Data classification method and apparatus, model training method and apparatus, and device and storage medium

Abstract

Provided in the present application are a data classification method and apparatus, a model training method and apparatus, and a device and a storage medium, which can be applied to various fields such as medical diagnostics and data classification. The method comprises: acquiring object data and Q pieces of candidate data; separately extracting object feature information of the object data and candidate feature information of the Q pieces of candidate data; on the basis of the object feature information and the candidate feature information, selecting, from the Q pieces of candidate data, at least one piece of target candidate data that meets a preset correlation condition with the object data; on the basis of the candidate feature information of the at least one piece of target candidate data, adjusting the object feature information; and performing classification processing on the adjusted object feature information, so as to obtain a classification result of the object data. The at least one piece of target candidate data obtained by means of screening is data of another object similar to the classification object, such that features of similar objects can be integrated, feature information of the classification object can be enriched, and the accuracy of data classification can be improved.

Description

Data classification and model training methods, devices, equipment and storage media

[0001] Priority information

[0002] This application claims priority to Chinese Patent Application No. 202411178661.3, filed on August 26, 2024, entitled “Data Detection and Model Training Method, Apparatus, Device and Storage Medium”, the entire contents of which are incorporated herein by reference. Technical Field

[0003] This application relates to the field of computer technology, and in particular to a data classification and model training method, apparatus, device and storage medium. Background Technology

[0004] With the rapid development of artificial intelligence (AI) and computer technology, various data detection models have been widely applied in various fields. In the medical field, accurate data detection can improve medical diagnosis. For example, early detection of cancer can significantly improve survival rates and reduce treatment costs, thus attracting widespread attention from academia and industry.

[0005] Current medical testing, such as early cancer detection, typically relies on specific cancer biological state parameters or image analysis methods. However, the results of current testing methods are not accurate enough. Summary of the Invention

[0006] This application provides a data classification and model training method, apparatus, device, and storage medium that can improve the accuracy of data classification, such as obtaining classification reference results for various cancers.

[0007] Firstly, this application provides a data classification method, including:

[0008] Obtain object data and Q candidate data, and extract object feature information from the object data and candidate feature information from the Q candidate data respectively. The object data includes parameter values ​​of L biological state parameters of the classified object, and the candidate data includes parameter values ​​of L biological state parameters of the reference object that is different from the classified object. Q and L are both positive integers greater than 1.

[0009] Based on the object feature information and the candidate feature information, at least one target candidate data that satisfies a preset correlation condition with the object data is selected from the Q candidate data;

[0010] Based on the candidate feature information of the at least one target candidate data, the object feature information is adjusted to obtain the adjusted object feature information of the object data;

[0011] The adjusted object feature information is classified to obtain the classification result of the object data. The classification result includes the probability reference value of the classified object belonging to multiple preset categories.

[0012] Secondly, this application provides a model training method, including:

[0013] A training sample and Q candidate samples are obtained, and the object feature information of the training sample and the candidate feature information of the Q candidate samples are extracted by the retrieval module in the data classification model. The training sample includes the parameter values ​​of L biological state parameters of the first object, and the candidate sample includes the parameter values ​​of L biological state parameters of the reference object that is different from the first object. Q and L are both positive integers greater than 1.

[0014] Based on the object feature information and the candidate feature information, at least one target candidate sample that satisfies the preset correlation condition with the training sample is selected from the Q candidate samples;

[0015] Based on the candidate feature information of the at least one target candidate sample, the object feature information is adjusted to obtain the adjusted object feature information of the training sample;

[0016] The adjusted object feature information is classified by the classification module in the data classification model to obtain the classification result of the training sample. The classification result includes the probability reference value of the first object belonging to multiple preset categories.

[0017] Based on the classification results of the training samples, the loss of the data classification model is determined, and the data classification model is trained based on the loss.

[0018] Thirdly, this application provides a data classification apparatus, comprising:

[0019] An extraction unit is used to acquire object data and Q candidate data, and to extract object feature information of the object data and candidate feature information of the Q candidate data respectively. The object data includes the state values ​​of L biological state parameters of the target object, and the candidate data are the state values ​​of L biological state parameters of other objects different from the target object. Q and L are both positive integers greater than 1.

[0020] The selection unit is configured to select at least one candidate data related to the object data from the Q candidate data based on the first feature information of the object data and the first feature information of the Q candidate data.

[0021] An adjustment unit is configured to adjust the first feature information of the object data based on the first feature information of the at least one candidate data, so as to obtain the adjusted feature information of the object data;

[0022] The classification unit is used to classify the adjusted feature information of the object data to obtain the classification result of the object data.

[0023] In some embodiments, the L biological state parameters include L gene mutations, and the classification result includes a probability reference value for the object data belonging to multiple cancer classifications.

[0024] In some embodiments, the classification unit specifically classifies the adjusted feature information of the object data using a data classification model. The data classification model includes a retrieval module and a classification module. The retrieval module includes R feature extraction units, where R is a positive integer greater than 1. The extraction unit is specifically used to extract features from the object data using the j-th feature extraction unit to obtain the j-th sub-object feature information of the object data, and to extract features from the Q candidate data using the j-th feature extraction units to obtain the j-th sub-candidate feature information of the Q candidate data, where j takes the value 1, 2, ..., R. The selection unit specifically... The system is configured to: select Kj target candidate data that satisfy a preset correlation condition with the object data from the Q candidate data based on R sub-object feature information of the object data and R sub-candidate feature information of the Q candidate data, where Kj is a positive integer; adjust the j-th sub-object feature information of the object data based on the j-th sub-candidate feature information of the Kj target candidate data to obtain the j-th adjusted sub-object feature information of the object data; and classify the R adjusted sub-object feature information of the object data through the classification module to obtain the classification result of the object data.

[0025] In some embodiments, the R feature extraction units are R encoders. The extraction unit is specifically used to extract features from the object data through the j-th encoder to obtain the j-th sub-object feature information of the object data, and to extract features from the Q candidate data through the j-th encoder to obtain the j-th sub-candidate feature information of the Q candidate data.

[0026] In some embodiments, the selection unit is specifically used to determine the similarity between the object data and each of the Q candidate data based on the R sub-object feature information of the object data and the R sub-candidate feature information of the Q candidate data, wherein the similarity is used to characterize the correlation between the object data and the candidate data; and to select the Kj target candidate data from the Q candidate data based on the similarity.

[0027] In some embodiments, the selection unit is specifically configured to, for the q-th candidate data among the Q candidate data, project the j-th sub-candidate feature information of the q-th candidate data using a first projection matrix to obtain the j-th projected feature information of the q-th candidate data, where the value of q is 1, 2, ..., Q; project the j-th sub-object feature information of the object data using the first projection matrix to obtain the j-th projected feature information of the object data; determine the distance between the j-th projected feature information of the object data and the j-th projected feature information of the q-th candidate data, and determine the similarity between the object data and the q-th candidate data based on the distance.

[0028] In some embodiments, the adjustment unit is specifically configured to, for the kth target candidate data among the Kj target candidate data, determine the distance between the jth projection feature information of the object data and the jth projection feature information of the kth target candidate data, and based on the distance, determine the first adjustment information of the object data by the kth target candidate data, where the value of k is 1, 2, ..., Kj; and adjust the jth sub-object feature information of the object data based on the first adjustment information of the object data by the Kj target candidate data respectively, to obtain the jth adjusted sub-object feature information of the object data.

[0029] In some embodiments, the adjustment unit is specifically used to perform linear activation processing on the distance and then random deactivation processing, and to perform lossless linear transformation on the random deactivation processing result to obtain the second adjustment information of the k-th target candidate data on the object data; based on the second adjustment information of the k-th target candidate data on the object data and the label information of the k-th target candidate data, to determine the first adjustment information of the k-th target candidate data on the object data.

[0030] In some embodiments, the adjustment unit is specifically used to project the label information of the kth target candidate data using a second projection matrix to obtain the label projection information of the kth target candidate data; and to determine the sum of the second adjustment information of the kth target candidate data on the object data and the label projection information of the kth target candidate data as the first adjustment information of the kth target candidate data on the object data.

[0031] In some embodiments, the adjustment unit is specifically configured to, for the kth target candidate data among the Kj target candidate data, determine the weight of the kth target candidate data based on the similarity between the object data and the kth target candidate data; based on the weights of the Kj target candidate data, perform weighted processing on the first adjustment information of the object data by the Kj target candidate data to obtain weighted adjustment information; and determine the jth adjusted sub-object feature information of the object data based on the jth sub-object feature information of the object data and the weighted adjustment information.

[0032] In some embodiments, the classification module includes a hybrid expert module, which includes a gating unit and E expert units. Specifically, the classification unit is used to obtain S input tokens based on R adjusted sub-object feature information of the object data, where S is a positive integer greater than 1; for the s-th input token among the S input tokens, the gating unit determines the probability value of the s-th input token being assigned to each of the E expert units, where E is a positive integer greater than 1, and s takes the value 1, 2, ..., S; based on the probability value, T expert units are selected from the E expert units to process the s-th input token, and the s-th input token is processed by the T expert units to obtain fused feature information of the s-th input token, where T is a positive integer less than E; based on the fused feature information of the S input tokens, hybrid feature information of the object data is obtained; based on the hybrid feature information, the object data is classified to obtain the classification result of the object data.

[0033] In some embodiments, the classification unit is specifically configured to, for each of the T expert units, perform feature fusion processing on each input token assigned to the expert unit to obtain output feature information of each input token, wherein each input token includes the s-th input token; determine the weights corresponding to the T expert units based on the probability value of the s-th input token being assigned to each of the T expert units; and perform weighted processing on the output feature information of the s-th input token output by the T expert units based on the weights corresponding to the T expert units to obtain the fused feature information of the s-th input token.

[0034] In some embodiments, the classification unit is specifically used to take each of the R adjusted sub-object feature information of the object data as a row of a first matrix to obtain a first matrix including R rows; and to take each column of the first matrix as an input token to obtain the S input tokens.

[0035] Fourthly, embodiments of this application provide a model training apparatus, comprising:

[0036] An extraction unit is used to acquire training samples and Q candidate samples, and to extract object feature information of the training samples and candidate feature information of the Q candidate samples through the retrieval module in the data classification model. The training samples include parameter values ​​of L biological state parameters of a first object, and the candidate samples include parameter values ​​of L biological state parameters of a reference object that is different from the first object. Q and L are both positive integers greater than 1.

[0037] The selection unit is used to select at least one target candidate sample from the Q candidate samples that satisfies a preset correlation condition with the training sample, based on the object feature information and the candidate feature information.

[0038] An adjustment unit is used to adjust the object feature information based on the candidate feature information of the at least one target candidate sample to obtain the adjusted object feature information of the training sample.

[0039] The classification unit is used to classify the adjusted object feature information of the training sample through the classification module in the data classification model to obtain the classification result of the training sample. The classification result includes the probability reference value of the first object belonging to multiple preset categories.

[0040] The training unit is used to determine the loss of the data classification model based on the classification results of the training samples, and to train the data classification model based on the loss.

[0041] In some embodiments, the retrieval module includes R feature extraction units, where R is a positive integer greater than 1; the object feature information includes R sub-object feature information; and the candidate feature information includes R sub-candidate feature information. The extraction unit is specifically configured to, for the j-th feature extraction unit among the R feature extraction units, extract features from the training sample to obtain the j-th sub-object feature information of the training sample, and extract features from the Q candidate samples respectively through the j-th feature extraction unit to obtain the j-th sub-candidate feature information of the Q candidate samples, where j takes values ​​of 1, 2, ..., R. The selection unit is specifically used to select Kj target candidate samples from the Q candidate samples that satisfy a preset correlation condition with the training sample, based on the R sub-object feature information of the training sample and the R sub-candidate feature information of the Q candidate samples; the adjustment unit is specifically used to adjust the j-th sub-object feature information of the training sample based on the j-th sub-sub-sub-feature information of the Kj target candidate samples to obtain the j-th adjusted sub-object feature information of the training sample; the classification unit is specifically used to classify the R adjusted sub-object feature information of the training sample through the classification module to obtain the classification result of the training sample.

[0042] In some embodiments, the classification module includes a hybrid expert module, which includes a gating unit and E expert units. Specifically, the classification unit is used to obtain S input tokens based on R adjusted sub-object feature information of the object data, where S is a positive integer greater than 1; for the s-th input token among the S input tokens, the gating unit determines the probability value of the s-th input token being assigned to each of the E expert units, where E is a positive integer greater than 1 and s is a positive integer less than or equal to S; based on the probability value, T expert units are selected from the E expert units to process the s-th input token, and the s-th input token is processed by the T expert units to obtain fused feature information of the s-th input token, where T is a positive integer less than E; based on the fused feature information of the S input tokens, hybrid feature information of the object data is obtained; based on the hybrid feature information, the object data is classified to obtain the classification result of the object data.

[0043] In some embodiments, the training unit is specifically configured to determine a first loss based on the classification results of the training samples and the label information of the training samples; determine a second loss based on the number of input labels assigned to each of the E expert units and the assignment probability value; and determine the loss of the data classification model based on the first loss and the second loss.

[0044] Fifthly, this application provides an electronic device including a processor and a memory. The memory is used to store a computer program, and the processor is used to invoke and run the computer program stored in the memory to perform the methods described in the first or second aspect above.

[0045] In a sixth aspect, a chip is provided for implementing the methods of various implementations of the first aspect described above. Specifically, the chip includes a processor for retrieving and running a computer program from a memory, causing a device equipped with the chip to perform the methods of the first or second aspect described above.

[0046] In a seventh aspect, a computer-readable storage medium is provided for storing a computer program that causes a computer to perform the methods described in the first or second aspect.

[0047] Eighthly, a computer program product is provided, including computer program instructions that cause a computer to perform the methods described in the first or second aspect.

[0048] Ninthly, a computer program is provided that, when run on a computer, causes the computer to perform the methods of the first or second aspect described above.

[0049] In summary, this application uses a data classification model for data classification, which includes a retrieval module and a classification module. Specifically, during data classification, object data and Q candidate data are first acquired. Object feature information of the object data and candidate feature information of the Q candidate data are extracted respectively. The object data includes parameter values ​​of L biological state parameters of the detected object, and the candidate data includes parameter values ​​of L biological state parameters of reference objects different from the classified object. Next, based on the object feature information of the object data and the candidate feature information of the Q candidate data, at least one target candidate data that satisfies a preset correlation condition with the object data is selected from the Q candidate data. Then, based on the candidate feature information of this at least one target candidate data, the object feature information of the object data is adjusted to obtain the adjusted object feature information of the object data. Finally, the adjusted object feature information of the object data is classified to obtain the classification result of the object data. The classification result includes a probability reference value of the classified object belonging to multiple preset categories. In this embodiment, at least one target candidate data selected from Q candidate data is data of other objects similar to the classification object. The candidate feature information of the at least one target candidate data is used to adjust the object feature information of the object data, which can integrate the features of similar objects, making the adjusted object feature information of the object data richer and more comprehensive. Therefore, when using the adjusted object feature information of the object data for data classification, it can cover multiple different classifications. For example, it can obtain classification reference results for multiple different types of early cancer and improve classification accuracy. Attached Figure Description

[0050] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0051] Figure 1 is a schematic diagram of the implementation environment of a data classification method provided in an embodiment of this application;

[0052] Figure 2 is a schematic flowchart of a model training method provided in an embodiment of this application;

[0053] Figure 3 is a schematic diagram of a data classification model;

[0054] Figure 4 is a schematic diagram of a training sample and a candidate sample;

[0055] Figure 5 is a schematic diagram of a retrieval model;

[0056] Figure 6A is a schematic diagram of selecting relevant candidate samples based on similarity;

[0057] Figure 6B is a schematic diagram of adjusting the features of training samples based on selected candidate samples;

[0058] Figure 7 is a schematic diagram of data classification using the classification module;

[0059] Figure 8 is a schematic diagram of the retrieval module including R classification units;

[0060] Figure 9 is a schematic flowchart of a model training method provided in an embodiment of this application;

[0061] Figure 10 is a schematic diagram of a feature extraction unit;

[0062] Figure 11 is a schematic diagram of the R adjusted feature information obtained using the classification module;

[0063] Figure 12 is a schematic diagram of a network structure for a classification module;

[0064] Figure 13 is a schematic diagram of data classification using a hybrid expert module;

[0065] Figure 14 is a flowchart illustrating a data classification method provided in an embodiment of this application;

[0066] Figure 15 is a flowchart illustrating a data classification method provided in an embodiment of this application;

[0067] Figure 16 is a flowchart illustrating a data classification method provided in an embodiment of this application;

[0068] Figure 17 is a schematic diagram of the overall data classification process;

[0069] Figure 18 is a schematic block diagram of a data classification device provided in an embodiment of this application;

[0070] Figure 19 is a schematic block diagram of a model training apparatus provided in an embodiment of this application;

[0071] Figure 20 is a schematic block diagram of an electronic device provided in an embodiment of this application. Detailed Implementation

[0072] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0073] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. In embodiments of the invention, "B corresponding to A" means that B is associated with A. In one implementation, B can be determined based on A. However, it should also be understood that determining B based on A does not mean determining B solely based on A; B can also be determined based on A and / or other information. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or server that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to these processes, methods, products, or devices. In the description of this application, unless otherwise stated, "a plurality of" means two or more.

