Clustering model training method, clustering method, device, equipment and storage medium

By adjusting the neural network model by using labeled clustering results and predicting clustering results, the problem of low accuracy of clustering models is solved, and the accuracy of clustering results is significantly improved.

CN114462540BActive Publication Date: 2024-11-19TENCENT TECHNOLOGY (SHENZHEN) CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202210126447.8
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2022-02-10
Publication Date
2024-11-19
Estimated Expiration
2042-02-10

AI Technical Summary

Technical Problem

The clustering results obtained by existing clustering models are poor in the training process, resulting in low accuracy of clustering results.

Method used

By acquiring the annotated clustering results of the multiple first object data and the second object data, and the predicted clustering results based on these data, the neural network model is adjusted to optimize the clustering model to improve its accuracy.

Benefits of technology

The model is guided to accurately cluster by labeling clustering results with a small amount of data, which significantly improves the accuracy of the clustering model and thus improves the accuracy of the clustering results.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN114462540B_ABST
    Figure CN114462540B_ABST
Patent Text Reader

Abstract

The present application discloses a training method, clustering method, device, equipment and storage medium for a clustering model, and belongs to the field of artificial intelligence technology. The method includes: obtaining a plurality of first object data, a plurality of second object data and annotated clustering results of each first object data; obtaining a first predicted clustering result of each first object data and a first predicted clustering result of each second object data according to a first network model; adjusting the first network model based on the annotated clustering results of each first object data, the first predicted clustering results and the first predicted clustering results of each second object data to obtain a clustering model. The annotated clustering results of a small amount of data are used to guide the model to perform accurate clustering, thereby improving the accuracy of the clustering model and thus improving the accuracy of the clustering results.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The embodiments of the present application relate to the field of artificial intelligence technology, and in particular to a clustering model training method, clustering method, device, equipment and storage medium. Background Art

[0002] With the development of artificial intelligence technology, the application of neural network models is becoming more and more extensive. Neural network models can perform clustering processing on multiple object data to obtain clustering results of each object data. This neural network model used for clustering processing is also called a clustering model.

[0003] In the related technology, when training a clustering model, multiple sample object data are obtained. The multiple sample object data are input into a neural network model, and the neural network model outputs a predicted clustering result of each sample object data. The neural network model is adjusted based on the predicted clustering result of each sample object data to obtain a clustering model.

[0004] The above technology only adjusts the neural network model based on the predicted clustering results of each sample object data, resulting in poor accuracy of the clustering model, thereby reducing the accuracy of the clustering results. Summary of the invention

[0005] The embodiments of the present application provide a clustering model training method, clustering method, device, equipment and storage medium, which can be used to solve the problem of poor accuracy of clustering models and low accuracy of clustering results. The technical solution includes the following contents.

[0006] On the one hand, an embodiment of the present application provides a method for training a clustering model, the method comprising:

[0007] Acquire a plurality of first object data, a plurality of second object data, and annotated clustering results of each first object data;

[0008] Acquire, according to the first network model, a first prediction clustering result of each first object data and a first prediction clustering result of each second object data;

[0009] The first network model is adjusted based on the labeled clustering results of each first object data, the first predicted clustering results, and the first predicted clustering results of each second object data to obtain a clustering model.

[0010] On the other hand, an embodiment of the present application provides a clustering method, the method comprising:

[0011] Get multiple target object data;

[0012] The clustering result of each target object data is determined according to the clustering model, wherein the clustering model is trained according to any of the clustering model training methods described above.

[0013] On the other hand, an embodiment of the present application provides a clustering model training device, the device comprising:

[0014] An acquisition module, used to acquire a plurality of first object data, a plurality of second object data and annotated clustering results of each first object data;

[0015] The acquisition module is further used to acquire the first prediction clustering result of each first object data and the first prediction clustering result of each second object data according to the first network model;

[0016] The adjustment module is used to adjust the first network model based on the labeled clustering results of each first object data, the first predicted clustering results and the first predicted clustering results of each second object data to obtain a clustering model.

[0017] In a possible implementation, the acquisition module is used to acquire the cluster to which each sample object data belongs; based on the cluster to which each sample object data belongs, divide the each sample object data into the multiple first object data and the multiple second object data; and determine the cluster to which each first object data belongs as the labeled clustering result of each first object data.

[0018] In a possible implementation, the acquisition module is used to determine, for any sample object data, a reference cluster of the sample object data from the clusters to which the various sample object data belong; in response to the cluster to which the sample object data belongs being consistent with the reference cluster, taking the sample object data as the first object data; in response to the cluster to which the sample object data belongs being inconsistent with the reference cluster, taking the sample object data as the second object data.

[0019] In a possible implementation, the acquisition module is used to determine the distance between any one of the sample object data and each other object data, where the other object data is the sample object data other than the any one of the sample object data; based on the distance between the any one of the sample object data and the each other object data, determine each other object data whose distance is not greater than a reference distance from the each other object data; and from the clusters to which the each other object data whose distance is not greater than the reference distance belongs, determine the cluster with the largest number of other object data whose distance is not greater than the reference distance as the reference cluster for the any one of the sample object data.

[0020] In a possible implementation, the acquisition module is configured to acquire the probability of each sample object data belonging to a cluster; and divide each sample object data into the plurality of first object data and the plurality of second object data based on the probability of each sample object data belonging to a cluster.

[0021] In a possible implementation, the acquisition module is used to determine, for any sample object data, cluster center data based on each sample object data in the cluster to which the any sample object data belongs, the cluster center data being used to characterize the center of the cluster to which the any sample object data belongs; determine the distance between the any sample object data and the cluster center data; and normalize the distance between the any sample object data and the cluster center data to obtain the probability of the cluster to which the any sample object data belongs.

[0022] In a possible implementation, the acquisition module is configured to, for any sample object data, in response to a probability that the probability of the cluster to which the any sample object data belongs is not less than a probability threshold, use the any sample object data as the first object data; and in response to a probability that the probability of the cluster to which the any sample object data belongs is less than the probability threshold, use the any sample object data as the second object data.

[0023] In a possible implementation, the adjustment module is used to obtain the second predicted clustering result of each first object data according to the second network model; adjust the second network model based on the labeled clustering results of each first object data, the first predicted clustering result, the second predicted clustering result and the first predicted clustering result of each second object data to obtain an adjusted second network model; adjust the first network model based on the adjusted second network model to obtain a clustering model.

[0024] In a possible implementation, the adjustment module is used to perform data enhancement processing on each of the first object data to obtain each of the data-enhanced first object data; input the each of the data-enhanced first object data into the second network model, and the second network model outputs the second predicted clustering result of the each of the first object data.

[0025] In a possible implementation, the adjustment module is used to determine a first loss value based on the labeled clustering results of the respective first object data and the second predicted clustering results; determine a second loss value based on the first predicted clustering results of the respective first object data, the second predicted clustering results and the first predicted clustering results of the respective second object data; and adjust the second network model based on the first loss value and the second loss value to obtain an adjusted second network model.

