined based on the target prediction index.

[0144] In an embodiment of the present application, an acquisition module acquires N kinds of object attribute information of a target object; a processing module extracts features of the target object based on the N kinds of object attribute information to obtain the object features of the target object under each kind of object attribute information; a processing module generates prediction index distributions of the target object under the target index based on the object features corresponding to each kind of object attribute information; a determination module determines the target prediction index of the target object and the target accuracy index of the target prediction index according to the prediction index distributions corresponding to each kind of object attribute information. Through the above-mentioned device, the object feature is obtained by the joint action of N kinds of attribute information, which can make the object feature contain more feature information, use the same resources when predicting each prediction index distribution, thereby avoiding the prediction bias caused by different resource usage, and can determine the target prediction index and the target accuracy index together through the prediction index distribution. The target accuracy index can be used to represent uncertainty, thereby improving the prediction efficiency and accuracy of the uncertainty of the indicator and the corresponding indicator result.

[0145] The functional modules in each embodiment of the present application may be integrated into one module, or each module may exist physically separately, or two or more modules may be integrated into one module. The above-mentioned integrated modules may be implemented in the form of hardware or software functional modules, which is not limited in this application.

[0146] See Figure 7 , Figure 7 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present application. Figure 7 As shown, the electronic device 700 includes: at least one processor 701 and a memory 702. Optionally, the electronic device may further include a network interface. The processor 701, the memory 702, and the network interface can exchange data. The network interface is controlled by the processor 701 to send and receive messages. The memory 702 is used to store a computer program, which includes program instructions. The processor 701 is used to execute the program instructions stored in the memory 702. The processor 701 is configured to invoke the program instructions to execute the above method.

[0147] Among them, the memory 702 may include a volatile memory (volatile memory), such as a random-access memory (RAM); the memory 702 may also include a non-volatile memory (non-volatile memory), such as a flash memory (flash memory), a solid-state drive (SSD), etc.; the memory 702 may also include a combination of the above types of memory.

[0148] The processor 701 may be a central processing unit (CPU). In one embodiment, the processor 701 may also be a graphics processing unit (GPU). The processor 701 may also be a combination of a CPU and a GPU.

[0149] In one possible implementation, the memory 702 is used to store program instructions, and the processor 701 may call the program instructions to perform the following steps:

[0150] Obtaining a target object for which a target indicator needs to be predicted, and obtaining N types of object attribute information of the target object, where N is an integer greater than 1; the target indicator has an indicator range;

[0151] Extract features of the target object based on N types of object attribute information to obtain the object features of the target object under each type of object attribute information;

[0152] Based on the object features corresponding to each object attribute information, the predicted indicator distribution of the target object under the target indicator is generated respectively; the predicted indicator distribution corresponding to any object attribute information is within the indicator range of the target object under the target indicator;

[0153] The target prediction index of the target object and the target accuracy index of the target prediction index are determined according to the prediction index distribution corresponding to each object attribute information.

[0154] In some embodiments, any one of the N types of object attribute information is represented as the i-th type of object attribute information, where i is a positive integer less than or equal to N;

[0155] When the processor 701 is configured to extract features of the target object based on N types of object attribute information and obtain the object features of the target object under each type of object attribute information, it is specifically configured to:

[0156] Obtain a trained prediction network; the trained prediction network includes R feature extractors;

[0157] Call R feature extractors to extract features of the i-th object attribute information based on N types of object attribute information, and obtain R initial extracted features of the i-th object attribute information;

[0158] The object features of the target object under the i-th object attribute information are generated based on the R initial extracted features.

[0159] In some embodiments, when the processor 701 is configured to generate the object feature of the target object under the i-th object attribute information based on the R initial extracted features, it is specifically configured to:

[0160] Perform feature splicing on the R initial extracted features to obtain the target extracted features of the i-th object attribute information;

[0161] Perform feature dimensionality reduction on the target extracted features to obtain the object features of the target object under the i-th object attribute information.

[0162] In some embodiments, the processor 701 is further configured to:

[0163] Obtain a sample object and obtain N types of sample attribute information of the sample object; the sample object has an indicator label;

[0164] Call the prediction network to extract features of the sample object based on N types of sample attribute information, and obtain the sample object features under each type of sample attribute information;

[0165] Generate the sample index distribution of the sample object under the target index based on the sample object characteristics corresponding to each sample attribute information;

[0166] Generate the prediction network's prediction deviation for the sample indicator distribution based on the sample indicator distribution and indicator label corresponding to each sample attribute information;

[0167] The network parameters of the prediction network are corrected according to the prediction deviation to obtain a trained prediction network.