[0074] The data classification method provided in this application can be applied to various fields such as medical diagnosis and data detection. It can achieve multi-dimensional classification of object data and improve classification accuracy. For example, it can obtain classification reference results for various early cancers, thereby improving the detection accuracy of early cancers.

[0075] To facilitate understanding of the embodiments of this application, the relevant concepts involved in the embodiments of this application will first be introduced:

[0076] In medical testing, early cancer detection can significantly improve survival rates and reduce treatment costs, thus attracting widespread attention from academia and industry. Currently, early cancer detection methods mainly include: methods based on specific cancer biological state parameters (also known as biomarkers) and methods based on image analysis. Methods based on specific cancer biological state parameters rely on known cancer-related biological state parameters for medical testing. Currently available specific biological state parameters for cancer detection include prostate-specific antigen (PSA) for prostate cancer detection and CA-125 for ovarian cancer detection. Image analysis-based methods utilize various imaging techniques, such as low-dose computed tomography (CT) for lung cancer detection, colonoscopy for colorectal cancer detection, and cervical smears for cervical cancer detection, to assist in the detection of related cancers.

[0077] However, the aforementioned methods based on specific cancer biological state parameters are limited to specific types of cancer, and the relevant biological state parameters for these cancers are clearly known. They lack sensitivity and specificity and cannot detect cancers associated with unknown cancer biological state parameters. Furthermore, image-based analysis methods can only target specific cancers and cannot be applied to convenient, low-cost early cancer screening. They cannot perform pan-cancer analysis, and some early stages of cancer do not have obvious imaging features, resulting in low detection rates using image-based analysis methods.

[0078] To address the aforementioned technical problems, this application proposes a data classification method. During data classification, firstly, object data of the object to be classified and Q candidate data of Q reference objects are obtained. Object feature information of the object data and candidate feature information of the Q candidate data are extracted respectively. The object data includes parameter values ​​of L biological state parameters of the object to be classified, and the candidate data includes parameter values ​​of L biological state parameters of the reference objects. The object to be classified differs from the reference objects. Next, based on the object feature information and the candidate feature information, at least one target candidate data that satisfies a preset correlation condition with the object data is selected from the Q candidate data. Then, based on the candidate feature information of this at least one target candidate data, the object feature information is adjusted to obtain adjusted object feature information of the object data. Finally, the adjusted object feature information of the object data is classified to obtain the classification result of the object data. This classification result includes a probability reference value indicating whether the object belongs to multiple preset categories. In this embodiment, at least one target candidate data selected from Q candidate data is data of other reference objects similar to the classification object. The candidate feature information of the at least one target candidate data is used to adjust the object feature information of the object data, which can integrate the features of similar objects, making the adjusted object feature information of the object data richer and more comprehensive. Therefore, when using the adjusted object feature information of the object data for data classification, it can cover multiple different classifications. For example, it can obtain classification reference results for multiple different types of early cancer, and can improve classification accuracy.

[0079] The implementation environment of the data classification method provided in the embodiments of this application is described below.

[0080] Figure 1 is a schematic diagram of the implementation environment of a data classification method provided in an embodiment of this application. As shown in Figure 1, the implementation environment includes: a terminal device 101 and a server 102.

[0081] The terminal device 101 is connected to the server 102 via wired or wireless means.

[0082] In some embodiments, a client for a data classification system is installed on the terminal device 101, and the server 102 can be understood as the server-side or backend of the data classification system. Users can interact with the client of the data classification system installed on the terminal device 101. The server 102 stores a data classification model used for data classification, such as cancer classification in early cancer screening.

[0083] In some embodiments, the data classification model is trained before data classification. The process of training the data classification model using server 102 is described below. Specifically, a user can specify a training dataset to server 102 via a terminal device. Server 102 obtains training samples and Q candidate samples from the training dataset. In this embodiment, both the training samples and candidate samples include parameter values ​​for M biological state parameters (e.g., gene mutation) of different objects. Taking gene mutation as an example, the parameter value for gene mutation includes two values: mutated and non-mutated. For example, for a certain gene, if the gene mutates, the parameter value for gene mutation is 1; if the gene does not mutate, the parameter value for gene mutation is 0. Server 102 extracts object feature information from training samples and candidate feature information from Q candidate samples through the retrieval module in the data classification model. The training samples include parameter values ​​of L biological state parameters of a first object, and the candidate samples include state values ​​of L biological state parameters of a reference object different from the first object. Q and M are both positive integers greater than 1. Based on the object feature information and the candidate feature information, at least one target candidate sample that meets the preset correlation condition with the training samples is selected from the Q candidate samples. Based on the candidate feature information of the at least one target candidate sample, the object feature information is adjusted to obtain the adjusted object feature information of the training samples. The adjusted object feature information of the training samples is classified through the classification module in the data classification model to obtain the classification result of the training samples. Based on the classification result of the training samples, the loss of the data classification model is determined, and the data classification model is trained based on the loss.

[0084] In some embodiments, the above-described data classification model can also be trained by the terminal device 101, and this application embodiment does not limit this.

[0085] In this embodiment of the application, after the data classification model is trained, the trained data classification model is used to execute the data classification method of this embodiment of the application.

[0086] In some embodiments, the data classification method provided in this application is jointly executed by terminal device 101 and server 102. For example, a user starts a client of the data classification system on terminal device 101. The client displays an input interface for the data to be classified. The user inputs the object data to be processed in the input interface. The object data is the parameter values ​​of L biological state parameters of the object to be classified, such as the parameter values ​​of gene mutations in L of the object to be classified. Terminal device 101 sends the object data to be classified to server 102. Server 102 loads a candidate dataset and obtains Q candidate data from the candidate dataset. Next, the object feature information of the object data and the candidate feature information of Q candidate data are extracted respectively. For example, server 102 extracts the object feature information of the object data and the candidate feature information of Q candidate data through the retrieval module in the data classification model. The object data includes the parameter values ​​of L biological state parameters of the classified object, and the candidate data includes the parameter values ​​of L biological state parameters of reference objects different from the classified object. Q and L are both positive integers greater than 1. Based on the object feature information of the object data and the candidate feature information of the Q candidate data, at least one target candidate data that meets the preset correlation condition with the object data is selected from the Q candidate data. Based on the candidate feature information of this at least one target candidate data, the object feature information of the object data is adjusted to obtain the adjusted object feature information of the object data. The adjusted object feature information of the object data is then classified to obtain the classification result of the object data. For example, the adjusted object feature information of the object data is classified through the classification module in the data classification model to obtain the classification result of the object data. Then, server 102 sends the classification result of the object data (e.g., the predicted value or probability reference value of the object belonging to multiple preset categories) to terminal device 101 for display. In this embodiment, at least one target candidate data is selected from Q candidate data. This target candidate data is data of other reference objects similar to the first object. The candidate feature information of the target candidate data is used to adjust the object feature information of the object data, integrating the features of similar objects. This makes the adjusted object feature information of the object data richer and more comprehensive. The adjusted object feature information of the object data is then used for data classification, which can cover multiple classifications and improve the accuracy of object data classification.

[0087] In some embodiments, the data classification method provided in this application is executed by terminal device 101. In this case, terminal device 101 can be any terminal device capable of providing data classification services. Server 102 can be any server capable of providing data download services. In one example, the server 102 stores a trained data classification model. A client for the data classification system is installed on terminal device 101, which can request the data classification model from server 102 based on a user's request. For example, terminal device 101 displays an input box for data to be processed, and the user inputs object data to be processed in this input interface. In response to the user's input operation on the object data, terminal device 101 sends a request to server 102 to request a data classification model. Server 102 sends the trained data classification model to terminal device 101. Terminal device 101 loads a candidate dataset and selects Q candidate data from it. Then, it extracts the object feature information of the object data and the candidate feature information of the Q candidate data, for example, through the retrieval module in a data classification model. The object data includes the parameter values ​​of L biological state parameters of the classified object, and the candidate data includes the parameter values ​​of L biological state parameters of a reference object different from the classified object. Both Q and L are positive integers greater than 1. Based on this object feature information and the candidate feature information, at least one target candidate data that satisfies a preset correlation condition with the object data is selected from the Q candidate data. Based on the candidate feature information of this at least one target candidate data, the object feature information of the object data is adjusted to obtain adjusted object feature information. The adjusted object feature information of the object data is then classified to obtain the classification result, for example, through the classification module in a data classification model. Finally, terminal device 101 displays the classification result of the object data to the object, which includes a probability reference value indicating whether the classified object belongs to multiple preset categories.

[0088] In some embodiments, the terminal device 101 includes, but is not limited to, desktop computers, laptops, smartphones, tablets, IoT devices, and portable wearable devices. IoT devices may include smart speakers, smart TVs, smart air conditioners, and smart in-vehicle systems. Portable wearable devices may include smartwatches, smart bracelets, and head-mounted devices. Terminal devices are often equipped with a display device, which may be a monitor, display screen, touchscreen, etc., and the touchscreen may be a touchscreen, touch panel, etc.

[0089] In some embodiments, the server 102 described above can be one or more servers. When there are multiple servers, at least two servers are used to provide different services, and / or at least two servers are used to provide the same service, such as providing the same service in a load-balanced manner. This application embodiment does not limit this. The server described above can be an independent physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN (Content Delivery Network), and big data and artificial intelligence platforms. The server can also be a node in a blockchain.

[0090] It should be noted that the implementation environment of this application embodiment includes, but is not limited to, the one shown in FIG1.

[0091] The technical solutions of the embodiments of this application will be described in detail below through some examples. The following embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0092] First, referring to Figure 2, the training process of the data classification model in this application embodiment will be introduced.

[0093] Figure 2 is a schematic flowchart of a model training method provided in an embodiment of this application. The execution entity of this embodiment can be a device with model training functionality, such as a model training apparatus. This model training apparatus can be an electronic device such as the server 102 or terminal device 101 shown in Figure 1, or it can be a system composed of the server 102 or 101 shown in Figure 1. For ease of description, the following embodiments use an electronic device as an example to illustrate the method of this embodiment.

[0094] As shown in Figure 2, the model training method of this application embodiment includes:

[0095] S101. Obtain training samples and Q candidate samples, and extract the object feature information of the training samples and the candidate feature information of the Q candidate samples through the retrieval module in the data classification model.

[0096] The training samples include the parameter values ​​of L biological state parameters of the first object, and the candidate samples include the parameter values ​​of L biological state parameters of reference objects different from the first object. For example, the i-th candidate sample among Q candidate samples can be understood as the parameter values ​​of L biological state parameters of the i-th reference object, which is different from the first object. Q and L are both positive integers greater than 1, and i takes the values ​​1, 2, ..., Q.

[0097] It should be noted that the data used in the embodiments of this application and the process of obtaining this data comply with the relevant laws and regulations.

[0098] The data classification model of this application can be applied to different fields to classify data from those fields. For example, it can be applied to the biological and medical fields to classify data related to these fields, such as for auxiliary detection of early-stage cancer.

[0099] In the fields of biology and medicine, the first object and the reference object can represent different patients. Biological status parameters can include parameters such as serological markers and methylation status. Taking gene mutation as an example, the biological status parameter can include the parameter values ​​of L gene mutations of patient 1, and the Q candidate samples can be the parameter values ​​of L gene mutations of Q patients. The method of this embodiment can detect the probability reference value of patient 1 belonging to multiple cancer categories.

[0100] In the embodiments of this application, when the data classification model is used in different fields, the data classification model can be trained using training samples from the corresponding fields.

[0101] In this embodiment of the application, before training the data classification model, the electronic device first acquires a training dataset, which includes N training samples, which can be represented as {S1, S2, ..., S...}. N Each training sample contains the parameter values ​​of L possible biological state parameters of an object. For example, the i-th training sample contains the parameter values ​​of the L biological state parameters of the i-th object, and this i-th training sample can be represented as... in The parameter value represents the first biological state parameter of the i-th object. This represents the parameter value of the second biological state parameter of the i-th object, and so on. The parameter value that expresses the Lth biological state parameter of the i-th object.

[0102] For example, the N training samples included in the training dataset can be shown in Table 1:

[0103] Table 1

[0104] Each row in Table 1 above represents a training sample. For example, the first row of Table 1 is the training sample S1, which includes the parameter values ​​of L biological state parameters of object 1. The second row in Table 1 above is training sample S2, which includes parameter values ​​for L biological state parameters of object 2. For example, the training dataset in this application embodiment can be formed into an N-row L-column matrix, where each row of the matrix is ​​a training sample.

[0105] The embodiments of this application do not limit the specific types of the above-mentioned L biological state parameters.

[0106] In one possible scenario, the aforementioned L biological state parameters may include gene mutations, where the parameter value of a gene mutation can be understood as a state value indicating whether a gene has mutated. In this case, the state value includes two possible values: mutated and not mutated. Optionally, 1 can be used to represent a mutation and 0 to represent no mutation; alternatively, 1 can be used to represent no mutation and 0 to represent a mutation. This application does not limit the specific type and number of gene mutations. For example, gene mutations may include mutations in genes such as ABL1, ACVR1, and AKT1. ABL1 is a proto-oncogene, ACVR1 encodes the activin type I receptor gene, and AKT1 is a thymomavirus proto-oncogene.

[0107] In one possible scenario, the aforementioned L biological state parameters may also include serological markers and methylation status, etc. In this case, the parameter values ​​of the biological state parameters can be determined based on the specific circumstances, such as the detection values ​​of specific biological state parameters.

[0108] In this embodiment, label information also needs to be determined for the aforementioned N training samples. For each training sample, the label information represents the true value of the training sample belonging to the preset category. Assuming that the preset categories include C categories, the label information of the N training samples can be defined as {Y1, Y2, ..., Y...} N}, where Y i ∈{0,…,C}, where C represents the number of preset categories, and the specific value of C varies depending on the different training datasets.

[0109] In some embodiments, the above-mentioned L biological state parameters include L gene mutations, and the corresponding preset classifications include multiple cancer classifications.

[0110] The data classification model in this application attempts to construct a mapping function Y. i =F(S) i This method allows for the classification of any object into multiple preset categories based on the parameter values ​​of L biological state parameters. For example, if the preset classification includes multiple cancer categories, then this embodiment of the application is used to develop a highly sensitive and specific multi-cancer early detection (MCED) model. The biological state parameters may involve parameters such as gene mutations, serological markers, and methylation status.

[0111] In this embodiment of the application, the training process of the data classification model using each training sample in the above-mentioned training dataset is basically the same. For ease of description, the model training process of this embodiment of the application will be introduced here using a training sample as an example.

[0112] In this embodiment of the application, for each training sample in the training dataset, when using the training sample to train the data classification model, it is necessary to obtain Q candidate samples corresponding to the training sample.

[0113] This application does not restrict the specific selection method for obtaining the Q candidate samples corresponding to the training sample in the embodiments.

[0114] In one possible implementation, for each of the N training samples, for example, the i-th training sample S i The electronic device extracts N training samples from the training dataset mentioned above, excluding the training sample S. i From the N-1 training samples other than the given samples, select Q training samples as the training sample S. i The corresponding Q candidate samples. For example, if Q is less than N-1, then Q training samples can be randomly selected from these N-1 training samples as the training sample S. i The corresponding Q candidate samples. For example, if Q equals N-1, then the remaining N-1 training samples in the training dataset, excluding the training sample Si, can be used as the training sample S. i The corresponding Q candidate samples.

[0115] In one possible implementation, when training the data classification model, multiple rounds of training are performed. In each round, a batch of training samples is selected from N training samples for the current round of training. In this case, for each round of training, the training sample S... i This can be understood as any training sample in the current batch of training samples. In this case, the Q candidate samples include all training samples in the current batch except for the given training sample S. i All samples outside of this.