[0026] In one possible implementation, the adjustment module is used to obtain the second predicted clustering result of each second object data according to the second network model; based on the labeled clustering results, the first predicted clustering results, the second predicted clustering results of each first object data and the first predicted clustering results, the second predicted clustering results of each second object data, the second network model is adjusted to obtain an adjusted second network model.

[0027] In a possible implementation, the adjustment module is used to perform data enhancement processing on each of the second object data to obtain each of the data-enhanced second object data; input the each of the data-enhanced second object data into the second network model, and the second network model outputs the second predicted clustering result of each of the second object data.

[0028] In a possible implementation, the acquisition module is used to obtain, for any first object data, respective object features of the first object data according to respective feature extraction networks in the first network model, any two feature extraction networks in the first network model are neural networks with different structures, and any object feature of the first object data is used to describe the first object data; the respective object features of the first object data are fused to obtain fused features of the first object data; and first predicted clustering results of the first object data are determined based on the fused features of the first object data.

[0029] On the other hand, an embodiment of the present application provides a clustering device, the device comprising:

[0030] An acquisition module, used to acquire data of multiple target objects;

[0031] The determination module is used to determine the clustering result of each target object data according to the clustering model, and the clustering model is trained according to any of the clustering model training methods described above.

[0032] In a possible implementation, the determination module is used to obtain, for any target object data, respective object features of the target object data according to respective feature extraction networks in the clustering model, any two feature extraction networks in the clustering model are neural networks with different structures, and any object feature of the target object data is used to describe the target object data; to fuse the respective object features of the target object data to obtain fused features of the target object data; and to determine clustering results of the target object data based on the fused features of the target object data.

[0033] On the other hand, an embodiment of the present application provides an electronic device, which includes a processor and a memory, wherein the memory stores at least one program code, and the at least one program code is loaded and executed by the processor so that the electronic device implements any of the above-mentioned clustering model training methods or any of the above-mentioned clustering methods.

[0034] On the other hand, a computer-readable storage medium is also provided, in which at least one program code is stored, and the at least one program code is loaded and executed by a processor so that a computer implements any of the above-mentioned clustering model training methods or any of the above-mentioned clustering methods.

[0035] On the other hand, a computer program or a computer program product is also provided, wherein at least one computer instruction is stored in the computer program or the computer program product, and the at least one computer instruction is loaded and executed by a processor so that a computer implements any of the above-mentioned clustering model training methods or any of the above-mentioned clustering methods.

[0036] The technical solution provided by the embodiments of the present application brings at least the following beneficial effects:

[0037] The technical solution provided in the embodiment of the present application is to adjust the first network model based on the labeled clustering results of each first object data, the first predicted clustering results and the first predicted clustering results of each second object data to obtain a clustering model. The labeled clustering results of a small amount of data can be used to guide the model to perform accurate clustering, thereby improving the accuracy of the clustering model and thus improving the accuracy of the clustering results. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0039] Figure 1 It is a schematic diagram of a clustering model training method or an implementation environment of a clustering method provided in an embodiment of the present application;

[0040] Figure 2 It is a flow chart of a clustering model training method provided in an embodiment of the present application;

[0041] Figure 3 It is a clustering schematic diagram of a DESC algorithm provided in an embodiment of the present application;

[0042] Figure 4 is a schematic diagram of a first network model provided in an embodiment of the present application;

[0043] Figure 5 is a flow chart of a clustering method provided in an embodiment of the present application;

[0044] Figure 6 It is a training diagram of a clustering model provided in an embodiment of the present application;

[0045] Figure 7 It is a schematic diagram of a clustering result obtained by clustering a pancreatic islet cell dataset using a DESC algorithm provided in an embodiment of the present application;

[0046] Figure 8 It is a schematic diagram of a clustering result obtained by clustering a pancreatic islet cell dataset using a clustering model provided in an embodiment of the present application;

[0047] Fig. 9 This is a schematic diagram of actual clustering results of a pancreatic islet cell dataset provided in an embodiment of the present application;

[0048] Fig.10 It is a structural schematic diagram of a clustering model training device provided in an embodiment of the present application;

[0049] Fig.11 It is a structural schematic diagram of a clustering device provided in an embodiment of the present application;

[0050] Fig.12 It is a structural diagram of a terminal device provided in an embodiment of the present application;

[0051] Fig.13 It is a structural diagram of a server provided in an embodiment of the present application. DETAILED DESCRIPTION

[0052] In order to make the objectives, technical solutions and advantages of the present application clearer, the implementation methods of the present application will be further described in detail below with reference to the accompanying drawings.

[0053] Figure 1Schematic diagram of a clustering model training method or clustering method implementation environment provided in an embodiment of the present application, such as Figure 1 The implementation environment shown includes an electronic device 11, and the clustering model training method or clustering method in the embodiment of the present application can be executed by the electronic device 11. Exemplarily, the electronic device 11 can include at least one of a terminal device or a server.

[0054] The terminal device may be at least one of a smart phone, a desktop computer, a tablet computer, and a laptop computer. The server may be a single server, or a server cluster consisting of multiple servers, or any one of a cloud computing platform and a virtualization center, which is not limited in the embodiments of the present application. The server may be connected to the terminal device through a wired network or a wireless network. The server may have functions such as data processing, data storage, and data transmission and reception, which are not limited in the embodiments of the present application.

[0055] Each optional embodiment of the present application is implemented based on artificial intelligence technology. Artificial Intelligence (AI) is the theory, method, technology and application system that uses digital computers or machines controlled by digital computers to simulate, extend and expand human intelligence, perceive the environment, acquire knowledge and use knowledge to obtain the best results. In other words, artificial intelligence is a comprehensive technology in computer science that attempts to understand the essence of intelligence and produce a new intelligent machine that can respond in a similar way to human intelligence. Artificial intelligence is to study the design principles and implementation methods of various intelligent machines so that the machines have the functions of perception, reasoning and decision-making.

[0056] Artificial intelligence technology is a comprehensive discipline that covers a wide range of fields, including both hardware-level and software-level technologies. Basic artificial intelligence technologies generally include sensors, dedicated artificial intelligence chips, cloud computing, distributed storage, big data processing technology, operating / interactive systems, mechatronics and other technologies. Artificial intelligence software technologies mainly include computer vision technology, speech processing technology, natural language processing technology, as well as machine learning / deep learning, autonomous driving, smart transportation and other major directions.

[0057] With the research and advancement of artificial intelligence technology, artificial intelligence technology has been studied and applied in many fields, such as common smart homes, smart wearable devices, virtual assistants, smart speakers, smart marketing, unmanned driving, automatic driving, drones, robots, smart medical care, smart customer service, Internet of Vehicles, automatic driving, smart transportation, etc. I believe that with the development of technology, artificial intelligence technology will be applied in more fields and play an increasingly important role.

[0058] Based on the above implementation environment, the present application embodiment provides a clustering model training method. Figure 2 Taking the flowchart of a clustering model training method provided in the embodiment of the present application as an example, the method can be performed by Figure 1 The electronic device 11 in the embodiment is executed. Figure 2 As shown, the method comprises the following steps:

[0059] Step 201 : obtaining a plurality of first object data, a plurality of second object data, and annotated clustering results of each first object data.