[0168] In some embodiments, when the processor 701 is configured to determine the target prediction index and the target accuracy index of the target prediction index of the target object according to the prediction index distribution corresponding to each type of object attribute information, the processor 701 is specifically configured to:

[0169] According to the distribution of prediction indicators corresponding to each object attribute information, the initial prediction indicator of the target object under each object attribute information and the initial accuracy index of each initial prediction indicator are determined respectively;

[0170] Determine a target prediction index based on the initial prediction index of the target object under each object attribute information;

[0171] A target accuracy index of the target predictor is determined according to the initial accuracy index of each initial predictor.

[0172] In some embodiments, the target object has M types of object attribute information, and the N types of object attribute information are extracted from the M types of object attribute information, where M is an integer greater than N;

[0173] The processor 701 is further configured to:

[0174] If the target accuracy index is less than the index threshold, the object attribute information to be replaced is determined from the N types of object attribute information according to the initial accuracy index corresponding to each type of object attribute information;

[0175] Selecting target object attribute information from the M types of object attribute information excluding the N types of object attribute information;

[0176] Based on the object attribute information other than the object attribute information to be replaced and the target object attribute information in the N types of object attribute information, a prediction index of the target object and an accuracy index of the prediction index are re-determined.

[0177] In some embodiments, the target object is a molecular object, and the target prediction indicator is used to indicate the stability of the molecular object;

[0178] The processor 701 is further configured to:

[0179] If the target accuracy index is greater than or equal to the index threshold, the stability level of the molecular object is determined based on the target prediction index.

[0180] In a specific implementation, the above-described devices, processors, memories, etc. can execute the implementation methods described in the above-mentioned method embodiments, and can also execute the implementation methods described in the embodiments of this application, which will not be repeated here.

[0181] The present application also provides a computer (readable) storage medium that stores a computer program. The computer program includes program instructions. When the program instructions are executed by a processor, the processor can execute some or all of the steps performed in the above method embodiments. Optionally, the computer storage medium can be volatile or non-volatile. The computer-readable storage medium can mainly include a program storage area and a data storage area. The program storage area can store an operating system, at least one application required for a function, etc.; the data storage area can store data created based on the use of the blockchain node, etc.

[0182] An embodiment of the present application further provides a computer program product, which includes computer instructions. When the computer instructions are executed by a processor, some or all of the steps in the above method can be implemented.

[0183] In this document, "plurality" refers to two or more. "And / or" describes a relationship between associated objects, indicating that three possible relationships exist. For example, "A and / or B" can mean: A exists alone, A and B exist simultaneously, or B exists alone. The character " / " generally indicates an "or" relationship between the associated objects.

[0184] Those skilled in the art will appreciate that all or part of the processes in the above-described method embodiments can be implemented by instructing related hardware through a computer program. The aforementioned program can be stored in a computer storage medium, which can be a computer-readable storage medium. When the program is executed, it can include the processes in the above-described method embodiments. The aforementioned storage medium can be a magnetic disk, an optical disk, a read-only memory (ROM), or a random access memory (RAM).

[0185] The above disclosure is only part of the embodiments of the present application, and it is certainly not intended to limit the scope of the rights of the present application. Ordinary technicians in this field can understand that all or part of the processes of the above embodiments and equivalent changes made in accordance with the claims of the present application are still within the scope covered by the present application.

Claims

1. A data processing method, characterized in that: The method comprises: Acquire N types of object attribute information of a target object, where N is an integer greater than 1, and the target object is a chemical molecule object or a medical image; Extracting features of the target object based on the N types of object attribute information to obtain object features of the target object under each type of object attribute information; The object features corresponding to each type of object attribute information are respectively used to perform distribution prediction processing to generate a predicted indicator distribution of the target object under the target indicator; the predicted indicator distribution corresponding to each type of object attribute information is within the indicator range of the target object under the target indicator; if the target object is the chemical molecular object, the target indicator is an indicator of the stability of the chemical molecular object; if the target object is the medical image, the target indicator is an indicator of the abnormality index of the medical image; The prediction indicator distributions corresponding to each type of object attribute information are combined and calculated to obtain a target prediction indicator of the target object and a target accuracy index of the target prediction indicator; wherein, if the target object is the chemical molecular object, the target prediction indicator is used to indicate the stability of the chemical molecular object; if the target object is the medical image, the target prediction indicator is used to indicate the abnormality index of the medical image.