[0116] In one possible implementation, the electronic device can also load other datasets, which include multiple candidate samples, where different candidate samples include parameter values ​​of L biological state parameters of different objects. In this way, the electronic device can select Q candidate samples corresponding to the training sample from this dataset.

[0117] In this embodiment of the application, after the electronic device obtains the training sample and the Q candidate samples corresponding to the training sample, it uses the training sample and the Q candidate samples to train the data classification model.

[0118] As shown in Figure 3, the data classification model of this embodiment includes a retrieval module and a classification module. The main function of the retrieval module is to select at least one target candidate sample related to the training sample from Q candidate samples. Optionally, when determining whether a candidate sample is related to the training sample, it can be determined by judging whether the candidate sample and the training sample meet a preset correlation condition. Then, the object feature information of the training sample is adjusted using the at least one target candidate sample. By adjusting the features of the training sample and integrating the features of other related candidate samples, the accuracy of data classification can be improved when using the adjusted features of the training sample for subsequent data classification. The classification module is used to classify the adjusted features of the training sample.

[0119] In some embodiments, for ease of description, the currently used training sample can be represented as S. input The S input This includes the parameter values ​​of L biological state parameters for a specific object, such as the first object. The training sample S... input The corresponding Q candidate samples can be represented as S 1 candidate S 2 candidate …S Q candidate The Q candidate samples include the parameter values ​​of L biological state parameters from Q other reference objects that are different from the first object. For example, Figure 4 shows the training sample S when L equals 6 and Q equals 5. input The diagram below illustrates the training and candidate samples. In Figure 4, each small square represents an element value, specifically the parameter value of a biological state parameter for a given object. Different colors of the squares indicate different parameter values. Taking gene mutation as an example, a dark square (e.g., black) indicates a mutated gene, while a light square (e.g., white) indicates no mutation. Conversely, a dark square indicates no mutation, and a light square indicates a mutated gene.

[0120] As shown in Figure 4, the electronic device inputs the training sample and Q candidate samples into the retrieval module, which extracts the object feature information of the training sample and the candidate feature information of each candidate sample.

[0121] In this embodiment of the application, the specific methods by which the retrieval module extracts the feature information of the training sample objects and the candidate feature information of Q candidate samples include at least the following:

[0122] In Method 1, the retrieval module does not consider the Q candidate samples when extracting object feature information from the training samples. Similarly, when extracting candidate feature information for each of the Q candidate samples, it does not consider the training samples or other candidate samples among the Q candidate samples. Method 1 can be understood as the electronic device inputting the training samples and the Q candidate samples (a total of Q+1 samples) one by one into the retrieval module for feature extraction. During the feature extraction process, these samples do not interact or fuse features.

[0123] In Method 2, the retrieval module considers Q candidate samples when extracting object feature information from the training samples. However, when extracting candidate feature information for each of the Q candidate samples, it does not consider the training samples or other candidate samples among the Q candidate samples. Method 2 can be understood as the electronic device inputting the training samples and the Q candidate samples (a total of Q+1 samples) into the retrieval module for feature extraction. During the feature extraction process, these samples interact or fuse features.

[0124] The embodiments of this application do not impose restrictions on the specific network structure of the retrieval module.

[0125] In one example, as shown in Figure 5, the retrieval module includes an encoder for feature extraction. Thus, the electronic device can input training samples and Q candidate samples into the encoder for feature extraction, obtaining object feature information from the training samples and candidate feature information from each of the Q candidate samples.

[0126] For example, the process by which an electronic device extracts object feature information from training samples and candidate feature information from Q candidate samples through an encoder can be represented by the following formula (1): h input =E(S) input h i candidate =E(S) i candidate ),i∈Q candidate (1)

[0127] Where E represents the encoder, h input ∈R 1×d S represents the training sample input Object feature information, h i candidate ∈R 1xd S represents the i-th candidate sample. i candidate Candidate feature information, Q candidate This represents the candidate sample set consisting of Q candidate samples. d represents the hidden dimension, that is, the object feature information h of the training samples. inputFor a d-dimensional vector, similarly, the candidate feature information h of the i-th candidate sample... i candidate It is also a d-dimensional vector.

[0128] In this embodiment of the application, the electronic device obtains a training sample and Q candidate samples corresponding to the training sample based on the above steps, and extracts the object feature information of the training sample and the candidate feature information of each of the Q candidate samples through the retrieval module in the data classification model, and then executes the following step S102.

[0129] S102. Based on object feature information and candidate feature information, select at least one target candidate sample from Q candidate samples that meets the preset correlation condition with the training sample.

[0130] In this embodiment of the application, in order to improve the accuracy of data classification, at least one target candidate sample that meets the preset correlation condition with the training sample is selected from Q candidate samples. This target candidate sample is also called a sample related to the training sample. Then, the features of the training sample are adjusted using the features of the at least one target candidate sample. In this way, the adjusted features of the training sample not only include the features of the first object, but also the features of other objects similar to the first object, thereby making the features of the first object richer and more comprehensive. Thus, when data classification is performed based on the rich and comprehensive feature information of the first object, the accuracy of data classification can be improved.

[0131] The following describes the specific process by which an electronic device selects at least one target candidate sample that is related to the training sample from Q candidate samples.

[0132] In this embodiment of the application, the electronic device selects at least one target candidate sample related to the training sample from the Q candidate samples based on the object feature information of the training sample and the candidate feature information of each of the Q candidate samples.

[0133] The preset relevance condition is used to determine whether there is a correlation between candidate samples and training samples. This preset relevance condition can be whether the clustering results of the candidate samples and training samples belong to the same cluster, or whether the similarity between the candidate samples and training samples meets a set condition (e.g., greater than a preset similarity threshold, or selecting the top N candidate samples with higher similarity from Q candidate samples). Essentially, clustering algorithms classify samples by measuring similarity in the feature space; samples belonging to the same cluster have higher similarity. Similarity is used to characterize the correlation between candidate samples and training samples; the higher the similarity, the stronger the correlation between them.

[0134] In some embodiments, the electronic device clusters the object feature information of the training samples and the candidate feature information of Q candidate samples. Candidate samples belonging to the same cluster as the training samples are identified as target candidate samples that satisfy a preset correlation condition with the training samples, i.e., target candidate samples related to the training samples. Specifically, training samples and candidate samples belonging to the same cluster have high similarity; therefore, candidate samples belonging to the same cluster as the training samples are identified as target candidate samples related to the training samples.

[0135] In some embodiments, S102 above includes the following steps S102-A and S102-B:

[0136] S102-A: Based on the object feature information of the training samples and the candidate feature information of Q candidate samples, determine the similarity between the training samples and each of the Q candidate samples;

[0137] S102-B: Based on the similarity between the training samples and Q candidate samples, select at least one target candidate sample from the Q candidate samples.

[0138] In this implementation, for each of the Q candidate samples, the similarity between the candidate sample and the training sample is determined based on the candidate feature information of the candidate sample and the object feature information of the training sample.

[0139] This application does not limit the specific method by which an electronic device determines the similarity between a candidate sample and a training sample based on the candidate feature information of the candidate sample and the object feature information of the training sample.

[0140] In one possible implementation, the electronic device uses cosine distance or L2 distance to calculate the distance between the object feature information of the candidate sample and the candidate feature information of the training sample, and then determines the similarity between the candidate sample and the training sample based on this distance. A larger distance indicates less similarity between the training sample and the candidate sample, resulting in a smaller similarity value. Conversely, a smaller distance indicates greater similarity between the training sample and the candidate sample, resulting in a larger similarity value. In one example, the reciprocal of the distance between the candidate feature information of the candidate sample and the object feature information of the training sample can be used to determine the similarity between the candidate sample and the training sample.

[0141] In one possible implementation, S102-A may include the following steps S102-A1 to S102-A3:

[0142] S102-A1. For the qth candidate sample among Q candidate samples, the candidate feature information of the qth candidate sample is projected using the first projection matrix to obtain the projected feature information of the qth candidate sample, where the value of q is 1, 2, ..., Q.

[0143] S102-A2: Project the object feature information of the training samples using the first projection matrix to obtain the projected feature information of the training samples;

[0144] S102-A3. Determine the distance between the projection feature information of the training sample and the projection feature information of the q-th candidate sample, and use it as the similarity between the training sample and the q-th candidate sample.

[0145] In this implementation, when calculating the similarity between candidate samples and training samples, a projection matrix is ​​used to project the candidate feature information of the candidate samples and the object feature information of the training samples to obtain the projected feature information of the candidate samples and training samples. Then, the similarity between the candidate samples and training samples is calculated using the projected feature information.

[0146] Specifically, for the q-th candidate sample out of Q candidate samples, the first projection matrix W is used. k The candidate feature information of the q-th candidate sample is projected. For ease of description, this projected candidate feature information is denoted as the projected feature information of the q-th candidate sample. Similarly, the first projection matrix W is used. k The object feature information of the training sample is projected to obtain the projected feature information of the training sample. Then, the distance between the projected feature information of the training sample and the projected feature information of the q-th candidate sample is calculated, and this distance is determined as the similarity between the training sample and the q-th candidate sample.

[0147] In one example, the electronic device uses the L2 distance as the similarity between the projected feature information of the training sample and the projected feature information of the q-th candidate sample. The L2 distance is also called the Euclidean distance. The similarity calculation formula is shown in formula (2):

[0148] in, W represents the similarity between the projected feature information of the training sample and the projected feature information of the q-th candidate sample. k ∈R d×d Let be the first projection matrix. W represents the candidate feature information for the q-th candidate sample. k (h input ) represents the projected feature information of the training samples. This represents the projected feature information of the q-th candidate sample.

[0149] As shown in Figure 6A, the electronic device can determine the similarity between each of the Q candidate samples and the training sample based on the above method. Then, based on the similarity between each of the Q candidate samples and the training sample, the electronic device selects at least one target candidate sample related to the training sample from the Q training samples. In one example, the electronic device selects at least one candidate sample with the highest similarity to the training sample from the Q candidate samples (e.g., the top k most similar candidate samples shown in Figure 6), as at least one target candidate sample related to the training sample.

[0150] S103. Based on the candidate feature information of at least one target candidate sample, adjust the object feature information of the training sample to obtain the adjusted object feature information of the training sample.

[0151] In this embodiment, adjusting the object feature information of the training samples may include expanding (or supplementing) the object feature information of the training samples to enrich the object feature information of the training samples, so that the adjusted object feature information includes not only the feature information of the first object, but also the feature information of other objects related to the first object. Adjusting the object feature information of the training samples may also include modifying or deleting the object features of the training samples, such as deleting or modifying some erroneous feature information. By deleting or modifying erroneous feature information, the adjusted object feature information becomes more accurate.

[0152] This application embodiment does not limit the specific method by which an electronic device adjusts the object feature information of a training sample based on the candidate feature information of at least one candidate sample to obtain the adjusted object feature information of the training sample.

[0153] In some embodiments, the electronic device can fuse the object feature information of the training sample with the candidate feature information of the at least one target candidate sample to obtain the adjusted object feature information of the training sample. For example, based on the similarity between the at least one target candidate sample and the training sample, the weight of the at least one target candidate sample is determined, and then based on the weight, the object feature information of the training sample is weighted with the candidate feature information of the at least one target candidate sample to obtain the adjusted object feature information of the training sample.

[0154] In some embodiments, the electronic device can compare and analyze the object feature information of the training sample with the candidate feature information of at least one target candidate sample to determine whether the object feature information of the training sample needs to be modified or deleted. If modification or deletion is required, the object feature information of the training sample is deleted or modified according to the candidate feature information of the at least one target candidate sample to obtain the adjusted object feature information of the training sample. By deleting or modifying erroneous feature information, the adjusted object feature information becomes more accurate.

[0155] In some embodiments, S103 above includes the following steps S103-A and S103-B:

[0156] S103-A. For the kth target candidate sample in at least one target candidate sample, determine the distance between the projection feature information of the training sample and the projection feature information of the kth target candidate sample, and based on the distance, determine the first adjustment information of the kth target candidate sample to the training sample. The value of k is 1, 2, ..., K, where K is the number of target candidate samples that meet the preset correlation condition with the training sample, that is, the number of target candidate samples related to the training sample.

[0157] S103-B: Based on the first adjustment information of the training samples from at least one target candidate sample, adjust the object feature information of the training samples to determine the adjusted object feature information of the training samples.

[0158] In this implementation, the adjustment information of the candidate sample to the training sample is determined in the hidden layer of the data classification model using the distance between the training sample and at least one target candidate sample. Specifically, for the k-th target candidate sample among the at least one target candidate sample, the distance between the projected feature information of the training sample and the projected feature information of the k-th target candidate sample is determined, and based on this distance, the first adjustment information of the k-th target candidate sample to the training sample is determined.

[0159] This application does not limit the specific method by which the first adjustment information of the training sample for the k-th target candidate sample is determined based on the distance.

[0160] In some embodiments, the electronic device can determine the distance between the projection feature information of the training sample and the projection feature information of the kth target candidate sample as the first adjustment information of the kth target candidate sample to the training sample.

[0161] In some embodiments, the electronic device performs activation processing on the distance between the projection feature information of the training sample and the projection feature information of the k-th target candidate sample, followed by random deactivation processing. A lossless linear transformation is then performed on the random deactivation result to obtain the projection adjustment information of the k-th target candidate sample relative to the training sample. This second adjustment information of the k-th target candidate sample relative to the training sample is then determined as the first adjustment information of the k-th target candidate sample relative to the training sample.

[0162] For example, the electronic device uses the following formula (3) to determine the second adjustment information:

[0163] in, This refers to the projection adjustment information of the k-th target candidate sample onto the training samples. For the k-th candidate sample, W is the L1 distance between the projected feature information of the training sample and the second feature information of the k-th candidate sample. T (·) = LinearWithoutBias(Dropout(ReLU(Linear(·)))), where Linear(·) is the linear processing function, ReLU is the activation function, Dropout(·) is the random deactivation function, and LinearWithoutBias(·) is the lossless linear transformation. The L1 distance is also known as the Manhattan distance.

[0164] It should be noted that the above method of processing the distance between the projection feature information of the training sample and the projection feature information of the kth target candidate sample to obtain the second adjustment information of the kth target candidate sample on the training sample is only one example. Electronic devices can also process the distance between the projection feature information of the training sample and the projection feature information of the kth candidate sample to obtain the second adjustment information of the kth target candidate sample on the training sample in other ways.

[0165] In some embodiments, since at least one target candidate sample is similar to the training sample, the label information of at least one target candidate sample can also provide additional information for the features of the training sample. Based on this, determining the first adjustment information of the k-th target candidate sample to the training sample based on this distance in S103-A includes the following steps S103-A1 and S103-A2:

[0166] S103-A1. The distance between the projection feature information of the training sample and the projection feature information of the kth target candidate sample is subjected to linear activation processing and then random deactivation processing. The result of random deactivation processing is subjected to lossless linear transformation to obtain the second adjustment information of the kth target candidate sample on the training sample.

[0167] S103-A2, Based on the second adjustment information of the kth target candidate sample to the training sample and the label information of the kth target candidate sample, determine the first adjustment information of the kth candidate sample to the training sample.

[0168] The specific process of performing linear activation processing on the distance between the projection feature information of the training sample and the projection feature information of the kth target candidate sample, followed by random deactivation processing, and then performing lossless linear transformation on the random deactivation processing result to obtain the second adjustment information of the kth target candidate sample on the training sample can be referred to the description of the above embodiment, and will not be repeated here.

[0169] Next, the electronic device determines the first adjustment information of the k-th target candidate sample to the training sample based on the second adjustment information of the k-th target candidate sample to the training sample and the label information of the k-th target candidate sample.

[0170] In one possible implementation, the electronic device determines the first adjustment information of the k-th target candidate sample to the training sample by summing the second adjustment information of the k-th target candidate sample to the training sample and the label information of the k-th target candidate sample.

[0171] In one possible implementation, the label information of the kth target candidate sample is projected using a second projection matrix to obtain the label projection information of the kth target candidate sample; the sum of the second adjustment information of the kth target candidate sample on the training sample and the label projection information of the kth target candidate sample is determined as the first adjustment information of the kth target candidate sample on the training sample.

[0172] For example, the electronic device determines the first adjustment information of the k-th target candidate sample to the training sample using the following formula (4):

[0173] in, W represents the first adjustment information of the k-th target candidate sample relative to the training samples. v ∈R d×d Y is the second projection matrix. k For the label information of the k-th target candidate sample, W v (Y k ) represents the label projection information of the k-th target candidate sample.