[0060] The embodiment of the present application does not limit the first object data and the second object data. Exemplarily, the first object data and the second object data are both gene expression data of cells, or the first object data and the second object data are both image data of cells. In the embodiment of the present application, clustering results of multiple sample object data can be obtained, and multiple first object data, multiple second object data, and labeled clustering results of each first object data can be determined based on the clustering results of each sample object data.

[0061] In a possible implementation, the clustering result of any sample object includes the clustering cluster to which any sample object belongs. In this case, obtaining the plurality of first object data, the plurality of second object data, and the labeled clustering result of each first object data includes: obtaining the clustering cluster to which each sample object data belongs; dividing each sample object data into a plurality of first object data and a plurality of second object data based on the clustering cluster to which each sample object data belongs; and determining the clustering cluster to which each first object data belongs as the labeled clustering result of each first object data.

[0062] In the embodiment of the present application, any clustering algorithm can be used to perform clustering processing on each sample object data to obtain the clustering result of each sample object data. The embodiment of the present application does not limit the clustering algorithm. For example, the clustering algorithm can be a traditional clustering algorithm such as the Leiden algorithm and the Louvain algorithm, or a clustering algorithm based on deep learning such as the Deep Embedding For Single-cell Clustering (DESC) algorithm.

[0063] When different clustering algorithms are used for clustering processing, the clustering results of any sample object data are also different. It is understandable that when any clustering algorithm is used for clustering processing, the clustering result of any sample object data includes the cluster cluster to which any sample object data belongs. When some clustering algorithms such as the DESC algorithm are used for clustering processing, the clustering result of any sample object data includes not only the cluster cluster to which any sample object data belongs, but also the probability of the cluster cluster to which any sample object data belongs.

[0064] The following describes in detail how the DESC algorithm is used in an embodiment of the present application to obtain the cluster to which each sample object data belongs and the probability of the cluster to which each sample object data belongs.

[0065] The embodiment of the present application can use the DESC algorithm to implement clustering processing of multiple sample object data using a neural network model. The DESC algorithm uses autoencoding to extract the object features of each sample object data, so as to describe any sample object data using the object features of any sample object data. Based on the object features of each sample object data, the Louvain algorithm is used to perform clustering processing on each sample object data to obtain an initial clustering result of each sample object data, that is, the neural network model outputs the initial clustering result of each sample object data. Afterwards, based on the initial clustering result of each sample object data, a convolutional neural network and a loss function are used to implement iterative clustering of each sample object data until the neural network model obtains and outputs the final clustering result of each sample object data.

[0066] The clustering result of a sample object data includes the cluster to which the sample object data belongs, the probability of each sample object data belonging to the cluster to which it belongs, and the batch number of each sample object data. The DESC algorithm determines the probability of a sample object data belonging to each cluster to determine the cluster to which the sample object data belongs as the cluster corresponding to the maximum probability. The probability of the sample object data belonging to the cluster to which it belongs is recorded as the maximum probability, and the batch number of the sample object data is the sampling batch of the sample object data.

[0067] See also Figure 3 , Figure 3 It is a clustering schematic diagram of a DESC algorithm provided in an embodiment of the present application. In an embodiment of the present application, the sample object data is cell data, and the cell data can be a gene expression matrix of a cell. First, the cell data is input into the automatic encoder in the neural network model as input information, and each cell data is encoded by the automatic encoder to obtain cell features, and the cell features are used to describe the cell data, wherein the cell data is multiple. Afterwards, the neural network model performs clustering initialization (i.e., clustering process) to each cell data, and obtains the clustering result of each cell data, which is the initial clustering result.

[0068] Next, based on the initial clustering results of each cell data, the neural network model and the loss function are used to implement iterative clustering of each cell data until the neural network model obtains and outputs the final clustering results of each cell data.

[0069] During iterative clustering, the output information of the neural network model is used to determine the loss value of the loss function, and the neural network model is updated using the loss value. In this way, the neural network model is iteratively updated, thereby iteratively updating the output information of the neural network model. The neural network model determines the cell features through an automatic encoder, and determines and outputs the clustering results of each cell data based on the cell features. The loss value of the neural network model is determined based on the clustering results of each cell data and the loss function, so as to adjust the parameters of the neural network model using the loss value to update the neural network model.

[0070] Optionally, the loss function is Figure 3 Loss = KL(P||Q) shown in the figure, where Loss is the loss function, KL represents KL divergence (Kullback-Leibler Divergence), also called relative entropy (Relative Entropy), P is the true probability distribution (ie, the probability that the cell data belongs to each cluster), and Q is the pseudo-distribution of P.

[0071] It should be noted that if Figure 3 As shown, the output information of the neural network model includes three parts. Among them, the first part is the cluster (i.e., cluster cluster), and the numbers 0-5 in the cluster represent 6 clusters, each cluster includes at least one cell data, that is, the cluster is the cluster cluster to which each cell data belongs. The second part is the probability of the cell data. By determining the probability that a cell data belongs to each cluster, the maximum probability of this cell data is determined, and thus the probability of the cell data in the cluster is determined to be this maximum probability. The third part is the batch of cell data, and the same cluster includes at least one batch of cell data.

[0072] In the embodiment of the present application, when clustering each sample object data using a clustering algorithm, the cluster to which each sample object data belongs can be directly obtained. Afterwards, each sample object data can be divided into a plurality of first object data and a plurality of second object data based on the cluster to which each sample object data belongs.

[0073] In a possible implementation, based on the clusters to which the respective sample object data belong, the respective sample object data are divided into a plurality of first object data and a plurality of second object data, including: for any sample object data, determining a reference cluster of any sample object data from the clusters to which the respective sample object data belong; in response to the cluster to which any sample object data belongs being consistent with the reference cluster, taking any sample object data as first object data; in response to the cluster to which any sample object data belongs being inconsistent with the reference cluster, taking any sample object data as second object data.

[0074] The embodiment of the present application determines a reference clustering cluster of any sample object data from the clustering clusters to which each sample object data belongs according to a metric-based sampling strategy, so as to determine whether any sample object data is first object data or second object data based on the reference clustering cluster of any sample object data, thereby realizing dividing each sample object data into multiple first object data and multiple second object data.

[0075] Optionally, determining a reference clustering cluster for any sample object data from the clustering clusters to which each sample object data belongs includes: determining the distance between any sample object data and each other object data, where the other object data are sample object data other than any sample object data among the sample object data; determining each other object data having a distance not greater than a reference distance from each other object data based on the distance between any sample object data and each other object data; and determining, from the clustering clusters to which each other object data having a distance not greater than the reference distance belongs, a clustering cluster having the largest number of other object data having a distance not greater than the reference distance as the reference clustering cluster for any sample object data.

[0076] The embodiment of the present application may use a K-Nearest Neighbor (KNN) algorithm to determine a reference cluster of any sample object data from the clusters to which each sample object data belongs.