2. The method according to claim 1, characterized in that Any one of the N types of object attribute information is represented as the i-th type of object attribute information, where i is a positive integer less than or equal to N; The extracting features of the target object based on the N types of object attribute information to obtain object features of the target object under each type of object attribute information includes: Obtaining a trained prediction network; the trained prediction network includes R feature extractors; Calling the R feature extractors to perform feature extraction on the i-th object attribute information based on the N types of object attribute information, respectively, to obtain R initial extracted features of the i-th object attribute information; An object feature of the target object under the i-th object attribute information is generated according to the R initial extracted features.

3. The method according to claim 2, characterized in that Generating the object feature of the target object under the i-th object attribute information according to the R initial extracted features includes: Performing feature splicing on the R initial extracted features to obtain a target extracted feature of the i-th object attribute information; Performing feature dimensionality reduction on the target extracted features to obtain object features of the target object under the i-th object attribute information.

4. The method according to claim 2, characterized in that The method further comprises: Obtaining a sample object and obtaining N types of sample attribute information of the sample object; the sample object has an indicator label; Calling the prediction network to perform feature extraction on the sample object based on the N types of sample attribute information to obtain sample object features under each type of sample attribute information; Generating a sample indicator distribution of the sample object under the target indicator based on the sample object characteristics corresponding to each type of sample attribute information; Generate a prediction deviation of the prediction network for the sample indicator distribution according to the sample indicator distribution and the indicator label corresponding to each type of sample attribute information; The network parameters of the prediction network are corrected according to the prediction deviation to obtain the trained prediction network.

5. The method according to claim 1, characterized in that The combining and calculating the prediction index distributions corresponding to each type of object attribute information to obtain the target prediction index of the target object and the target accuracy index of the target prediction index includes: Determining, according to the prediction indicator distribution corresponding to each object attribute information, an initial prediction indicator of the target object under each object attribute information and an initial accuracy index of each initial prediction indicator; Determining the target prediction index according to the initial prediction index of the target object under each object attribute information; A target accuracy index of the target prediction indicator is determined according to the initial accuracy index of each initial prediction indicator.

6. The method according to claim 5, characterized in that The target object has M types of object attribute information, and the N types of object attribute information are extracted from the M types of object attribute information, where M is an integer greater than N; The method further comprises: If the target accuracy index is less than the index threshold, determining the object attribute information to be replaced from the N types of object attribute information according to the initial accuracy index corresponding to each type of object attribute information; Selecting target object attribute information from the M types of object attribute information excluding the N types of object attribute information; Based on the object attribute information other than the object attribute information to be replaced in the N types of object attribute information and the target object attribute information, a prediction index of the target object and an accuracy index of the prediction index are re-determined.

7. The method according to claim 1, characterized in that If the target object is the chemical molecule object, the method further includes: If the target accuracy index is greater than or equal to the index threshold, the stability level of the chemical molecule object is determined according to the target prediction index.

8. A data processing device, characterized in that: The device comprises: an acquisition module, configured to acquire N types of object attribute information of a target object, wherein N is an integer greater than 1, and the target object is a chemical molecule object or a medical image; a processing module, configured to extract features of the target object based on the N types of object attribute information, and obtain object features of the target object under each type of object attribute information; The processing module is further configured to respectively perform distribution prediction processing using the object features corresponding to each type of object attribute information to generate a predicted indicator distribution of the target object under the target indicator; the predicted indicator distribution corresponding to each type of object attribute information is within the indicator range of the target object under the target indicator; if the target object is the chemical molecule object, the target indicator is an indicator of the stability of the chemical molecule object; if the target object is the medical image, the target indicator is an indicator of the abnormality index of the medical image; A determination module is used to combine and calculate the prediction indicator distribution corresponding to each type of object attribute information to obtain a target prediction indicator of the target object and a target accuracy index of the target prediction indicator; wherein, if the target object is the chemical molecular object, the target prediction indicator is used to indicate the stability of the chemical molecular object; if the target object is the medical image, the target prediction indicator is used to indicate the abnormality index of the medical image.

9. An electronic device, characterized in that: The method comprises a processor and a memory, wherein the memory is used to store a computer program, the computer program includes program instructions, and the processor is configured to call the program instructions to execute the method according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a computer program, wherein the computer program includes program instructions, and when the program instructions are executed by a processor, the processor is caused to perform the method according to any one of claims 1 to 7.

11. A computer program product, characterized in that The computer program product comprises computer instructions, and when the computer instructions are executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images