[0174] The above describes the process by which an electronic device determines the first adjustment information of the k-th target candidate sample in at least one target candidate sample relative to the training sample. Referring to the above steps, the electronic device can determine the first adjustment information of each target candidate sample in at least one target candidate sample relative to the training sample, and then execute step S103-B above, adjusting the object feature information of the training sample based on the first adjustment information of the at least one target candidate sample relative to the training sample, to obtain the adjusted object feature information of the training sample.

[0175] The embodiments of this application do not limit the specific method of adjusting the object feature information of the training sample based on the first adjustment information of at least one target candidate sample to determine the adjusted object feature information of the training sample.

[0176] In some embodiments, the electronic device adds the sum or average of the first adjustment information of the training samples to the object feature information of the training samples for each of the at least one target candidate samples, thereby obtaining the adjusted object feature information of the training samples.

[0177] In some embodiments, S103-B above includes the following steps S103-B1 to S103-B3:

[0178] S103-B1. For the kth target candidate sample in at least one target candidate sample, determine the weight of the kth target candidate sample based on the similarity between the training sample and the kth target candidate sample.

[0179] S103-B2: Based on the weights of at least one target candidate sample, the first adjustment information of the training sample for at least one target candidate sample is weighted to obtain the weighted adjustment information.

[0180] S103-B3. Based on the object feature information of the training samples and the weighted adjustment information, determine the adjusted object feature information of the training samples.

[0181] In this implementation, as shown in Figure 6B, for each target candidate sample among at least one target candidate sample, such as the k-th target candidate sample, the weight of the k-th target candidate sample is determined based on the similarity between the k-th target candidate sample and the training samples. The weight is positively correlated with the similarity; that is, the higher the similarity, the greater the weight, and vice versa.

[0182] This application does not limit the specific method by which the electronic device determines the weight of the k-th target candidate sample based on the similarity sim between the k-th target candidate sample and the training sample. For example, the similarity sim between the k-th target candidate sample and the training sample can be determined as the weight of the k-th target candidate sample.

[0183] In one possible implementation, as shown in Figure 6, the similarity sim between the k-th target candidate sample and the training sample is normalized to obtain the weight of the k-th target candidate sample.

[0184] For example, an electronic device determines the weight of the k-th target candidate sample using the following formula (5): P k =Softmax(Sim(h input ,h k candidate (5)

[0185] Among them, P k Sim(h) represents the weight of the k-th target candidate sample. input ,h k candidate ) represents the similarity between the training sample and the k-th target candidate sample, and Softmax is the normalization function.

[0186] The electronic device can determine the weight of each target candidate sample in at least one target candidate sample based on the above steps, as shown in Figure 6B. Based on the weights of the at least one target candidate sample, the first adjustment information of the training sample by the at least one target candidate sample is weighted to obtain the weighted adjustment information. Then, based on the object feature information of the training sample and the weighted adjustment information, the adjusted object feature information of the training sample is determined. For example, as shown in Figure 6, the sum of the object feature information of the training sample and the weighted adjustment information is determined as the adjusted object feature information of the training sample.

[0187] For example, an electronic device can determine the adjusted object feature information of the training samples using the following formula (6):

[0188] Where h represents the adjusted object feature information of the training samples, h input For the object feature information of the training samples (i.e., the object feature information of the training samples before adjustment), P k This represents the weight (also known as the normalized similarity) of the k-th target candidate sample among at least one target candidate sample. Y represents the candidate feature information of the k-th target candidate sample. k The label information for the k-th target candidate sample. This refers to the adjustment information of the training samples for the k-th target candidate sample. This is the weighted adjustment information.

[0189] S104. The adjusted object feature information of the training samples is classified through the classification module in the data classification model to obtain the classification result of the training samples. The classification result includes the probability reference value of the first object belonging to multiple preset categories.

[0190] In this embodiment of the application, after the electronic device determines the adjusted object feature information of the training sample based on the above steps, as shown in Figure 7, the adjusted object feature information of the training sample is input into the classification module for data classification processing to obtain the classification result of the training sample. The classification result includes the probability reference value of the first object belonging to multiple preset categories.

[0191] Understandably, these multiple preset classifications differ when applied to different data classification scenarios. For example, if L biological state parameters include L gene mutations, the classification result includes a probability reference value for the first object belonging to multiple cancer classifications. The probability reference value for the first object belonging to multiple preset classifications represents the likelihood of the first object belonging to each preset classification. For instance, the sum of the probability reference values ​​for the first object belonging to multiple preset classifications equals 1.

[0192] S105. Based on the classification results of the training samples, determine the loss of the data classification model, and train the data classification model based on the loss.

[0193] As described above, in this embodiment, the label information of the training sample has already been obtained when training the data classification model using the training sample. The electronic device then inputs the training sample into the data classification model for data classification, obtaining the classification result of the training sample output by the data classification model. Based on the deviation between the classification result and the label information of the training sample, the loss of the data classification model is determined. This embodiment does not limit the type of loss function used by the electronic device when determining the loss of the data classification model based on the classification result and label information of the training sample. The classification result includes a probability reference value for the first object belonging to multiple cancer categories, and the label of the training sample is the actual probability that the first object belongs to multiple preset categories. Based on the probability reference value and the actual probability of the first object belonging to multiple cancer categories, the loss of the data classification model is determined. After determining the loss, the parameters in the data classification model are updated based on the loss. Then, a new training sample is selected from the training dataset, and the updated data classification model is iteratively trained. This training is repeated multiple times until the model training termination condition is met. The conditions for ending model training may include reaching a preset number of training iterations and / or the model's loss reaching a preset value.

[0194] The model training method provided in this application training method trains a data classification model, which includes a retrieval module and a classification module. During model training, training samples and Q candidate samples are first acquired. The retrieval module in the data classification model extracts object feature information from the training samples and candidate feature information from the Q candidate samples. The training samples include parameter values ​​of L biological state parameters of a first object, and the candidate samples include parameter values ​​of L biological state parameters of a reference object different from the first object. Next, based on the object feature information of the training samples and the candidate feature information of the Q candidate samples, at least one target candidate sample that satisfies a preset correlation condition with the training samples is selected from the Q candidate samples. Then, based on the candidate feature information of this at least one target candidate sample, the object feature information of the training samples is adjusted to obtain adjusted object feature information of the training samples. Finally, the classification module in the data classification model classifies the adjusted object feature information of the training samples to obtain the classification result of the training samples. Based on the classification result of the training samples, the loss of the data classification model is determined, and the data classification model is trained based on this loss. In this embodiment, at least one target candidate sample selected from Q candidate samples is a sample of other objects similar to the first object. The candidate feature information of these at least one target candidate samples is used to adjust the object feature information of the training samples, which can integrate the features of similar objects, making the adjusted object feature information of the training samples richer and more comprehensive. When using the adjusted object feature information of the training samples for data classification, multiple different classifications can be covered. For example, multiple different types of early cancer classification reference results can be obtained, and the classification accuracy can be improved.

[0195] The overall training process of the data classification model has been introduced above. In some embodiments, to further improve the accuracy of data classification, the retrieval module can be used to retrieve the training samples multiple times, projecting the training samples onto the d-dimensional plane using different linear projections, thereby further extracting the features of the training samples and improving the richness of the extracted features. As shown in Figure 8, the retrieval module of this embodiment includes R feature extraction units. The model training method of this embodiment will be further described below with reference to Figure 9.

[0196] Figure 9 is a flowchart illustrating a model training method provided in an embodiment of this application.

[0197] As shown in Figure 9, the training process of the data detection model in this embodiment includes the following steps:

[0198] S201. Obtain training samples and Q candidate samples.

[0199] The training samples are the parameter values ​​of L biological state parameters of the first object, and the candidate samples are the parameter values ​​of L biological state parameters of a reference object that is different from the first object.

[0200] The specific implementation process of S201 can be referred to the relevant description of S101 above, and will not be repeated here.

[0201] S202. For the j-th feature extraction unit among the R feature extraction units, the j-th feature extraction unit extracts features from the training samples to obtain the j-th sub-object feature information of the training samples, and the j-th feature extraction unit extracts features from the Q candidate samples to obtain the j-th sub-candidate feature information of the Q candidate samples.

[0202] As shown in Figure 8, the retrieval module of this application embodiment includes R feature extraction units. These R feature extraction units operate in parallel. The input training sample and Q candidate samples are processed separately by these R feature extraction units. Finally, the R sub-object feature information of the training sample and the R sub-object feature information of each candidate sample can be obtained. That is, the object feature information of the training sample includes R sub-object feature information, and the candidate feature information of each candidate sample includes R sub-candidate feature information.

[0203] In one possible implementation, the object feature information of the training sample is obtained by concatenating the R sub-object feature information extracted by the R feature extraction units, and the candidate feature information of the candidate sample is obtained by concatenating the R sub-candidate feature information extracted by the R feature extraction units. For example, this concatenation can be performed using the `concat` function.

[0204] In this embodiment, the processing of the training sample and the Q candidate samples corresponding to the training sample by each of the R feature extraction units is the same. For ease of description, the j-th feature extraction unit is used as an example for illustration.

[0205] In this embodiment, the data processing process of each of the R feature extraction units is consistent with the data processing process of the retrieval module in the above embodiment, and the specific process can be referred to the description of the above embodiment.

[0206] As shown in Figure 8, the electronic device inputs the training sample and Q candidate samples into R feature extraction units. For each feature extraction unit, such as the j-th feature extraction unit, the j-th feature extraction unit extracts the object feature information of the training sample and the candidate feature information of each of the Q candidate samples. For ease of description, the object feature information of the training sample extracted by the j-th feature extraction unit is denoted as the j-th sub-object feature information of the training sample. Similarly, the candidate feature information of the candidate sample extracted by the j-th feature extraction unit is denoted as the j-th sub-candidate feature information of the candidate sample.

[0207] In this embodiment of the application, the specific methods by which the j-th feature extraction unit extracts the j-th sub-object feature information of the training samples and the j-th sub-candidate feature information of the Q candidate samples include at least the following methods:

[0208] Method 1: When the j-th feature extraction unit extracts the feature information of the j-th sub-object of the training sample, it does not consider the Q candidate samples. At the same time, when extracting the feature information of the j-th sub-candidate of each candidate sample in the Q candidate samples, it does not consider the training sample and other candidate samples in the Q candidate samples (i.e., the remaining Q-1 candidate samples).

[0209] Method 2: When the j-th feature extraction unit extracts the feature information of the j-th sub-object of the training sample, it considers Q candidate samples. At the same time, when extracting the feature information of the j-th sub-candidate of each of the Q candidate samples, it does not consider the training sample and other candidate samples in the Q candidate samples (i.e., the remaining Q-1 candidate samples).

[0210] The embodiments of this application do not impose restrictions on the specific network structure of the j-th feature extraction unit.

[0211] In some embodiments, as shown in FIG10, each feature extraction unit includes an encoder. The encoder included in the j-th feature extraction unit is denoted as the j-th encoder. Correspondingly, training samples (S) are extracted through the j-th feature extraction unit. input The feature information of the j-th sub-object and the Q candidate samples (S) candidate The j-th sub-candidate feature information includes: extracting features from the training samples using the j-th encoder to obtain the j-th sub-object feature information of the training samples; and extracting features from the Q candidate samples using the j-th encoder to obtain the j-th sub-candidate feature information of the Q candidate samples.

[0212] As shown in Figure 10, each of the R feature extraction units in this embodiment includes an encoder, and the encoders included in different feature extraction units can be different. Thus, through the different encoders included in the R feature extraction units, the input training samples and candidate samples are linearly projected into a d-dimensional feature space. During the retrieval process, different encoders are used to encode the training samples and candidate samples, and then candidate samples related to the training samples are identified from different perspectives. The features of the related candidate samples are used to adjust the features of the training samples, thereby enriching the features of the training samples.

[0213] The embodiments of this application do not impose restrictions on the grid structure of the encoder.

[0214] In one example, the encoder included in the j-th feature extraction unit mentioned above may contain a linear layer that projects training samples and candidate samples onto the hidden dimension.

[0215] S203. Based on the R sub-object feature information of the training samples and the R sub-candidate feature information of the Q candidate samples, select Kj target candidate samples from the Q candidate samples that meet the preset correlation conditions with the training samples.

[0216] In this embodiment, for each of the R feature extraction units, such as the j-th feature extraction unit, the electronic device extracts the j-th sub-object feature information of the training sample and the j-th sub-candidate feature information of the candidate sample through the j-th feature extraction unit. Then, based on the R sub-object feature information of the training sample and the R sub-candidate feature information of the Q candidate samples, Kj target candidate samples related to the training sample are selected from the Q candidate samples. The R sub-object feature information of the training sample constitutes the object feature information of the training sample, and the R sub-candidate feature information of the Q candidate samples constitutes the candidate feature information of the candidate sample.

[0217] It should be noted that for different feature extraction units, the number of corresponding Kj target candidate samples may be different, and the selected target candidate samples may also be different.

[0218] This application does not limit the specific method by which an electronic device selects Kj target candidate samples related to the training samples from the Q candidate samples based on the R sub-object feature information of the training samples and the R sub-candidate feature information of the Q candidate samples.

[0219] In some embodiments, the electronic device clusters the R sub-object feature information of the training samples and the R sub-candidate feature information of the Q candidate samples, and determines the candidate samples that belong to the same cluster as the training samples as the Kj target candidate samples related to the training samples.

[0220] In some embodiments, S203 above includes the following steps S203-A and S203-B:

[0221] S203-A: Based on the feature information of R sub-objects in the training samples and the feature information of R sub-candidates in Q candidate samples, determine the similarity between the training samples and each candidate sample in the Q candidate samples.

[0222] S203-B: Based on the similarity between the training samples and each candidate sample, select Kj target candidate samples from Q candidate samples.

[0223] This application does not limit the specific method by which an electronic device determines the similarity between a candidate sample and a training sample based on the R sub-object feature information of the candidate sample and the R sub-candidate feature information of the training sample.

[0224] In one possible implementation, the electronic device uses cosine distance or L2 distance to calculate the distance between R sub-candidate features of the candidate sample and R sub-object features of the training sample, and then determines the similarity between the candidate sample and the training sample based on this distance. In one example, the reciprocal of the distance between the R sub-candidate features of the candidate sample and the R sub-object features of the training sample can be used to determine the similarity between the candidate sample and the training sample.

[0225] In some embodiments, S203-A above includes the following steps S203-A1 to S203-A3:

[0226] S203-A1. For the qth candidate sample among Q candidate samples, the jth sub-candidate feature information of the qth candidate sample is projected using the first projection matrix to obtain the jth projected feature information of the qth candidate sample, where the value of q is 1, 2, ..., Q.

[0227] S203-A2: Project the j-th sub-object feature information of the training sample using the first projection matrix to obtain the j-th projected feature information of the training sample;

[0228] S203-A3: Determine the distance between the R projected features of the training sample and the R projected features of the q-th candidate sample, and determine the similarity between the training sample and the q-th candidate sample based on the distance.

[0229] Specifically, for the q-th candidate sample out of Q candidate samples, the first projection matrix W is used. kProjecting the j-th sub-candidate feature information of the q-th candidate sample, for ease of description, we denote this projected j-th sub-candidate feature information as the j-th projected feature information of the q-th candidate sample. Similarly, we use the first projection matrix W. k Projecting the feature information of the j-th sub-object of the training sample yields the j-th projected feature information of the training sample. Next, the distance between the R projected feature information of the training sample and the R projected feature information of the q-th candidate sample is calculated, and this distance is determined as the similarity between the training sample and the q-th candidate sample.

[0230] In one example, the electronic device uses the L2 distance as the distance between the R projected features of the training sample and the R projected features of the q-th candidate sample. Specifically, the method shown in formula (2) above can be referred to.

[0231] Electronic devices can use the above method to determine the similarity between each of the Q candidate samples and the training samples.

[0232] Next, based on the similarity between each candidate sample and the training sample in the Q candidate samples, the electronic device selects Kj target candidate samples from the Q training samples that meet the preset correlation conditions with the training samples.

[0233] In one example, the electronic device selects the Kj target candidate samples from Q candidate samples that have the highest similarity to the training samples, as the Kj candidate samples associated with the training samples.