[0077] First, any sample object data and any other object data can be directly used to calculate the distance between any sample object data and any other object data. Alternatively, each sample object data can be first subjected to dimensionality reduction processing to obtain each sample object data after dimensionality reduction, and any sample object data after dimensionality reduction and any other object data after dimensionality reduction can be used to calculate the distance between any sample object data and any other object data. The embodiment of the present application does not limit the dimensionality reduction processing. Exemplarily, the dimensionality reduction processing can be principal component analysis (PCA) dimensionality reduction processing, and the dimension of any sample object data after PCA dimensionality reduction processing can be 1*50.

[0078] Alternatively, the object features of each sample object data may be extracted first, and the sample object features of the sample object data may be used to describe the sample object data, and the distance between any sample object data and any other object data may be calculated using the sample object features of any sample object data and the sample object features of any other object data. The DESC algorithm may be used to extract the object features of each sample object data, and the dimension of the object features of any sample object data may be 1*32.

[0079] When the distance between any sample object data and any other object data is not greater than the reference distance, the any other object data can be used as the neighboring object data of the any sample object data. When the distance between any sample object data and any other object data is greater than the reference distance, the any other object data can be used as the non-neighboring object data of the any sample object data. The number of neighboring object data of any sample object data is recorded as k, and they are other object data that are not greater than the reference distance, and k is a positive integer.

[0080] It should be noted that the reference distance mentioned above is the reference distance of any sample object data, and the reference distances of any two sample object data may be the same or different. The embodiment of the present application does not limit the method for determining the reference distance. Exemplarily, based on manual experience, the reference distance is determined to be 0.75, or the distances between any sample object data and each other object data are sorted from small to large, and the kth distance after sorting is used as the reference distance, where k is a positive integer.

[0081] According to the above method, each other object data corresponding to any sample object data and not greater than the reference distance can be determined. Since the cluster to which each sample object data belongs can be determined by using any clustering algorithm, the cluster to which each other object data not greater than the reference distance belongs can be determined. Afterwards, according to the voting principle, the cluster containing the largest number of other object data not greater than the reference distance is determined as the reference cluster of any sample object.

[0082] For example, the other object data corresponding to any sample object data that are not greater than the reference distance are other object data 1 to 5. Among them, other object data 1 belongs to cluster A, other object data 2 belongs to cluster A, other object data 3 belongs to cluster B, other object data 4 belongs to cluster C, and other object data 5 belongs to cluster A. Since cluster A contains three other object data that are not greater than the reference distance, and clusters B and C each contain only one other object data that is not greater than the reference distance, cluster A is the reference cluster for any sample object data.

[0083] After determining the reference cluster of any sample object, if the cluster to which any sample object data belongs is consistent with the reference cluster, any sample object data is used as the first object data, and the cluster to which the first object data belongs is determined as the labeled clustering result of the first object data. If the cluster to which any sample object data belongs is inconsistent with the reference cluster, any sample object data is used as the second object data.

[0084] In the embodiment of the present application, the set composed of each sample object data can be recorded as {f i ,i=1,…,N},f i is the i-th sample object data, and N is the number of sample object data. The clustering algorithm can be used to determine the cluster {y i ,i=1,…,N},y i is the cluster to which the i-th sample object data belongs. By determining the reference distance of any sample object data f, we can determine the k other object data corresponding to any sample object data f that are not greater than the reference distance. The set of these k other object data is recorded as N K According to the voting principle, use Determine the reference cluster of any sample object f vote is the function symbol corresponding to the voting principle. If any sample object data f belongs to the same cluster y as the reference cluster If the cluster y to which any sample object data f belongs is consistent with the reference cluster y, then any sample object data f is the first object data. If they do not match, any sample object data f is the second object data.

[0085] It should be noted that when the reference cluster of any sample object data is determined by the above method, the accuracy of the reference cluster is closely related to the number of other object data that is not greater than the reference distance (i.e., k). The more the number of other object data that is not greater than the reference distance, the higher the accuracy of the reference cluster; the fewer the number of other object data that is not greater than the reference distance, the lower the accuracy of the reference cluster.

[0086] Through the above method, the reference clustering cluster of each sample object data can be determined, and by determining whether the clustering cluster to which each sample object data belongs is consistent with the reference clustering cluster of each sample object data, multiple sample object data are divided into multiple first object data and multiple second object data, and the labeled clustering results of each first object data are obtained.

[0087] An embodiment of the present application further provides another method for dividing sample object data. Optionally, a cluster to which each sample object data belongs is obtained, and each sample object data is divided into multiple first object data and multiple second object data, including: obtaining the probability of the cluster to which each sample object data belongs; and dividing each sample object data into multiple first object data and multiple second object data based on the probability of the cluster to which each sample object data belongs.

[0088] In the embodiment of the present application, when clustering the data of each sample object using certain clustering algorithms, the probability that any sample object belongs to each cluster cluster can be obtained. At this time, the maximum probability is determined from the probabilities that any sample object belongs to each cluster cluster, and the cluster cluster corresponding to the maximum probability is used as the cluster cluster to which any sample object belongs, and the maximum probability is used as the probability of the cluster cluster to which any sample object belongs. In this way, the cluster cluster to which each sample object belongs and the probability of each sample object belonging to the cluster cluster can be determined.

[0089] When clustering each sample object data using certain clustering algorithms such as the DESC algorithm, the cluster to which each sample object data belongs and the probability of each sample object data belonging to the cluster can be directly obtained. When clustering each sample object data using certain clustering algorithms such as the Leiden algorithm, the cluster to which each sample object data belongs can be directly obtained. At this time, the probability of each sample object data belonging to the cluster can be determined based on the cluster to which each sample object data belongs.

[0090] Optionally, obtaining the probability of each sample object data belonging to a cluster includes: for any sample object data, based on each sample object data in the cluster to which any sample object data belongs, determining cluster center data, the cluster center data being used to characterize the center of the cluster to which any sample object data belongs; determining the distance between any sample object data and the cluster center data; and normalizing the distance between any sample object data and the cluster center data to obtain the probability of the cluster to which any sample object data belongs.

[0091] In the embodiment of the present application, the cluster to which each sample object data belongs can be determined. The clusters to which any two sample object data belong can be the same or different, so any cluster includes at least one sample object data, that is, any cluster to which any sample object data belongs includes at least one sample object data.

[0092] Based on each sample object data in the cluster to which any sample object data belongs, cluster center data for representing the center of the cluster to which any sample object data belongs is determined. The embodiment of the present application does not limit the method for determining the cluster center data.

[0093] For example, the cluster to which sample object data 1 belongs is cluster A, and cluster A includes sample object data 1 to 20. Then cluster center data can be determined based on sample object data 1 to 20, and the cluster center data is used to characterize the center of cluster A.

[0094] After the cluster center data is determined, the distance between any sample object data and the cluster center data is calculated based on any sample object data and the cluster center data, and the distance is normalized to [0, 1] to obtain the normalized distance. The normalized distance can be used as the probability of the cluster to which any sample object data belongs.

[0095] It can be understood that, in addition to using the normalized distance to characterize the probability of any sample object data belonging to a cluster, the embodiments of the present application may also use other indicators other than the normalized distance to represent it.

[0096] According to the above method, the probability of each sample object data belonging to a cluster can be determined. Then, based on the probability of each sample object data belonging to a cluster, each sample object data is divided into a plurality of first object data and a plurality of second object data.

[0097] Optionally, based on the probability of the cluster to which each sample object data belongs, each sample object data is divided into multiple first object data and multiple second object data, including: for any sample object data, in response to the probability that any sample object data belongs to a cluster greater than or equal to a probability threshold, any sample object data is used as the first object data; in response to the probability that any sample object data belongs to a cluster less than the probability threshold, any sample object data is used as the second object data.

[0098] The embodiment of the present application is based on a confidence-based sampling strategy (Confidence-based SamplingStrategy), based on the probability of the cluster to which any sample object data belongs, to determine whether any sample object data is the first object data or the second object data, thereby realizing dividing each sample object data into multiple first object data and multiple second object data.

[0099] Among them, the embodiments of the present application do not limit the numerical value of the probability threshold. For example, the probability threshold is determined to be 0.5 based on manual experience, or the probability average is determined based on the probability of the cluster to which each sample object data belongs, and the probability average is used as the probability threshold.

[0100] For any sample object data, if the probability of the cluster to which any sample object data belongs is greater than or equal to the probability threshold, any sample object data is used as the first object data, and at the same time, the cluster to which the first object data belongs is determined as the labeled clustering result of the object data. If the probability of the cluster to which any sample object data belongs is less than the probability threshold, any sample object data is used as the second object data.

[0101] Through the above method, it is possible to determine whether each sample object data is the first object data or the second object data, thereby realizing the division of each sample object data into each first object data and each second object data. At the same time, the cluster cluster to which each first object data belongs can be respectively determined as the labeled clustering result of the first object data.

[0102] The embodiment of the present application may also combine a confidence-based sampling strategy with a metric-based sampling strategy to implement a hybrid-based sampling strategy, dividing each sample object data into a plurality of first object data and a plurality of second object data.

[0103] Optionally, if the probability that any sample object data belongs to a cluster is not less than a probability threshold and the cluster to which any sample object data belongs is consistent with a reference cluster, any sample object data is used as the first object data. If the probability that any sample object data belongs to a cluster is less than a probability threshold and / or the cluster to which any sample object data belongs is inconsistent with the reference cluster, any sample object data is used as the second object data.

[0104] When the first object data is determined based on a mixed sampling strategy, the intersection of the first object data determined by the confidence-based sampling strategy and the first object data determined by the metric-based sampling strategy is obtained. By determining the first object data and the second object data based on a mixed sampling strategy, the division of the first object data and the second object data can be made more accurate. Since the accuracy of the first object data is improved, the accuracy of the labeled clustering results of the first object data is improved. When the clustering model is obtained by using the labeled clustering results of the first object data, since the accuracy of the clustering model depends on the accuracy of the labeled clustering results of the first object data, the accuracy of the clustering model can be improved, thereby improving the accuracy of the clustering results.

[0105] In the embodiment of the present application, the first object data may be referred to as high confidence data (Clean Data) or labeled data (Labeled Data), and the second object data may be referred to as low confidence data (Unclean Data) or unlabeled data (Unlabeled Data). When each sample object data is divided into a plurality of first object data and a plurality of second object data based on the cluster to which each sample object data belongs, the cluster to which the first object data belongs may be accurate or inaccurate, and similarly, the cluster to which the second object data belongs may also be accurate or inaccurate. Since the first object data is high confidence data, the embodiment of the present application determines the cluster to which the first object data belongs as the labeled clustering result of each first object data, and obtains a clustering model based on the training of the labeled clustering result of each first object data, so that the clustering model can learn the processing capability of any clustering algorithm for high confidence data, so as to improve the clustering accuracy of the clustering model.

[0106] Step 202: Acquire a first predicted clustering result of each first object data and a first predicted clustering result of each second object data according to the first network model.

[0107] In the embodiment of the present application, each first object data and each second object data are respectively input into the first network model, and the first network model determines and outputs the first predicted clustering result of each first object data and the first predicted clustering result of each second object data. The embodiment of the present application does not limit the structure, size, etc. of the first network model. For example, the first network model can be a multi-layer perceptron (MLP).

[0108] MLP contains three hidden layers. The dimension of the first hidden layer is 128, which means that this hidden layer can extract 128-dimensional object features. The dimension of the second hidden layer is 64, which means that this hidden layer can extract 64-dimensional object features. The dimension of the third hidden layer is n, which means that this hidden layer can output n clusters, where n is a positive integer. The activation function used by MLP is the Rectified Linear Unit (ReLU), which is also called the rectified linear unit and is a commonly used activation function in neural networks.

[0109] In a possible implementation, obtaining a first predicted clustering result for each first object data according to the first network model includes: for any first object data, obtaining each object feature of any first object data according to each feature extraction network in the first network model, any two feature extraction networks in the first network model are neural networks with different structures, and any object feature of any first object data is used to describe any first object data; fusing each object feature of any first object data to obtain a fused feature of any first object data; and determining a first predicted clustering result for each first object data based on the fused features of each first object data.

[0110] The first network model may include at least one feature extraction network, any one of which is used to extract object features of object data input into the first network model, so as to express the object data using the object features of the object data. Any two feature extraction networks of the first network model are neural networks with different structures, for example, the two feature extraction networks may be at least one of an MLP and a Graph Convolution Network (GCN), respectively.

[0111] In an embodiment of the present application, any first object data is input into the first network model, and each feature extraction network of the first network model performs feature extraction on the any first object data respectively to obtain each object feature of the any first object data. Then, each object feature of any first object data is fused to obtain a fused feature of any first object data, so as to use the fused feature of any first object data to describe any first object data, thereby improving accuracy. In this way, the fused features of each first object data can be determined. Next, the fused features of each first object data are clustered using the fully connected layer of the first network model to obtain the first predicted clustering results of each first object data. Since the accuracy of the fused features of the first object data is high, the accuracy of the first predicted clustering results of the first object data is also high, thereby improving the accuracy of the clustering model and further improving the accuracy of the clustering results.

[0112] See also Figure 4 , Figure 4 is a schematic diagram of a first network model provided in an embodiment of the present application, the first network model includes a multi-layer perceptron, a graph convolutional network and a fully connected layer. The first object data is input into the first network model, the multi-layer perceptron in the first network model performs feature extraction on the first object data to obtain object feature 1, and the graph convolutional network in the first network model performs feature extraction on the first object data to obtain object feature 2. Then, object feature 1 and object feature 2 are fused to obtain a fused feature.

[0113] In the embodiment of the present application, there are multiple first object data, and each first object data corresponds to a fusion feature. The fusion features of each first object data are input into the fully connected layer of the first network model, and the fully connected layer of the first network model performs clustering processing on each first object data to obtain a first predicted clustering result of each first object data.

[0114] It should be noted that the manner in which the first network model determines the first predicted clustering result of the second object data can be found in the description of the first network model determining the first predicted clustering result of the first object data in the above text, and will not be repeated here. That is to say, for any second object data, the various object features of any second object data can be obtained respectively according to the various feature extraction networks in the first network model, and any object feature of any second object data is used to describe any second object data. Then, the various object features of any second object data are fused to obtain the fused features of any second object data. In this way, the fused features of each second object data can be determined. Afterwards, based on the fused features of each second object data, the first predicted clustering result of each second object data is determined.