[0234] S204. Based on the j-th sub-candidate feature information of Kj target candidate samples, adjust the j-th sub-object feature information of the training samples to obtain the j-th adjusted sub-object feature information of the training samples.

[0235] In this embodiment of the application, for the j-th feature extraction unit among the R feature extraction units, the electronic device can select Kj target candidate samples related to the training sample from the Q candidate samples based on the above steps. The electronic device adjusts the object feature information of the training sample based on the candidate feature information of the Kj target candidate samples to obtain the adjusted object feature information of the training sample. Here, the specific method of adjustment is not limited.

[0236] In some embodiments, the electronic device can fuse the j-th sub-object feature information of the training sample with the j-th sub-candidate feature information of the Kj target candidate samples to obtain the j-th adjusted sub-object feature information of the training sample. For example, based on the similarity between the Kj target candidate samples and the training sample, the weight of each target candidate sample in the Kj target candidate samples is determined, and then based on the weight, the j-th sub-object feature information of the training sample is weighted with the j-th sub-candidate feature information of the Kj target candidate samples to obtain the j-th adjusted sub-object feature information of the training sample.

[0237] In some embodiments, S204 above includes the following steps S204-A and S204-B:

[0238] S204-A. For the kth target candidate sample among Kj target candidate samples, determine the distance between the jth projection feature information of the training sample and the jth projection feature information of the kth target candidate sample, and based on the distance, determine the first adjustment information of the kth target candidate sample to the training sample, where the value of k is 1, 2, ..., Kj;

[0239] S204-B: Based on the first adjustment information of the training sample from the Kj target candidate samples, adjust the feature information of the j-th sub-object of the training sample, and determine the adjusted feature information of the j-th sub-object of the training sample.

[0240] In this implementation, the distance between the training samples and Kj target candidate samples is used in the hidden layer to determine the feature adjustment information of each target candidate sample relative to the training samples. Specifically, for the k-th target candidate sample among the Kj target candidate samples, the distance between the j-th projected feature information of the training sample and the j-th projected feature information of the k-th target candidate sample is determined, and based on this distance, the first adjustment information of the k-th target candidate sample relative to the training samples is determined.

[0241] This application does not limit the specific method by which the first adjustment information of the k-th candidate sample to the training sample is determined based on the distance.

[0242] In some embodiments, the electronic device can determine the distance between the j-th projection feature information of the training sample and the j-th projection feature information of the k-th target candidate sample as the first adjustment information of the k-th candidate sample to the training sample.

[0243] In some embodiments, the electronic device performs activation processing on the distance between the j-th projected feature information of the training sample and the j-th projected feature information of the k-th candidate sample, followed by random deactivation processing. A lossless linear transformation is then performed on the random deactivation result to obtain the second adjustment information of the k-th target candidate sample relative to the training sample. This second adjustment information of the k-th candidate sample relative to the training sample is then determined as the first adjustment information of the k-th candidate sample relative to the training sample.

[0244] For example, the electronic device uses the above formula (3) to determine the second adjustment information, and then determines the second adjustment information as the first adjustment information of the kth candidate sample to the training sample.

[0245] In some embodiments, since the Kj target candidate samples are similar to the training samples, the label information of the Kj target candidate samples can also provide additional information for the features of the training samples. Based on this, the determination of the first adjustment information of the k-th target candidate sample to the training samples based on distance in S204-A above includes the following steps S204-A1 and S204-A2:

[0246] S204-A1: After linear activation processing of the distance, random deactivation processing is performed, and lossless linear transformation is performed on the result of random deactivation processing to obtain the second adjustment information of the k-th target candidate sample on the training sample.

[0247] S204-A2, Based on the second adjustment information of the k-th target candidate sample to the training sample and the label information of the k-th target candidate sample, determine the first adjustment information of the k-th candidate sample to the training sample.

[0248] The specific process of performing linear activation processing on the distance between the j-th projection feature information of the training sample and the j-th projection feature information of the k-th candidate sample, followed by random deactivation processing, and then performing lossless linear transformation on the random deactivation processing result to obtain the second adjustment information of the k-th target candidate sample on the training sample can be referred to the description of the above embodiment, and will not be repeated here.

[0249] Next, the electronic device determines the first adjustment information of the k-th candidate sample to the training sample based on the second adjustment information of the k-th target candidate sample to the training sample and the label information of the k-th target candidate sample.

[0250] In one possible implementation, the electronic device determines the first adjustment information of the k-th candidate sample to the training sample by summing the second adjustment information of the k-th candidate sample to the training sample and the label information of the k-th candidate sample.

[0251] In one possible implementation, the label information of the k-th candidate sample is projected using a second projection matrix to obtain the label projection information of the k-th candidate sample; the sum of the second adjustment information of the k-th candidate sample relative to the training sample and the label projection information of the k-th candidate sample is determined as the first adjustment information of the k-th candidate sample relative to the training sample.

[0252] The above describes the process by which an electronic device determines the first adjustment information of the k-th target candidate sample among Kj target candidate samples in relation to the training sample. Referring to the above steps, the electronic device can determine the first adjustment information of each of the Kj target candidate samples in relation to the training sample, and then execute step S204-B as described above. Based on the first adjustment information of the Kj target candidate samples in relation to the training sample, the j-th sub-object feature information of the training sample is adjusted to determine the j-th adjusted sub-object feature information of the training sample. The R adjusted sub-object feature information of the training sample constitutes the adjusted object feature of the training sample.

[0253] In this embodiment, the first adjustment information of the training sample based on Kj target candidate samples is used to adjust the feature information of the j-th sub-object of the training sample. The specific method of determining the j-th adjusted sub-object feature information of the training sample is not limited.

[0254] In some embodiments, the electronic device adds the sum or average of the first adjustment information of the Kj target candidate samples to the training sample, and adds the j-th sub-object feature information of the training sample to obtain the j-th adjusted sub-object feature information of the training sample.

[0255] In some embodiments, S204-B above includes the following steps S204-B1 to S204-B3:

[0256] S204-B1. For the kth target candidate sample among Kj target candidate samples, determine the weight of the kth target candidate sample based on the similarity between the training sample and the kth target candidate sample.

[0257] S204-B2: Based on the weights of the Kj target candidate samples, the first adjustment information of the Kj target candidate samples on the training samples is weighted to obtain the weighted adjustment information;

[0258] S204-B3. Based on the feature information of the j-th sub-object in the training sample and the weighted adjustment information, determine the adjusted feature information of the j-th sub-object in the training sample.

[0259] In this implementation, for each of the Kj target candidate samples, such as the k-th target candidate sample, the weight of the k-th target candidate sample is determined based on the similarity between the k-th target candidate sample and the training samples. The weight is positively correlated with the similarity.

[0260] This application does not limit the specific method by which the electronic device determines the weight of the k-th target candidate sample based on the similarity sim between the k-th target candidate sample and the training sample. For example, the similarity sim between the k-th target candidate sample and the training sample can be determined as the weight of the k-th target candidate sample.

[0261] In one possible implementation, the similarity sim between the k-th target candidate sample and the training sample is normalized to obtain the weight of the k-th target candidate sample. For example, the weight of each of the Kj target candidate samples can be determined by the method shown in formula (5) above.

[0262] The electronic device can determine the weight of each of the Kj target candidate samples based on the above steps. Based on these weights, it can weight the first adjustment information of the training sample for each of the Kj target candidate samples to obtain weighted adjustment information. Then, based on the j-th sub-object feature information of the training sample and the weighted adjustment information, it can determine the j-th adjusted sub-object feature information of the training sample. For example, the sum of the j-th sub-object feature information and the weighted adjustment information can be used to determine the j-th adjusted sub-object feature information of the training sample.

[0263] The above describes the process of determining the j-th adjusted sub-object feature information of the training sample through the j-th feature extraction unit, using the j-th feature extraction unit as an example. Referring to the above method, the electronic device can determine the adjustment information of the training sample by each feature extraction unit in the R feature extraction units, thereby obtaining the R adjusted sub-object feature information of the training sample, denoted as h1, h2…h j ...h r , where h j This refers to the adjusted sub-object feature information of the training sample corresponding to the j-th feature extraction unit.

[0264] In some embodiments, the electronic device can compare and analyze the R sub-object feature information of the training sample with the R sub-candidate feature information of Kj target candidate samples. For example, it can calculate the variance of the j-th sub-object feature information with that of the j-th sub-candidate feature information of the Kj target candidate samples. If the variance corresponding to the j-th sub-object feature information is higher than a threshold, the j-th sub-object feature information is deleted to obtain the adjusted object feature information of the training sample. Alternatively, if the variance corresponding to the j-th sub-object feature information is higher than a threshold, it can determine to modify the j-th sub-object feature information. For example, it can calculate the mean or weighted mean of the j-th sub-candidate feature information of the Kj target candidate samples and modify the j-th sub-object feature information to this mean or weighted mean.

[0265] S205. The classification module classifies the R adjusted sub-object feature information of the training samples to obtain the classification result of the training samples.

[0266] As shown in Figure 11, in this embodiment of the application, the electronic device uses a classification module to classify the R adjusted sub-object feature information of the training samples to obtain the classification result of the training samples.

[0267] In some embodiments, the electronic device fuses the R adjusted sub-object feature information of the training sample (e.g., by addition, concatenation, multiplication, etc.), and then classifies the fused feature information through a classification module to obtain the classification result of the training sample.

[0268] In some embodiments, as shown in FIG12, the classification module includes a hybrid expert module, which includes a gating unit and E expert units. The gating unit is used to determine the expert unit that assigns each input token to the top-n probabilities. In this case, S205 above includes the following steps S205-A to S205-C:

[0269] S205-A: Based on the R adjusted sub-object feature information from the training samples, S input tokens are obtained, where S is a positive integer greater than 1.

[0270] In this embodiment, for the j-th feature extraction unit among R feature extraction units, the j-th adjusted sub-object feature information of the training sample corresponding to the j-th classification unit includes the adjusted feature information of L biological state parameters. The j-th adjusted sub-object feature information includes d elements, i.e., a vector of length d. In this way, the R adjusted sub-object feature information of the training sample can form a matrix.

[0271] In some embodiments, each of the R adjusted sub-object feature information from the training samples is taken as a row of a first matrix to obtain a first matrix with R rows; each column of the first matrix is ​​taken as an input token to obtain S input tokens. If the dimension of each adjusted sub-object feature information is d, then S is equal to d.

[0272] For example, the first matrix H can be represented as: H = Concat(h1, h2, ..., h r ) T , where H∈R d×r Cancat represents the concatenation operation. Each element in the d-dimensional array can be considered as an input token (also called a mark) in a subsequent Hybrid Expert (MoE) module. For example, in this embodiment, the letter A is used to represent the input token. The r-dimensional array is embedded as its hidden layer.

[0273] S205-B: For the s-th input token among S input tokens, the probability value of the s-th input token being assigned to each of the E expert units is determined through the gating unit.

[0274] Where E is a positive integer greater than 1, and s takes the values ​​1, 2, ..., S.

[0275] For example, the electronic device can determine the probability value of each input token A being assigned to each expert unit using the following formula (7): G = H·W Gate (7)

[0276] Among them, W Gate Let G be the learnable parameter matrix, where G represents the probability of the input token being assigned to each expert unit. H is the first matrix.

[0277] For example, as shown in Figure 13, which assumes that the number of input tokens S equals 6 and the number of expert units E equals 4, the electronic device determines the probability value of assigning the 6 input tokens A1, A2, A3, A4, A5, and A6 to each of the 4 expert units through a gating unit. For example, the probability value of input token A1 being assigned to the 4 expert units is as follows: in The probability value for assigning input token A1 to the first expert unit e1 for processing, and so on, is given. The probability value for assigning input token A1 to the 4th expert unit e4 for processing.

[0278] S205-C: Based on probability values, select T expert units from E expert units to process the s-th input token, and process the s-th input token through the T expert units respectively to obtain the fusion feature information of the s-th input token.

[0279] Where T is a positive integer less than E.

[0280] Based on the above steps, for the s-th input token out of S input tokens, the electronic device can determine the probability value of the s-th input token being assigned to each of the E expert units. Then, according to the probability values, T expert units are selected from the E expert units to process the s-th input token. That is, the T expert units with the highest probability values ​​are selected from the E expert units.

[0281] For example, the electronic device can determine the T expert units for processing the s-th input token using the following formula (8):

[0282] Among them, G s Let be the probability value that the s-th input token is assigned to E expert units, and topN(·) represent the probability value that each input token (d) in the expert dimension is assigned to W. Gate The position of the Nth element with the highest probability. These T top-level expert units will receive the s-th input token, topT_indice. s s represents the T expert units used to process the s-th input token.

[0283] Referring to Figure 13, taking the s-th input token as the first input token A1 as an example, as shown in Figure 13, the probability value of the first input token A1 being assigned to the first expert unit e1 is among the probability values ​​of the four expert units. and the probability value assigned to the second expert unit e2 Therefore, it can be determined that the first input token A1 will be assigned to the first expert unit e1 and the second expert unit e2 for processing, at which point T equals 2.

[0284] In practice, some expert units may receive a large number of input tokens, while others receive only a small number. This phenomenon is called unbalanced expert load. This situation not only contradicts the initial design intent of the data classification model in this application, but also affects computational efficiency (e.g., causing uneven communication load in distributed training). Therefore, this application sets a fixed capacity value C for each expert unit (i.e., the maximum number of tokens each expert unit can accept). If the number of input tokens currently received by an expert unit exceeds its capacity, that expert unit will no longer receive input tokens. This excess input token is called overflow.

[0285] Based on the above steps, the electronic device selects T expert units from E expert units to process the s-th input token, and then assigns the s-th input token to the T expert units for processing to obtain the fusion feature information of the s-th input token.

[0286] This application embodiment does not limit the specific method by which the electronic device processes the s-th input token through T expert units to obtain the fused feature information of the s-th input token.

[0287] In some embodiments, the electronic device inputs the s-th input token into T expert units respectively, and obtains feature information by processing the s-th input token in each of the T expert units, thereby obtaining T feature information. Then, these T feature information are fused, for example, by adding the T feature information to obtain the fused feature information of the s-th input token, or by determining the average value of the T feature information as the fused feature information of the s-th input token.

[0288] In some embodiments, the steps in S205-C1 to S205-C3 above, in which T expert units process the s-th input token respectively to obtain the fused feature information of the s-th input token, include the following steps:

[0289] S205-C1. For each of the T expert units, feature fusion processing is performed on each input token assigned to the expert unit to obtain the output feature information of each input token, including the s-th input token.

[0290] S205-C2, Based on the probability value of the s-th input token being assigned to each of the T expert units, determine the weights corresponding to the T expert units respectively;

[0291] S205-C3: Based on the weights corresponding to the T expert units, the output feature information of the s-th input token output by the T expert units is weighted to obtain the fused feature information of the s-th input token.

[0292] In this implementation, for each of the T expert units, the electronic device performs feature fusion processing on each input token assigned to the expert unit to obtain the output feature information of each input token, wherein each input token includes the s-th input token.

[0293] The embodiments of this application do not limit the specific network structure of the expert unit. In one example, the expert unit includes two linear transformations and an activation function, such as the ReLU activation function, or other types of activation functions.

[0294] In some embodiments, the E expert units included in the data classification model of this application have a unified structure, rather than different structures specifically designed for certain classifications (e.g., cancer types), so that the data classification model proposed in this application can be generalized to other datasets.

[0295] In this embodiment of the application, the electronic device further determines the weights corresponding to the T expert units based on the probability value of the s-th input token being assigned to each of the T expert units.

[0296] For example, the electronic device determines the weights corresponding to the expert units using the following formula (9): W = softmax(TopT(G)) (9)

[0297] In this example, the electronic device uses the softmax function to normalize the probability value of the s-th input token being assigned to the T expert units, thus deriving the weight of each input token being assigned to the T expert units.

[0298] Next, based on the weights corresponding to the T expert units, the output feature information of the s-th input token output by the T expert units is weighted to obtain the fused feature information of the s-th input token.

[0299] The above describes the process of determining the fusion feature information of the s-th input token out of S input tokens, using it as an example. Following this method, an electronic device can determine the fusion feature information of each of the S input tokens.

[0300] For example, the electronic device determines the fusion feature information of the s-th input token using the following formula (10).

[0301] in, Its shape is the same as the first matrix H. The output feature information of the s-th input token output by the expert unit.