[0115] Step 203 : adjusting the first network model based on the labeled clustering results of each first object data, the first predicted clustering results, and the first predicted clustering results of each second object data to obtain a clustering model.

[0116] In an embodiment of the present application, the loss value of the first network model can be determined based on the labeled clustering results of each first object data, the first predicted clustering results, and the first predicted clustering results of each second object data, and the first network model can be adjusted based on the loss value of the first network model to obtain an adjusted first network model.

[0117] If the adjusted first network model meets the training end condition (such as reaching the training times 500), the adjusted first network model is used as the clustering model. If the adjusted first network model does not meet the training end condition, the adjusted first network model is further adjusted according to steps 201-203 until the training end condition is met to obtain the clustering model.

[0118] In a possible implementation, the first network model is adjusted based on the labeled clustering results of each first object data, the first predicted clustering results, and the first predicted clustering results of each second object data to obtain a clustering model, including: obtaining the second predicted clustering results of each first object data according to the second network model; adjusting the second network model based on the labeled clustering results of each first object data, the first predicted clustering results, the second predicted clustering results, and the first predicted clustering results of each second object data to obtain an adjusted second network model; adjusting the first network model based on the adjusted second network model to obtain a clustering model.

[0119] In the embodiment of the present application, each first object data can be directly input into the second network model, and the second network model performs clustering processing on each first object data to obtain a second predicted clustering result of each first object data. Among them, the embodiment of the present application does not limit the structure, size, etc. of the second network model. Exemplarily, the second network model can be MLP.

[0120] Optionally, obtaining the second predicted clustering result of each first object data according to the second network model includes: performing data enhancement processing on each first object data to obtain each data-enhanced first object data; inputting each data-enhanced first object data into the second network model, and outputting the second predicted clustering result of each first object data by the second network model.

[0121] In the embodiment of the present application, data enhancement processing can be performed on each first object data to obtain each data-enhanced first object data. The embodiment of the present application does not limit the method of data enhancement. Exemplarily, the method of data enhancement includes but is not limited to cropping data, adding noise (such as Gaussian noise), data interpolation, etc.

[0122] Afterwards, each data-enhanced first object data is input into the second network model, and the second network model performs clustering processing on each data-enhanced first object data to obtain the second predicted clustering result of each data-enhanced first object data. The second predicted clustering result of any data-enhanced first object data is the second predicted clustering result of any first object data.

[0123] By performing data enhancement processing on the first object data to introduce noise into the first object data, clustering processing is performed on each data-enhanced first object data using the second network model, which can improve the clustering ability of the second network model and improve the robustness and generalization ability of the model.

[0124] It should be noted that the second network model may include at least one feature extraction network, any one of which is used to extract object features of object data input into the second network model, so as to express the object data using the object features of the object data. Wherein, any two feature extraction networks of the second network model are neural networks with different structures, for example, the two feature extraction networks may be at least one of an MLP and a graph convolution network (GCN).

[0125] In an embodiment of the present application, for any first object data, each object feature of any first object data is obtained respectively according to each feature extraction network in the second network model, and any object feature of any first object data is used to describe any first object data. The various object features of any first object data are fused according to the second network model to obtain the fused features of any first object data. Based on the fused features of each first object data, the second predicted clustering result of each first object data is determined according to the second network model. Among them, any first object data can be the first object data after data enhancement. The method of determining the second predicted clustering result of each first object data according to the second network model can be seen in the above description of obtaining the first predicted clustering result of each first object data according to the first network model. The implementation principles of the two are similar and will not be repeated here.

[0126] After determining the second predicted clustering results of each first object data, the second network model is adjusted based on the labeled clustering results of each first object data, the first predicted clustering results, the second predicted clustering results and the first predicted clustering results of each second object data to obtain an adjusted second network model.

[0127] Optionally, based on the labeled clustering results of each first object data, the first predicted clustering results, the second predicted clustering results and the first predicted clustering results of each second object data, the second network model is adjusted to obtain an adjusted second network model, including: determining a first loss value based on the labeled clustering results and the second predicted clustering results of each first object data; determining a second loss value based on the first predicted clustering results of each first object data, the second predicted clustering results and the first predicted clustering results of each second object data; adjusting the second network model based on the first loss value and the second loss value to obtain an adjusted second network model.

[0128] In the embodiment of the present application, the first loss value is determined according to the first loss function, the labeled clustering results of each first object data, and the second predicted clustering results of each first object data. Among them, the first loss function is a supervised loss function. The embodiment of the present application does not limit the function type of the first loss function. Exemplarily, the first loss function is a cross entropy function. The first loss value can be expressed as in, is the function symbol of the first loss function, is the second predicted clustering result of each first object data, and y is the labeled clustering result of each first object data.

[0129] The second loss value is determined according to the second loss function, the first predicted clustering result of each first object data, the second predicted clustering result of each first object data, and the first predicted clustering result of each second object data. The second loss function is an unsupervised loss function. The embodiment of the present application does not limit the function type of the second loss function. Exemplarily, the second loss function is a mutual information function. The second loss value can be expressed as in, is the function symbol of the second loss function, is the second predicted clustering result of each first object data, p c is the first predicted clustering result of each first object data, p u It is a first predicted clustering result of each second object data.

[0130] Next, based on the first loss value, the second loss value, and the weight of the second loss value, the loss value of the second network model is determined. The loss value of the second network model can be expressed as in, is the loss value of the second network model, and α is the weight of the second loss value.

[0131] The embodiment of the present application does not limit the size of the weight of the second loss value. Exemplarily, α∈[0.2, 0.99]. Optionally, the weight of the second loss value is proportional to the number of training times. As the number of training times increases, the value of the second loss value gradually increases from 0.2 to 0.99.

[0132] Optionally, based on the labeled clustering results, first predicted clustering results, second predicted clustering results of each first object data and the first predicted clustering results of each second object data, the second network model is adjusted to obtain an adjusted second network model, including: obtaining the second predicted clustering results of each second object data according to the second network model; based on the labeled clustering results, first predicted clustering results, second predicted clustering results of each first object data and the first predicted clustering results, second predicted clustering results of each second object data, the second network model is adjusted to obtain an adjusted second network model.

[0133] In the embodiment of the present application, each second object data may be directly input into the second network model, and the second network model may perform clustering processing on each second object data to obtain a second predicted clustering result of each second object data.

[0134] Optionally, obtaining the second predicted clustering result of each second object data according to the second network model includes: performing data enhancement processing on each second object data to obtain each data-enhanced second object data; inputting each data-enhanced second object data into the second network model, and outputting the second predicted clustering result of each second object data by the second network model.

[0135] In the embodiment of the present application, data enhancement processing can be performed on each second object data to obtain each data-enhanced second object data. The embodiment of the present application does not limit the method of data enhancement. Exemplarily, the method of data enhancement includes but is not limited to cropping data, adding noise (such as Gaussian noise), data interpolation, etc.

[0136] Afterwards, each data-enhanced second object data is input into the second network model, and the second network model performs clustering processing on each data-enhanced second object data to obtain the second predicted clustering result of each data-enhanced second object data. The second predicted clustering result of any data-enhanced second object data is the second predicted clustering result of any second object data.

[0137] By performing data enhancement processing on the second object data to introduce noise into the second object data, clustering processing is performed on each data-enhanced second object data using the second network model, which can improve the clustering ability of the second network model and improve the robustness and generalization ability of the model.

[0138] In an embodiment of the present application, for any second object data, each object feature of any second object data is obtained respectively according to each feature extraction network in the second network model, and any object feature of any second object data is used to describe any second object data. The various object features of any second object data are fused according to the second network model to obtain the fused features of any second object data. Based on the fused features of each second object data, the second predicted clustering result of each second object data is determined according to the second network model. Among them, any second object data can be the second object data after data enhancement. The method of determining the second predicted clustering result of each second object data according to the second network model can be seen in the above description of obtaining the first predicted clustering result of each first object data according to the first network model. The implementation principles of the two are similar and will not be repeated here.

[0139] After determining the second predicted clustering result of each second object data, a first loss value is determined based on the labeled clustering result and the second predicted clustering result of each first object data, and a third loss value is determined based on the first predicted clustering result and the second predicted clustering result of each first object data and the first predicted clustering result and the second predicted clustering result of each second object data. Then, based on the first loss value and the third loss value, the second network model is adjusted to obtain an adjusted second network model.

[0140] In the embodiment of the present application, the first loss function, the labeled clustering result y of each first object data, the second predicted clustering result y of each first object data, Determine the first loss value in, is the function symbol of the first loss function.

[0141] According to the second loss function, the first predicted clustering result of each first object data, the second predicted clustering result of each first object data, the first predicted clustering result of each second object data, and the second predicted clustering result of each second object data, the third loss value is determined. Among them, the third loss function is an unsupervised loss function. The embodiment of the present application does not limit the function type of the third loss function. Exemplarily, the third loss function is a mutual information function. The third loss value can be expressed as in, is the function symbol of the third loss function, A second predicted clustering result for each first object data, is the second predicted clustering result of each second object data, p c is the first predicted clustering result of each first object data, p u It is a first predicted clustering result of each second object data.

[0142] Next, based on the first loss value, the third loss value, and the weight of the third loss value, the loss value of the second network model is determined. The loss value of the second network model can be expressed as in, is the loss value of the second network model, and β is the weight of the third loss value.

[0143] The embodiment of the present application does not limit the weight of the third loss value. Exemplarily, β∈[0.2, 0.99]. Optionally, the weight of the third loss value is proportional to the number of training times, and as the number of training times increases, the value of the third loss value gradually increases from 0.2 to 0.99.

[0144] After determining the loss value of the second network model, the second network model is adjusted using the loss value of the second network model to obtain an adjusted second network model. Then, the first network model is adjusted based on the adjusted second network model to obtain a clustering model.

[0145] In an embodiment of the present application, the prediction results of the first network model and the prediction results of the second network model are both used to determine the loss value of the unsupervised loss function, and the loss value is used for gradient back propagation to update the second network model, and the second network model is used to update the first network model.

[0146] It is understandable that the process of obtaining the clustering model is a process of multiple trainings. Each training (i.e., each Epoch) is carried out in accordance with steps 201 to 203 to determine the loss value of the second network model, adjust the second network model based on the loss value of the second network model, and adjust the first network model based on the adjusted second network model to obtain the adjusted first network model.

[0147] The model parameters of the first network model during the tth (t is a positive integer or 0) training can be recorded as θ′ t-1 , the model parameters of the second network model adjusted during the tth training are recorded as θ t , according to the formula θ′ t =αθ′ t-1 +(1-α)θ t Calculate the model parameters θ′ of the first network model adjusted during the tth training t , where α is the model parameter weight of the first network model, and at the beginning of training, i.e., t = 0, θ′ t =θ t The embodiment of the present application does not limit the value of α. For example, according to manual experience, α=0.99 is taken.

[0148] It is understandable that in order to improve the training speed, only the second network model can be adjusted in each of the multiple trainings, and the model parameters of the first network model can be adjusted using the model parameters of the adjusted second network model obtained in each of the multiple trainings. In this way, the model parameters of the first network model are adjusted once every multiple trainings, the update frequency of the first network model is reduced, and the training speed is improved.

[0149] It should be noted that the embodiment of the present application can use the exponential moving average (EMA) method to adjust the first network model based on the adjusted second network model, and the first network model is also called the EMA model. In the EMA method, the model structure of the first network model and the second network model is consistent, and the two only differ in model parameters. The EMA method needs to calculate the average value of the model parameters of the adjusted second network model, and use the average value of the model parameters of the adjusted second network model to adjust the first network model. Through the EMA method, the model performance of the second network model can be migrated to the first network model, thereby improving the robustness of the first network model.

[0150] The above method adjusts the first network model based on the labeled clustering results of each first object data, the first predicted clustering results and the first predicted clustering results of each second object data to obtain a clustering model. The labeled clustering results of a small amount of data can be used to guide the model to perform accurate clustering, thereby improving the accuracy of the clustering model and thus improving the accuracy of the clustering results.

[0151] Based on the above implementation environment, the present application embodiment provides a clustering method. Figure 5 Taking the flowchart of a clustering method provided in the embodiment of the present application as an example, the method can be performed by Figure 1 The electronic device 11 in the embodiment is executed. Figure 5 As shown, the method comprises the following steps:

[0152] Step 501, obtaining multiple target object data.

[0153] The embodiments of the present application do not limit the target object data. For example, the target object data is the gene expression data of the cell, or the target object data is the image data of the cell. The embodiments of the present application do not limit the method for obtaining any target object data. For example, the gene expression data of the cell is obtained by performing gene sequencing on the cell slice, or the image data of the cell is obtained by taking an image of the cell slice presented under a microscope.

[0154] Step 502: Determine the clustering result of each target object data according to the clustering model.

[0155] The clustering model is obtained by training according to the clustering model training method provided in the above-mentioned optional embodiments.

[0156] In the embodiment of the present application, each target object data is input into the clustering model, and the clustering model determines and outputs the clustering result of each target object data. The structure of the clustering model is the same as that of the first network model, which can be seen in the description of the first network model above, and will not be repeated here.

[0157] In a possible implementation, clustering results of each target object data are determined according to a clustering model, including: for any target object data, according to each feature extraction network in the clustering model, each object feature of any target object data is obtained respectively, any two feature extraction networks in the clustering model are neural networks with different structures, and any object feature of any target object data is used to describe any target object data; each object feature of any target object data is fused to obtain a fused feature of any target object data; and based on the fused features of each target object data, the clustering results of each target object data are determined.

[0158] The clustering model may include at least one feature extraction network, any one of which is used to extract object features of object data input into the clustering model, so as to express the object data using the object features of the object data. Any two feature extraction networks of the clustering model are neural networks of different structures, for example, the two feature extraction networks may be at least one of MLP and GCN, respectively.