[0302] Referring to Figure 13, taking the s-th input token as the first input token A1 as an example, this first input token is assigned to the first expert unit e1 and the second expert unit e2, where the weight of the first expert unit e1 in processing the first input token A1 is... The weight of the second expert unit e2 in processing the first input token A1 is: Thus, the weight matrix corresponding to the first input token A1 The first expert unit processes the first input token A1, obtaining output feature information A11. The second expert unit processes the first input token A1, obtaining output feature information A12. At this point, the output feature information corresponding to the first input token A1... Thus, the fused feature information of the first input token A1

[0303] Referring to the method described above, the electronic device can determine the fusion feature information of each of the S input tokens. Next, steps S205-D are performed.

[0304] S205-D: Based on the fusion feature information of S input tokens, the mixed feature information of the training samples is obtained.

[0305] This application does not limit the specific method by which an electronic device obtains the hybrid feature information of training samples based on the fused feature information of S input tokens.

[0306] In some embodiments, the average or sum of the fused feature information of these S input tokens is determined as the fused feature information of the training samples.

[0307] In some embodiments, the fused feature information of these S input tokens is concatenated to obtain the mixed feature information of the training samples.

[0308] S205-E: Based on mixed feature information, the training samples are classified to obtain the classification results of the training samples.

[0309] This application embodiment does not limit the specific method by which electronic devices classify training samples based on mixed feature information to obtain the classification results of the training samples.

[0310] In some embodiments, the mixed feature information of the training samples is classified to obtain the classification result of the training samples.

[0311] In some embodiments, a linear layer is used to linearly transform the mixed feature information of the training samples to reduce dimensionality and aggregate features. Then, a simple linear classification head is used to map the hidden layers to the final output dimension. The probabilities are then normalized to generate the final output, predicting the probability value of the training sample belonging to a preset category, thus obtaining the classification result of the training samples.

[0312] For example, the electronic device determines the classification result of the training samples using the following formula (11): p y =softmax(FC head (Linear(h MOE (11)

[0313] Where, p y The classification result of the training samples, i.e., the probability value of the training samples belonging to the preset category. Linear is a linear layer, FC. head For linear classification, softmax is the normalization function.

[0314] S206. Based on the classification results of the training samples, determine the loss of the data classification model, and train the data classification model based on the loss.

[0315] In some embodiments, the electronic device determines a first loss based on the classification results and label information of the training samples, and then determines the first loss as the loss of the data classification model.

[0316] For example, the electronic device determines the first loss L using the following formula (12). main (p y ,y):

[0317] Where, n class y is the preset number of categories. i p represents the true probability that a training sample belongs to the i-th category in the preset classification. yi This is the probability value predicted by the data classification model for a training sample to belong to the i-th category.

[0318] It should be noted that electronic devices, in addition to using the classification cross-entropy loss function L... main Besides using other loss functions to calculate the first loss, this application does not limit the specific methods used in this embodiment.

[0319] In some embodiments, S206 above includes the following steps S206-A to S206-C:

[0320] S206-A, Determine the first loss based on the classification results and label information of the training samples;

[0321] S206-B: Determine the second loss based on the number of input labels assigned to each of the E expert units and the assignment probability value;

[0322] S206-C, Based on the first loss and the second loss, determine the loss of the data classification model.

[0323] In this implementation, to achieve a uniform distribution among expert units, prevent uneven distribution of input tokens, and avoid discarding too many input tokens due to overlap, this embodiment introduces a balanced loss function L. balance This balancing loss function aims to minimize the proportion c of input tokens received by each expert unit.e The squared difference of / S. However, due to c e It is obtained through a top-T process and is not differentiable. To address this issue, embodiments of this application use the average probability m of all tokens received by each expert unit. e As a differentiable alternative, replace the square with m. e (c e In this way, the embodiments of this application can optimize this loss function through gradient descent, effectively preserving the original concept, solving the problem of non-differentiability, and ensuring the balanced distribution of input tokens assigned by the expert unit.

[0324] For example, the electronic device can determine the second loss L using the following formula (13). balance :

[0325] Where E is the number of expert units, c e Let m be the number of input tokens assigned to the e-th expert unit out of E expert units, where S is the total number of input tokens. e Let be the average probability of all tokens assigned to the e-th expert unit.

[0326] After determining the first loss and the second loss based on the above steps, the electronic device determines the loss of the data classification model based on the first loss and the second loss.

[0327] In one example, the electronic device determines the loss of the data classification model as the sum of the first loss and the second loss.

[0328] In one example, the electronic device multiplies the second loss by the balance coefficient and then adds it to the first loss to obtain the loss of the data classification model.

[0329] For example, the electronic device determines the loss L of the data classification model according to the following formula (14): L = L main +α*L balance (14)

[0330] Where α is the balance coefficient.

[0331] Next, the electronic device adjusts the parameters in the data classification model based on the aforementioned loss. This process is repeated multiple times to obtain the trained data classification model.

[0332] The model training method provided in this application includes a data classification model comprising a retrieval module and a classification module, and the data classification model includes R feature extraction units. These R feature extraction units extract R sub-object feature information of the training samples, and R sub-candidate feature information of each of the Q candidate samples. For each feature extraction unit, such as the j-th feature extraction unit, based on the j-th sub-object feature information of the training samples and the j-th sub-candidate feature information of the Q training samples, Kj target candidate samples that satisfy a preset correlation condition with the training samples are selected from the Q candidate samples. Then, based on the j-th sub-candidate feature information of these Kj target candidate samples, the j-th sub-object feature information of the training samples is adjusted to obtain the j-th adjusted sub-object feature information of the training samples. This process is repeated to obtain R adjusted sub-object feature information of the training samples. Then, the classification module in the data classification model classifies the R adjusted sub-object feature information of the training sample to obtain the classification result of the training sample. For example, the hybrid expert module mixes the R adjusted sub-object feature information of the training sample before data classification to obtain the classification result of the training sample. Based on the classification result of the training sample, the loss of the data classification model is determined, and the model is trained based on this loss. Thus, by extracting features of similar objects multiple times from different linear transformation spaces, the R adjusted sub-object feature information of the training sample includes richer features. Therefore, when using the R adjusted sub-object feature information of the training sample for data classification, it can cover multiple different classifications, such as obtaining classification reference results for multiple different types of early cancer, and improving classification accuracy.

[0333] The training process of the data classification model has been introduced above. The following section introduces the data classification process based on the trained data classification model.

[0334] Figure 14 is a schematic flowchart of a data classification method provided in an embodiment of this application. The execution entity of this embodiment can be a device with data classification functionality, such as a data classification apparatus. This data classification apparatus can be an electronic device such as the server 102 or terminal device 101 shown in Figure 1, or it can be a system composed of the server 102 or 101 shown in Figure 1. For ease of description, the following embodiments use an electronic device as an example to illustrate the method of this embodiment.

[0335] As shown in Figure 14, the data classification method of this application embodiment includes:

[0336] S301. Obtain object data and Q candidate data, and extract object feature information from the object data and candidate feature information from the Q candidate data respectively.

[0337] The object data includes the parameter values ​​of L biological state parameters of the classified object, and the candidate data consists of the parameter values ​​of L biological state parameters of other reference objects that are different from the classified object. Q and L are both positive integers greater than 1.

[0338] In this embodiment, a candidate dataset is involved, which includes multiple different candidate data. When object data needs to be classified, the electronic device can obtain Q candidate data corresponding to the object data from the candidate dataset. For example, Q candidate data can be randomly selected from the candidate dataset.

[0339] This application does not limit the specific method by which the electronic device extracts object feature information from object data and candidate feature information from Q candidate data. For example, it can extract object feature information from object data and candidate feature information from Q candidate data using a trained feature extraction model.

[0340] In some embodiments, the electronic device extracts object feature information of the object data and subsequent feature information of Q candidate data through the retrieval module in the data classification model described above.

[0341] It should be noted that the specific implementation process of S301 is basically the same as that of S101. Referring to the relevant description of S101, it will not be repeated here.

[0342] S302. Based on the object feature information of the object data and the candidate feature information of Q candidate data, select at least one target candidate data from the Q candidate data that satisfies the preset correlation condition with the object data.

[0343] The specific implementation process of S302 is basically the same as that of S102. Referring to the relevant description of S102, it will not be repeated here.

[0344] S303. Based on the candidate feature information of at least one target candidate data, adjust the object feature information of the object data to obtain the adjusted object feature information of the object data.

[0345] In this embodiment of the application, adjusting the object feature information of the object data can be understood as expanding the object feature information of the object data to enrich the object feature information of the object data, so that the adjusted object feature information includes not only the feature information of the detected object, but also the feature information of similar reference objects.

[0346] The specific implementation process of S303 is basically the same as that of S103. Referring to the relevant description of S103, it will not be repeated here.

[0347] S304. The adjusted object feature information of the object data is classified to obtain the classification result of the object data. The classification result includes the probability reference value of the classified object belonging to multiple preset categories.

[0348] This application embodiment classifies the adjusted object feature information of object data by an electronic device, and the specific method for obtaining the classification result of the object data is not limited. For example, the adjusted object feature information of the object data can be classified using a trained data classification model to obtain the classification result of the object data.

[0349] In some embodiments, the electronic device uses the classification module in the data classification model described above to classify the adjusted object feature information of the object data to obtain the classification result of the object data.

[0350] It should be noted that if the data classification method of this application embodiment is applied to disease diagnosis, the classification result obtained by this application embodiment is a disease classification reference result, not a direct result of disease diagnosis. For example, when the data classification method of this application embodiment is applied to the detection of early cancer, this application embodiment can obtain a classification reference result indicating whether the classified object belongs to multiple types of early cancer.

[0351] The specific implementation process of S304 is basically the same as that of S104. Referring to the relevant description of S104, it will not be repeated here.

[0352] The data classification method provided in this application first acquires object data and Q candidate data during data classification. Object feature information of the object data and candidate feature information of the Q candidate data are extracted respectively. The object data includes parameter values ​​of L biological state parameters of the object to be classified, and the candidate data includes parameter values ​​of L biological state parameters of reference objects different from the object to be classified. Next, based on the object feature information of the object data and the candidate feature information of the Q candidate data, at least one target candidate data that satisfies a preset correlation condition with the object data is selected from the Q candidate data. Then, based on the candidate feature information of this at least one target candidate data, the object feature information is adjusted to obtain adjusted object feature information. Finally, the adjusted object feature information of the object data is classified to obtain the classification result of the object data. The classification result includes probability reference values ​​of the classified object belonging to multiple preset categories. In this embodiment, at least one target candidate data selected from Q candidate data is data of other objects similar to the classification object. The candidate feature information of the at least one target candidate data is used to adjust the object feature information of the object data, which can integrate the features of similar objects, making the adjusted object feature information of the object data richer and more comprehensive. Therefore, when using the adjusted object feature information of the object data for data classification, it can cover multiple different classifications. For example, it can obtain classification reference results for multiple different types of early cancer, and improve the classification accuracy.

[0353] The data classification process involved in the embodiments of this application has been described above. In some embodiments, as shown in FIG8, the retrieval module of this application embodiment includes R feature extraction units.

[0354] Figure 15 is a flowchart illustrating a data classification method provided in an embodiment of this application. As shown in Figure 15, the data classification process in this embodiment includes the following steps:

[0355] S401, Obtain object data and Q candidate data.

[0356] The specific implementation process of S401 is the same as that of S301. Referring to the relevant description of S301, it will not be repeated here.

[0357] S402. For the j-th feature extraction unit among the R feature extraction units, the j-th sub-object feature information of the object data and the j-th candidate sub-feature information of the Q candidate data are extracted through the j-th feature extraction unit.

[0358] Where j takes values ​​of 1, 2, ..., R. The object feature information of the object data includes R sub-object feature information, and the candidate feature information of each candidate data includes R sub-candidate feature information.

[0359] For example, the feature extraction unit can be an encoder, which extracts the j-th sub-object feature information of the object data and the j-th sub-candidate feature information of the Q candidate data through the j-th encoder.

[0360] The specific implementation process of S402 can be referred to the relevant description of S202 above, and will not be repeated here.

[0361] S403. Based on the R sub-object feature information of the object data and the R sub-candidate feature information of the Q candidate data, select Kj target candidate data from the Q candidate data that meet the preset correlation conditions with the object data.

[0362] For example, based on the R sub-object feature information of the object data and the R sub-candidate feature information of the Q candidate data, the similarity between the object data and each candidate data in the Q candidate data is determined; based on the similarity, Kj target candidate data are selected from the Q candidate data.

[0363] In some embodiments, determining the similarity between the object data and each of the Q candidate data based on the R sub-object feature information of the object data and the R sub-candidate feature information of the Q candidate data includes: for the q-th candidate data in the Q candidate data, projecting the j-th sub-candidate feature information of the q-th candidate data using a first projection matrix to obtain the j-th projected feature information of the q-th candidate data, where q takes values ​​of 1, 2, ..., Q; projecting the j-th sub-logarithmic feature information of the object data using the first projection matrix to obtain the j-th projected feature information of the object data; determining the distance between the R projected feature information of the object data and the R projected feature information of the q-th candidate data, and determining the similarity between the object data and the q-th candidate data based on the distance.

[0364] The specific implementation process of S403 can be referred to the relevant description of S203 above, and will not be repeated here.

[0365] S404. Based on the j-th sub-candidate feature information of Kj target candidate data, adjust the j-th sub-object feature information of the object data to obtain the j-th adjusted sub-object feature information of the object data.

[0366] For example, for the kth target candidate data among Kj target candidate data, determine the distance between the jth projection feature information of the object data and the jth projection feature information of the kth target candidate data, and based on the distance, determine the first adjustment information of the object data by the kth target candidate data, where k takes the value 1, 2, ..., Kj; based on the first adjustment information of the object data by the Kj target candidate data respectively, adjust the jth sub-object feature information of the object data, and determine the jth adjusted sub-object feature information of the object data.

[0367] In some embodiments, determining the first adjustment information of the k-th target candidate data on the object data based on distance includes: performing linear activation processing on the distance and then random deactivation processing, and performing lossless linear transformation on the random deactivation processing result to obtain the second adjustment information of the k-th target candidate data on the object data; and determining the first adjustment information of the k-th target candidate data on the object data based on the second adjustment information of the k-th target candidate data on the object data and the label information of the k-th target candidate data.

[0368] In some embodiments, determining the first adjustment information of the object data based on the second adjustment information of the object data based on the kth target candidate data and the label information of the kth target candidate data includes: projecting the label information of the kth target candidate data using a second projection matrix to obtain the label projection information of the kth target candidate data; and determining the sum of the second adjustment information of the object data based on the kth target candidate data and the label projection information of the kth target candidate data as the first adjustment information of the object data based on the kth target candidate data.

[0369] In some embodiments, the above-mentioned adjustment of the first adjustment information of the object data based on Kj target candidate data to the j-th sub-object feature information of the object data, and the determination of the j-th adjusted sub-object feature information of the object data, includes: for the k-th target candidate data among the Kj target candidate data, determining the weight of the k-th target candidate data based on the similarity between the object data and the k-th target candidate data; weighting the first adjustment information of the object data based on the weight of the Kj target candidate data to obtain weighted adjustment information; and determining the j-th adjusted sub-object feature information of the object data based on the j-th sub-object feature information of the object data and the weighted adjustment information.

[0370] The specific implementation process of S404 can be referred to the relevant description of S204 above, and will not be repeated here.

[0371] S405. Through the classification module, the R adjusted sub-object feature information of the object data are classified to obtain the classification result of the object data.

[0372] In some embodiments, the classification module includes a hybrid expert module, which comprises a gating unit and E expert units. The classification module performs classification processing on the R adjusted sub-object feature information of the object data to obtain the classification result of the object data, including:

[0373] Based on R adjusted sub-object feature information of the object data, S input tokens are obtained, where S is a positive integer greater than 1. For example, each adjusted sub-object feature information in the R adjusted sub-object feature information of the object data is taken as a row of the first matrix, resulting in a first matrix with R rows; each column of the first matrix is ​​taken as an input token, resulting in S input tokens.

[0374] Next, for the s-th input token out of S input tokens, a gating unit determines the probability value of the s-th input token being assigned to each of the E expert units, where E is a positive integer greater than 1 and s is a positive integer less than or equal to S. Based on the probability value, T expert units are selected from the E expert units to process the s-th input token, and the s-th input token is processed by each of the T expert units to obtain the fused feature information of the s-th input token, where T is a positive integer less than E. For example, for each of the T expert units, feature fusion processing is performed on each input token assigned to the expert unit to obtain the output feature information of each input token, including the s-th input token. Based on the probability value of the s-th input token being assigned to each of the T expert units, the weights corresponding to the T expert units are determined. Based on the weights corresponding to the T expert units, the output feature information of the s-th input token output by the T expert units is weighted to obtain the fused feature information of the s-th input token.

[0375] Then, based on the fused feature information of the S input tokens, the mixed feature information of the object data is obtained; based on the mixed feature information, the object data is classified to obtain the classification result of the object data.

[0376] The specific implementation process of S405 can be referred to the relevant description of S205 above, and will not be repeated here.

[0377] In this embodiment, multiple extractions are performed on each candidate data point to extract R sub-candidate feature information. Then, the features of the object data are adjusted based on these R sub-candidate features, allowing for different adjustments to the object data's features in different linear spaces, thereby enriching the object data's feature information. This approach, based on the rich feature information of the object data, can improve the accuracy of object data classification.

[0378] The following uses L biological state parameters, including L gene mutations, and classification results including probability reference values ​​of object data belonging to multiple cancer categories as an example, combined with Figure 14, to introduce the data classification process of this application embodiment.

[0379] Figure 16 is a flowchart illustrating a data classification method provided in an embodiment of this application. As shown in Figure 16, the data classification method of this embodiment includes the following steps:

[0380] S501, Obtain object data and Q candidate data.

[0381] The object data includes the status values ​​of L types of gene mutations of the classified object, and the candidate data includes the status values ​​of L types of gene mutations of other reference objects that are different from the classified object. Q and L are both positive integers greater than 1.

[0382] For example, as shown in Figure 17, the first dashed circle includes the status values ​​of L types of gene mutations of the classification object, including ABL1, ACVR1, AKT1, etc., and the remaining dashed circles include the status values ​​of L types of gene mutations of other reference objects.

[0383] S502. For the j-th feature extraction unit among the R feature extraction units, the j-th sub-object feature information of the object data and the j-th sub-candidate feature information of the Q candidate data are extracted through the j-th feature extraction unit.

[0384] The specific implementation process of S502 can be referred to the relevant description of S202 above, and will not be repeated here.

[0385] S503. Based on the R sub-object feature information of the object data and the R sub-candidate feature information of the Q candidate data, select Kj target candidate data that satisfy the correlation condition with the object data from the Q candidate data.

[0386] The specific implementation process of S503 can be referred to the relevant description of S203 above, and will not be repeated here.

[0387] S504. Based on the j-th sub-sub-feature information of Kj target candidate data, adjust the j-th sub-object feature information of the object data to obtain the j-th adjusted sub-object feature information of the object data.

[0388] As shown in Figure 17, the electronic device inputs the object data and Q candidate data into R feature extraction units respectively. Each feature extraction unit processes the object data and the Q candidate data, and finally obtains R adjusted sub-object feature information of the object data, such as h1, h2...hr. These R adjusted sub-object feature information are concatenated to obtain the first matrix H.

[0389] The specific implementation process of S504 can be referred to the relevant description of S204 above, and will not be repeated here.

[0390] S505. Through the classification module, the R adjusted sub-object feature information of the object data are classified to obtain the probability reference value of the classified object belonging to the preset cancer category.

[0391] In this embodiment of the application, the probability reference value of the classified object belonging to a preset cancer category can be understood as the classification reference result of the classification belonging to the preset cancer category, so that relevant personnel can diagnose the classified object based on the classification reference result.

[0392] In some embodiments, as shown in Figure 17, the classification module includes a hybrid expert module, which includes a gating unit and E expert units. The classification module performs classification processing on the R adjusted sub-object feature information of the object data under R feature extraction units to obtain a probability reference value for the object data belonging to a preset cancer category. This includes: obtaining S input tokens based on the adjusted sub-object feature information of the object data under R feature extraction units, where S is a positive integer greater than 1; for the s-th input token among the S input tokens, the gating unit determines that the s-th input token is classified. A probability value is assigned to each of the E expert units, where E is a positive integer greater than 1, and s takes the values ​​1, 2, ..., S. Based on the probability value, T expert units are selected from the E expert units to process the s-th input token, and the s-th input token is processed by the T expert units to obtain the fused feature information of the s-th input token, where T is a positive integer less than E. Based on the fused feature information of the S input tokens, the mixed feature information of the object data is obtained. As shown in Figure 17, classification processing is performed based on the mixed feature information to obtain the probability reference value of the classified object belonging to multiple cancer classifications.

[0393] The specific implementation process of S505 can be referred to the relevant description of S205 above, and will not be repeated here.

[0394] This application employs a hybrid expert module to extract the interrelationships between different gene mutations and incorporates interpretability considerations, which helps in discovering potential unknown mutations in specific cancer types. This application's embodiments start from two dimensions: patient-level and gene mutation status, learning the deep relationships within different patients and different gene mutation statuses to classify different types of cancer. This yields classification reference results showing that the classified object belongs to multiple different cancer types, thereby improving the detection range and accuracy of early-stage cancer.

[0395] The method embodiments of this application have been described in detail above with reference to Figures 2 to 17. The apparatus embodiments of this application have been described in detail below with reference to Figures 18 and 19.

[0396] Figure 18 is a schematic block diagram of a data classification device provided in an embodiment of this application.

[0397] As shown in Figure 18, the data classification device 10 includes:

[0398] Extraction unit 11 is used for object data and Q candidate data, and extracts object feature information of the object data and candidate feature information of the Q candidate data respectively. The object data includes parameter values ​​of L biological state parameters of the classified object, and the candidate data includes parameter values ​​of L biological state parameters of the reference object that is different from the classified object. Q and L are both positive integers greater than 1.

[0399] Selection unit 12 is used to select at least one target candidate data that satisfies a preset correlation condition with the object data from the Q candidate data based on the object feature information and the candidate feature information;

[0400] Adjustment unit 13 is used to adjust the feature information based on the candidate feature information of the at least one target candidate data to obtain the adjusted object feature information of the object data;

[0401] The classification unit 14 is used to classify the adjusted object feature information to obtain the classification result of the object data. The classification result includes the probability reference value of the classified object belonging to multiple preset categories.

[0402] In some embodiments, the L biological state parameters include L gene mutations, and the classification result includes a probability reference value for the classification object belonging to multiple cancer classifications.

[0403] In some embodiments, the classification unit 14 specifically classifies the adjusted object feature information using a data classification model. The data classification model includes a retrieval module and a classification module. The retrieval module includes R feature extraction units, where R is a positive integer greater than 1. The object feature information includes R sub-object feature information, and the candidate feature information includes R sub-candidate feature information. The extraction unit 11 is specifically used to extract features from the object data using the j-th feature extraction unit to obtain the j-th sub-object feature information of the object data, and to extract features from the Q candidate data using the j-th feature extraction unit to obtain the j-th sub-candidate feature information of the Q candidate data, where the value of j is... The numbers are 1, 2, ..., R; the selection unit 12 is specifically used to select Kj target candidate data that satisfy a preset correlation condition with the object data from the Q candidate data based on the R sub-object feature information of the object data and the R sub-candidate feature information of the Q candidate data, where Kj is a positive integer; the adjustment unit 13 is specifically used to adjust the j-th sub-object feature information of the object data based on the j-th sub-candidate feature information of the Kj target candidate data to obtain the j-th adjusted sub-object feature information of the object data; the classification unit 14 is specifically used to classify the R adjusted sub-object feature information of the object data through the classification module to obtain the classification result of the object data.

[0404] In some embodiments, the R feature extraction units are R encoders. The extraction unit 11 is specifically used to extract features from the object data through the j-th encoder to obtain the j-th sub-object feature information of the object data, and to extract features from the Q candidate data through the j-th encoder to obtain the j-th sub-candidate feature information of the Q candidate data.

[0405] In some embodiments, the selection unit 12 is specifically used to determine the similarity between the object data and each of the Q candidate data based on the R sub-object feature information of the object data and the R sub-candidate feature information of the Q candidate data; and to select the Kj target candidate data from the Q candidate data based on the similarity, wherein the similarity is used to characterize the correlation between the object data and the candidate data.

[0406] In some embodiments, the selection unit 12 is specifically used to: project the j-th sub-candidate feature information of the q-th candidate data using a first projection matrix to obtain the j-th projected feature information of the q-th candidate data, where q takes the value 1, 2, ..., Q; project the j-th sub-object feature information of the object data using the first projection matrix to obtain the j-th projected feature information of the object data; determine the distance between the R projected feature information of the object data and the R projected feature information of the q-th candidate data, and determine the similarity between the object data and the q-th candidate data based on the distance.

[0407] In some embodiments, the adjustment unit 13 is specifically configured to, for the kth target candidate data among the Kj target candidate data, determine the distance between the jth projection feature information of the object data and the jth projection feature information of the kth target candidate data, and based on the distance, determine the first adjustment information of the object data by the kth target candidate data, where the value of k is 1, 2, ..., Kj; and based on the first adjustment information of the object data by the Kj target candidate data respectively, adjust the jth sub-object feature information of the object data to obtain the jth adjusted sub-object feature information of the object data.

[0408] In some embodiments, the adjustment unit 13 is specifically used to perform linear activation processing on the distance and then random deactivation processing, and to perform lossless linear transformation on the random deactivation processing result to obtain the second adjustment information of the kth target candidate data on the object data; based on the second adjustment information of the kth target candidate data on the object data and the label information of the kth target candidate data, to determine the first adjustment information of the kth target candidate data on the object data.

[0409] In some embodiments, the adjustment unit 13 is specifically used to project the label information of the kth target candidate data using a second projection matrix to obtain the label projection information of the kth target candidate data; and to determine the sum of the second adjustment information of the kth target candidate data on the object data and the label projection information of the kth target candidate data as the first adjustment information of the kth target candidate data on the object data.

[0410] In some embodiments, the adjustment unit 13 is specifically configured to, for the kth target candidate data among the Kj target candidate data, determine the weight of the kth target candidate data based on the similarity between the object data and the kth target candidate data; based on the weight of the Kj target candidate data, perform weighted processing on the first adjustment information of the object data by the Kj target candidate data to obtain weighted adjustment information; and determine the jth adjusted sub-object feature information of the object data based on the jth sub-object feature information of the object data and the weighted adjustment information.

[0411] In some embodiments, the classification module includes a hybrid expert module, which includes a gating unit and E expert units. The classification unit 14 is specifically used to obtain S input tokens based on R adjusted sub-object feature information of the object data, where S is a positive integer greater than 1; for the s-th input token among the S input tokens, the gating unit determines the probability value of the s-th input token being assigned to each of the E expert units, where E is a positive integer greater than 1, and s takes the value 1, 2, ..., S; based on the probability value, T expert units are selected from the E expert units to process the s-th input token, and the s-th input token is processed by the T expert units to obtain fused feature information of the s-th input token, where T is a positive integer less than E; based on the fused feature information of the S input tokens, hybrid feature information of the object data is obtained; based on the hybrid feature information, the object data is classified to obtain the classification result of the object data.

[0412] In some embodiments, the classification unit 14 is specifically configured to, for each of the T expert units, perform feature fusion processing on each input token assigned to the expert unit to obtain output feature information of each input token, wherein each input token includes the s-th input token; determine the weights corresponding to the T expert units based on the probability value of the s-th input token being assigned to each of the T expert units; and perform weighted processing on the output feature information of the s-th input token output by the T expert units based on the weights corresponding to the T expert units to obtain the fused feature information of the s-th input token.

[0413] In some embodiments, the classification unit 14 is specifically used to take each of the R adjusted sub-object feature information from the object data as a row of a first matrix, to obtain a first matrix including R rows; and to take each column of the first matrix as an input token, to obtain the S input tokens.

[0414] It should be understood that the device embodiments and method embodiments can correspond to each other, and similar descriptions can be referred to the method embodiments. To avoid repetition, they will not be repeated here. Specifically, the device shown in FIG18 can execute the above-described data classification method embodiments, and the foregoing and other operations and / or functions of each module in the device are respectively for implementing the above-described method embodiments, which will not be repeated here for the sake of brevity.

[0415] Figure 19 is a schematic block diagram of a model training device provided in an embodiment of this application.

[0416] As shown in Figure 19, the model training device 20 includes:

[0417] Extraction unit 21 is used to acquire training samples and Q candidate samples, and extract object feature information of the training samples and candidate feature information of the Q candidate samples respectively through the retrieval module in the data classification model. The training samples include parameter values ​​of L biological state parameters of the first object, and the candidate samples include parameter values ​​of L biological state parameters of the reference object that is different from the first object. Q and L are both positive integers greater than 1.

[0418] Selection unit 22 is used to select at least one target candidate sample from the Q candidate samples that satisfies a preset correlation condition with the training sample based on the object feature information and the candidate feature information.

[0419] The adjustment unit 23 is used to adjust the object feature information based on the candidate feature information of the at least one target candidate sample to obtain the adjusted object feature information of the training sample.

[0420] The classification unit 24 is used to classify the adjusted object feature information through the classification module in the data classification model to obtain the classification result of the training sample. The classification result includes the probability reference value of the first object belonging to multiple preset categories.

[0421] Training unit 25 is used to determine the loss of the data classification model based on the classification results of the training samples, and to train the data classification model based on the loss.

[0422] In some embodiments, the retrieval module includes R feature extraction units, where R is a positive integer greater than 1; the object feature information includes R sub-object feature information; and the candidate feature information includes R sub-candidate feature information. The extraction unit 21 is specifically configured to, for the j-th feature extraction unit among the R feature extraction units, extract features from the training sample to obtain the j-th sub-object feature information of the training sample, and to extract features from the Q candidate samples through the j-th feature extraction unit to obtain the j-th sub-candidate feature information of the Q candidate samples, where j takes the value 1, 2, ..., R. The selection unit 22 has... The system is configured to select Kj target candidate samples from the Q candidate samples that satisfy a preset correlation condition with the training sample, based on the R sub-object feature information of the training sample and the R sub-candidate feature information of the Q candidate samples; the adjustment unit 23 is specifically configured to adjust the j-th sub-object feature information of the training sample based on the j-th sub-candidate feature information of the Kj target candidate samples, to obtain the j-th adjusted sub-object feature information of the training sample; the classification unit 24 is specifically configured to classify the R adjusted sub-object feature information of the training sample through the classification module, to obtain the classification result of the training sample.

[0423] In some embodiments, the classification module includes a hybrid expert module, which includes a gating unit and E expert units. The classification unit 24 is specifically used to obtain S input tokens based on R adjusted sub-object feature information of the object data, where S is a positive integer greater than 1; for the s-th input token among the S input tokens, the gating unit determines the probability value of the s-th input token being assigned to each of the E expert units, where E is a positive integer greater than 1 and s is a positive integer less than or equal to S; based on the probability value, T expert units are selected from the E expert units to process the s-th input token, and the s-th input token is processed by the T expert units to obtain fused feature information of the s-th input token, where T is a positive integer less than E; based on the fused feature information of the S input tokens, hybrid feature information of the object data is obtained; based on the hybrid feature information, the object data is classified to obtain the classification result of the object data.

[0424] In some embodiments, the training unit 25 is specifically configured to determine a first loss based on the classification results of the training samples and the label information of the training samples; determine a second loss based on the number of input labels assigned to each of the E expert units and the assignment probability value; and determine the loss of the data classification model based on the first loss and the second loss.

[0425] It should be understood that the device embodiments and method embodiments can correspond to each other, and similar descriptions can be referred to the method embodiments. To avoid repetition, they will not be repeated here. Specifically, the device shown in FIG19 can execute the above-described model training method embodiments, and the foregoing and other operations and / or functions of each module in the device are respectively for implementing the above-described method embodiments, which will not be repeated here for the sake of brevity.

[0426] The apparatus of this application embodiment has been described above from the perspective of functional modules in conjunction with the accompanying drawings. It should be understood that this functional module can be implemented in hardware, in software instructions, or in a combination of hardware and software modules. Specifically, the steps of the method embodiments in this application can be completed by integrated logic circuits in the processor's hardware and / or by software instructions. The steps of the method disclosed in this application embodiment can be directly embodied as being executed by a hardware decoding processor, or by a combination of hardware and software modules in the decoding processor. Optionally, the software module can reside in a mature storage medium in the art, such as random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, etc. This storage medium is located in memory, and the processor reads information from the memory and, in conjunction with its hardware, completes the steps in the above method embodiments.

[0427] Figure 20 is a schematic block diagram of an electronic device provided in an embodiment of this application. The electronic device may be the terminal device or server described above.

[0428] As shown in Figure 20, the electronic device 40 may include:

[0429] The system includes a memory 41 and a processor 42. The memory 41 stores a computer program 43 and transfers the program code 43 to the processor 42. In other words, the processor 42 can retrieve and run the computer program 43 from the memory 41 to implement the methods described in the embodiments of this application.

[0430] For example, the processor 42 can be used to execute the steps in the above method according to the instructions in the computer program 43.

[0431] In some embodiments of this application, the processor 42 may include, but is not limited to:

[0432] General-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc.

[0433] In some embodiments of this application, the memory 41 includes, but is not limited to:

[0434] Volatile memory and / or non-volatile memory. Non-volatile memory can be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), or flash memory. Volatile memory can be random access memory (RAM), which is used as an external cache. By way of example, but not limitation, many forms of RAM are available, such as Static RAM (SRAM), Dynamic RAM (DRAM), Synchronous DRAM (SDRAM), Double Data Rate SDRAM (DDR SDRAM), Enhanced SDRAM (ESDRAM), Synchronous Link DRAM (SLDRAM), and Direct Rambus RAM (DR RAM).

[0435] In some embodiments of this application, the computer program 43 may be divided into one or more modules, which are stored in the memory 41 and executed by the processor 42 to complete the page recording method provided in this application. The one or more modules may be a series of computer program instruction segments capable of performing specific functions, which describe the execution process of the computer program 43 in the electronic device.

[0436] As shown in Figure 20, the electronic device 40 may further include:

[0437] Transceiver 44, which can be connected to processor 42 or memory 41.

[0438] The processor 42 can control the transceiver 44 to communicate with other devices; specifically, it can send information or data to other devices or receive information or data sent by other devices. The transceiver 44 may include a transmitter and a receiver. The transceiver 44 may further include antennas, and the number of antennas may be one or more.

[0439] It should be understood that the various components in the electronic device 40 are connected through a bus system, which includes a data bus, a power bus, a control bus, and a status signal bus.

[0440] According to one aspect of this application, a computer storage medium is provided that stores a computer program thereon, which, when executed by a computer, enables the computer to perform the methods of the above-described method embodiments. Alternatively, embodiments of this application also provide a computer program product containing instructions that, when executed by a computer, cause the computer to perform the methods of the above-described method embodiments.

[0441] According to another aspect of this application, a computer program product or computer program is provided, comprising computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the method described in the above-described method embodiments.

[0442] In other words, when implemented using software, it can be implemented wholly or partially in the form of a computer program product. This computer program product includes one or more computer instructions. When these computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., digital video disc (DVD)), or a semiconductor medium (e.g., solid-state disk (SSD)).

[0443] Those skilled in the art will recognize that the modules and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0444] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or modules may be electrical, mechanical, or other forms.

[0445] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical modules; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. For example, the functional modules in the various embodiments of this application may be integrated into one processing module, or each module may exist physically separately, or two or more modules may be integrated into one module.

[0446] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A data classification method, characterized in that, include: Obtain object data and Q candidate data, and extract object feature information from the object data and candidate feature information from the Q candidate data respectively. The object data includes parameter values ​​of L biological state parameters of the classified object, and the candidate data includes parameter values ​​of L biological state parameters of the reference object that is different from the classified object. Q and L are both positive integers greater than 1. Based on the object feature information and the candidate feature information, at least one target candidate data that satisfies a preset correlation condition with the object data is selected from the Q candidate data; Based on the candidate feature information of the at least one target candidate data, the object feature information is adjusted to obtain the adjusted object feature information of the object data; The adjusted object feature information is classified to obtain the classification result of the object data. The classification result includes the probability reference value of the classified object belonging to multiple preset categories.

2. The method according to claim 1, characterized in that, The L biological state parameters include L gene mutations, and the classification results include probability reference values ​​for the classification object belonging to multiple cancer classifications.

3. The method according to claim 1 or 2, characterized in that, The adjusted object feature information is classified using a data classification model, which includes a retrieval module and a classification module. The retrieval module includes R feature extraction units, where R is a positive integer greater than 1. The object feature information includes R sub-object feature information, and the candidate feature information includes R sub-candidate feature information. The step of extracting object feature information of the object data and candidate feature information of the Q candidate data respectively includes: extracting features from the object data through the j-th feature extraction unit to obtain the j-th sub-object feature information of the object data, and extracting features from the Q candidate data through the j-th feature extraction unit to obtain the j-th sub-candidate feature information of the Q candidate data, wherein the value of j is 1, 2, ..., R; The step of selecting at least one object data that satisfies a preset correlation condition with the object data from the Q candidate data based on the object feature information and the candidate feature information includes: selecting Kj target candidate data that satisfy a preset correlation condition with the object data from the Q candidate data based on R sub-object feature information of the object data and R candidate feature information of the Q candidate data, where Kj is a positive integer; The step of adjusting the object feature information based on the candidate feature information of the at least one target candidate data to obtain the adjusted object feature information of the object data includes: adjusting the j-th sub-object feature information of the object data based on the j-th sub-candidate feature information of the Kj target candidate data to obtain the j-th adjusted sub-object feature information of the object data; The step of classifying the adjusted object feature information to obtain the classification result of the object data includes: classifying the R adjusted sub-object feature information of the object data through the classification module to obtain the classification result of the object data.

4. The method according to claim 3, characterized in that, The R feature extraction units are R encoders. The step of extracting features from the object data using the j-th feature extraction unit to obtain the j-th sub-object feature information of the object data, and extracting features from the Q candidate data using the j-th feature extraction unit to obtain the j-th sub-candidate feature information of the Q candidate data, includes: The j-th encoder is used to extract features from the object data to obtain the j-th sub-object feature information of the object data, and the j-th encoder is used to extract features from the Q candidate data to obtain the j-th sub-candidate feature information of the Q candidate data.

5. The method according to claim 3, characterized in that, The selection of Kj target candidate data that satisfy a preset correlation condition with the object data from the Q candidate data, based on the R sub-object feature information of the object data and the R sub-candidate feature information of the Q candidate data, includes: Based on the R sub-object feature information of the object data and the R sub-candidate feature information of the Q candidate data, the similarity between the object data and each of the Q candidate data is determined, and the similarity is used to characterize the correlation between the object data and the candidate data; Based on the similarity, Kj target candidate data are selected from the Q candidate data.

6. The method according to claim 5, characterized in that, The determination of the similarity between the object data and each of the Q candidate data based on the R sub-object feature information of the object data and the R sub-candidate feature information of the Q candidate data includes: For the q-th candidate data among the Q candidate data, the j-th sub-candidate feature information of the q-th candidate data is projected using the first projection matrix to obtain the j-th projected feature information of the q-th candidate data, where the value of q is 1, 2, ..., Q; The first projection matrix is ​​used to project the j-th sub-object feature information of the object data to obtain the j-th projected feature information of the object data; Determine the distance between the R projected feature information of the object data and the R projected feature information of the q-th candidate data, and determine the similarity between the object data and the q-th candidate data based on the distance.

7. The method according to claim 6, characterized in that, The step of adjusting the j-th sub-object feature information of the object data based on the j-th candidate sub-feature information of the Kj target candidate data to obtain the j-th adjusted sub-object feature information of the object data includes: For the kth target candidate data among the Kj target candidate data, determine the distance between the jth projection feature information of the object data and the jth projection feature information of the kth target candidate data, and based on the distance, determine the first adjustment information of the kth target candidate data on the object data, where the value of k is 1, 2, ..., Kj; Based on the first adjustment information of the object data according to the Kj target candidate data, the j-th sub-object feature information of the object data is adjusted to obtain the j-th adjusted sub-object feature information of the object data.

8. The method according to claim 7, characterized in that, The step of determining the first adjustment information of the object data based on the k-th target candidate data based on the distance includes: The distance is linearly activated and then randomly deactivated. The result of the random deactivation is then subjected to a lossless linear transformation to obtain the second adjustment information of the object data for the kth target candidate data. Based on the second adjustment information of the object data to the k-th target candidate data and the label information of the k-th target candidate data, the first adjustment information of the object data to the k-th target candidate data is determined.

9. The method according to claim 8, characterized in that, The determination of the first adjustment information of the object data based on the second adjustment information of the object data based on the kth target candidate data and the label information of the kth target candidate data includes: The label information of the kth target candidate data is projected using the second projection matrix to obtain the label projection information of the kth target candidate data; The sum of the second adjustment information of the kth target candidate data on the object data and the label projection information of the kth target candidate data is determined as the first adjustment information of the kth target candidate data on the object data.

10. The method according to claim 7, characterized in that, The step of adjusting the j-th sub-object feature information of the object data based on the first adjustment information of the Kj target candidate data, and determining the j-th adjusted sub-object feature information of the object data, includes: For the kth target candidate data among the Kj target candidate data, the weight of the kth target candidate data is determined based on the similarity between the object data and the kth target candidate data; Based on the weights of the Kj target candidate data, the first adjustment information of the object data by the Kj target candidate data is weighted to obtain the weighted adjustment information; Based on the feature information of the j-th sub-object of the object data and the weighted adjustment information, the adjusted feature information of the j-th sub-object of the object data is determined.

11. The method according to claim 3, characterized in that, The classification module includes a hybrid expert module, which comprises a gating unit and E expert units. The classification module performs classification processing on R adjusted object feature information of the object data to obtain the classification result of the object data, including: Based on the R adjusted sub-object feature information of the object data, S input tokens are obtained, where S is a positive integer greater than 1; For the s-th input token among the S input tokens, the gating unit determines the probability value of the s-th input token being assigned to each of the E expert units, where E is a positive integer greater than 1, and s takes the value 1, 2, ..., S; Based on the probability value, T expert units are selected from the E expert units to process the s-th input token, and the s-th input token is processed by the T expert units respectively to obtain the fusion feature information of the s-th input token, where T is a positive integer less than E; Based on the fusion feature information of the S input tokens, the hybrid feature information of the object data is obtained; The mixed feature information is classified to obtain the classification result of the object data.

12. The method according to claim 11, characterized in that, The process of processing the s-th input token through the T expert units to obtain the fusion feature information of the s-th input token includes: For each of the T expert units, feature fusion processing is performed on each input token assigned to the expert unit to obtain the output feature information of each input token, wherein each input token includes the s-th input token; Based on the probability value of the s-th input token being assigned to each of the T expert units, the weights corresponding to the T expert units are determined respectively; Based on the weights corresponding to the T expert units, the output feature information of the s-th input token output by the T expert units is weighted to obtain the fusion feature information of the s-th input token.

13. The method according to claim 11, characterized in that, Based on the R adjusted sub-object feature information of the object data, S input tokens are obtained, including: Each of the R adjusted sub-object feature information is used as a row of the first matrix to obtain a first matrix containing R rows. Each column of the first matrix is ​​used as an input token to obtain the S input tokens.

14. The method according to any one of claims 1-4, characterized in that, Based on the object feature information and the candidate feature information, at least one target candidate data that satisfies a preset correlation condition with the object data is selected from the Q candidate data, including: Cluster the object feature information and the candidate feature information of the Q candidate data; The candidate data that belongs to the same cluster as the object data among the Q candidate data are determined as the target candidate data.

15. The method according to any one of claims 1-4, characterized in that, Based on the object feature information and the candidate feature information, at least one target candidate data that satisfies a preset correlation condition with the object data is selected from the Q candidate data, including: Based on the object feature information and the candidate feature information of the Q candidate data, the similarity between the object data and each of the Q candidate data is determined; The at least one target candidate data is determined based on the similarity between the object data and the Q candidate data.

16. A model training method, characterized in that, include: A training sample and Q candidate samples are obtained, and the object feature information of the training sample and the candidate feature information of the Q candidate samples are extracted by the retrieval module in the data classification model. The training sample includes the parameter values ​​of L biological state parameters of the first object, and the candidate sample includes the parameter values ​​of L biological state parameters of the reference object that is different from the first object. Q and L are both positive integers greater than 1. Based on the object feature information and the candidate feature information, at least one target candidate sample that satisfies the preset correlation condition with the training sample is selected from the Q candidate samples; Based on the candidate feature information of the at least one target candidate sample, the object feature information is adjusted to obtain the adjusted object feature information of the training sample; The adjusted object feature information of the training samples is classified by the classification module in the data classification model to obtain the classification result of the training samples. The classification result includes the probability reference value of the first object belonging to multiple preset categories. Based on the classification results of the training samples, the loss of the data classification model is determined, and the data classification model is trained based on the loss.

17. The method according to claim 16, characterized in that, The retrieval module includes R feature extraction units, where R is a positive integer greater than 1; the object feature information includes R sub-object feature information; and the candidate feature information includes R sub-candidate feature information. The step of extracting object feature information of the training samples and candidate feature information of the Q candidate samples through the retrieval module in the data classification model includes: for the j-th feature extraction unit among the R feature extraction units, performing feature extraction on the training samples through the j-th feature extraction unit to obtain the j-th sub-object feature information of the training samples, and performing feature extraction on the Q candidate samples through the j-th feature extraction unit to obtain the j-th sub-candidate feature information of the Q candidate samples, where the value of j is 1, 2, ..., R; The step of selecting at least one target candidate sample from the Q candidate samples that satisfies a preset correlation condition with the training sample based on the object feature information and the candidate feature information includes: selecting Kj target candidate samples from the Q candidate samples that satisfy a preset correlation condition with the training sample based on R sub-object feature information of the training sample and R sub-candidate feature information of the Q candidate samples. The step of adjusting the object feature information of the training sample based on the candidate feature information of the at least one target candidate sample to obtain the adjusted object feature information of the training sample includes: adjusting the j-th sub-object feature information of the training sample based on the j-th sub-candidate feature information of the Kj target candidate samples to obtain the j-th adjusted sub-object feature information of the training sample. The step of classifying the adjusted object feature information of the training sample through the classification module in the data classification model to obtain the classification result of the training sample includes: classifying the R adjusted sub-object feature information of the training sample through the classification module to obtain the classification result of the training sample.

18. The method according to claim 17, characterized in that, The classification module includes a hybrid expert module, which comprises a gating unit and E expert units. Determining the loss of the data classification model based on the classification results of the training samples includes: Based on the classification results and label information of the training samples, a first loss is determined; The second loss is determined based on the number of input labels assigned to each of the E expert units and the assignment probability value; Based on the first loss and the second loss, the loss of the data classification model is determined.

19. A data classification device, characterized in that, include: An extraction unit is used to acquire object data and Q candidate data, and to extract object feature information of the object data and candidate feature information of the Q candidate data respectively. The object data includes parameter values ​​of L biological state parameters of a preset object, and the candidate data includes parameter values ​​of L biological state parameters of a reference object that is different from the classified object. Q and L are both positive integers greater than 1. The selection unit is used to select at least one target candidate data that satisfies a preset correlation condition with the object data from the Q candidate data based on the object feature information and the candidate feature information. An adjustment unit is configured to adjust the object feature information based on the candidate feature information of the at least one target candidate data, so as to obtain the adjusted object feature information of the object data; The classification unit is used to classify the adjusted object feature information of the object data to obtain the classification result of the object data. The classification result includes the probability reference value of the classified object belonging to multiple preset categories.

20. A model training device, characterized in that, include: An extraction unit is used to acquire training samples and Q candidate samples, and to extract object feature information of the training samples and candidate feature information of the Q candidate samples through the retrieval module in the data classification model. The training samples include preset parameter values ​​of L biological state parameters of a first object, and the candidate samples include reference parameter values ​​of L biological state parameters of a reference object different from the first object. Q and L are both positive integers greater than 1. The selection unit is used to select at least one target candidate sample from the Q candidate samples that satisfies a preset correlation condition with the training sample, based on the object feature information and the candidate first feature information. An adjustment unit is configured to adjust the object feature information based on the candidate feature information of the at least one target candidate sample to obtain the adjusted object feature information of the training sample. The classification unit is used to classify the adjusted object feature information through the classification module in the data classification model to obtain the classification result of the training sample. The classification result includes the probability reference value of the first object belonging to multiple preset categories. The training unit is used to determine the loss of the data classification model based on the classification results of the training samples, and to train the data classification model based on the loss.

21. An electronic device, characterized in that, Including processor and memory; The memory is used to store computer programs; The processor is configured to execute the computer program to implement the method as described in any one of claims 1 to 15 or 16 to 18.

22. A computer-readable storage medium, characterized in that, Used to store computer programs; The computer program causes the computer to perform the method as described in any one of claims 1 to 15 or 16 to 18.

Images