[0159] In an embodiment of the present application, any target object data is input into a clustering model, and each feature extraction network of the clustering model performs feature extraction on the target object data respectively to obtain each object feature of the target object data. Then, each object feature of the target object data is fused to obtain a fused feature of the target object data, so as to use the fused feature of the target object data to describe any target object data, thereby improving accuracy. In this way, the fused feature of each target object data can be determined. Next, the fused feature of each target object data is clustered using the fully connected layer of the clustering model to obtain a clustering result of each target object data. Since the accuracy of the fused feature of the target object data is high, the accuracy of the clustering result of the target object data is also high.

[0160] In the embodiments of the present application, the clustering method provided in the embodiments of the present application can be used as an algorithm module, and the algorithm module can be embedded in any software, such as embedding the algorithm module in single-cell analysis clustering software.

[0161] The clustering model in the above method is obtained by adjusting the first network model based on the labeled clustering results of each first object data, the first predicted clustering results and the first predicted clustering results of each second object data. It can guide the model to perform accurate clustering through the labeled clustering results of a small amount of data, thereby improving the accuracy of the clustering model and thus improving the accuracy of the clustering results.

[0162] The above describes the training method and clustering method of the clustering model of the embodiment of the present application from the perspective of method steps, and the following is a detailed description in combination with the scenario. In the scenario of the embodiment of the present application, the object data is cell data, the first network model is an exponential moving average model, and the second network model is a multilayer perceptron.

[0163] See also Figure 6 , Figure 6 : is a training diagram of a clustering model provided in an embodiment of the present application. In an embodiment of the present application, for multiple sample cell data x, clustering processing is performed using a deep embedding single cell clustering algorithm to obtain cluster clusters to which each sample cell data x belongs, wherein the method of clustering processing based on a deep embedding single cell clustering algorithm has been described in detail above and will not be repeated here.

[0164] Afterwards, sampling processing is performed based on the cluster to which each sample cell data x belongs, to obtain a plurality of first cell data x c , a plurality of second cell data x u And each first cell data x c The labeled clustering result y. Using multiple first cell data x c , a plurality of second cell data x u And each first cell data x c The clustering model is obtained by obtaining the labeled clustering result y of the first object data. The sampling processing method can be seen in the above description of "obtaining a plurality of first object data, a plurality of second object data and a labeled clustering result of each first object data", which will not be repeated here.

[0165] In using multiple first cell data x c , a plurality of second cell data x u And each first cell data x c When obtaining the clustering model, on the one hand, multiple first cell data x are c , a plurality of second cell data x u Input the exponential moving average model, and the exponential moving average model outputs each first cell data x c The first predicted clustering result p c And each second cell data x u The first predicted clustering result p uOn the other hand, firstly, a plurality of first cell data x c Perform data enhancement to obtain the data of each first cell after data enhancement For multiple second cell data x u Perform data enhancement to obtain the data of each second cell after data enhancement Next, the data of each first cell after data enhancement is And the data of each second cell after data enhancement Input the multi-layer perceptron, and the multi-layer perceptron outputs the data of each first cell after data enhancement The second predicted clustering result And the data of each second cell after data enhancement The second predicted clustering result

[0166] Then, using each first cell data x c The first predicted clustering result p c , each second cell data x u The first predicted clustering result p u , each first cell data after data enhancement The second predicted clustering result Data of each second cell after data enhancement The second predicted clustering result Determine the loss value of the multilayer perceptron. Use the loss value of the multilayer perceptron to update the multilayer perceptron, and use the updated multilayer perceptron to update the exponential moving average model. The method of determining the loss value of the multilayer perceptron can be seen in the above description of determining the loss value of the second network model, and the method of updating the exponential moving average model using the updated multilayer perceptron can be seen in the above description of adjusting the first network model based on the adjusted second network model, which will not be repeated here.

[0167] In the above manner, the multilayer perceptron and the exponential moving average model are iteratively updated until the training end condition is met. The exponential moving average model when the training end condition is met is used as the clustering model. The clustering model is a clustering model trained by the clustering model training method provided in the embodiment of the present application, hereinafter referred to as the clustering model of the present method.

[0168] The present application example uses three cell data sets, which include about 60,000 cell data in total. The three cell data sets include a macaque cell data set, a monocyte data set, and a pancreatic islet cell data set. The relevant information of the three cell data sets is shown in Table 1 as shown below.

[0169] Table 1

[0170] Dataset Cell Type Cell number Clustering results Macaque cell dataset Macaque retinal bipolar cells 30302 12 Monocyte dataset Human peripheral blood mononuclear cells 24679 8 Pancreatic islet cell dataset Human pancreatic cancer islet cells 6321 13

[0171] As can be seen from Table 1, the macaque cell dataset includes data of 30,302 macaque retinal bipolar cells, and its actual clustering result is 12 clusters. The monocyte dataset includes data of 24,679 human peripheral blood mononuclear cells, and its actual clustering result is 8 clusters. The islet cell dataset includes data of 6,321 human pancreatic cancer islet cells, and its actual clustering result is 13 clusters.

[0172] After obtaining the macaque cell dataset, monocyte dataset and islet cell dataset, the DESC algorithm is used to cluster the macaque cell dataset, monocyte dataset and islet cell dataset respectively, and the clustering results corresponding to each dataset are obtained. At the same time, the clustering model of this method is also used to cluster the macaque cell dataset, monocyte dataset and islet cell dataset respectively, and the clustering results of each dataset are obtained.

[0173] See also Figure 7 , Figure 7 is a schematic diagram of a clustering result obtained by clustering a pancreatic islet cell dataset using a DESC algorithm provided in an embodiment of the present application. The schematic diagram of the clustering result includes clusters numbered 0 to 12, wherein the dotted ellipse includes two clusters numbered 0 and 4. Figure 8 , Figure 8 is a schematic diagram of a clustering result obtained by clustering a pancreatic islet cell dataset using a clustering model provided in an embodiment of the present application. The schematic diagram of the clustering result includes clusters numbered 1 to 3, 5 to 6, and 8 to 12, respectively. The dotted ellipse includes a cluster numbered 6. Please refer to Fig. 9 , Fig. 9 1 is a schematic diagram of an actual clustering result of a pancreatic islet cell data set provided in an embodiment of the present application. The schematic diagram of the clustering result includes clusters numbered 1 to 13, wherein the dotted ellipse includes a cluster numbered 2.

[0174] contrast Figures 7 to 9 It can be seen from the dotted ellipse that the clustering result corresponding to the DESC algorithm is inconsistent with the actual clustering result, while the clustering result corresponding to the clustering model is consistent with the actual clustering result. Therefore, the clustering model trained by the clustering model training method provided in the embodiment of the present application (i.e., the clustering model of the present method) is more accurate.

[0175] It should be noted that Figure 7 The numbers 0 to 12 are used only to distinguish Figure 7 The different clusters in cannot be used to represent the names of cells in the clusters. Figure 8The numbers 1 to 3, 5 to 6, 8 to 12 are used only to distinguish Figure 8 The different clusters in cannot be used to represent the names of cells in the clusters. Fig. 9 The numbers 1 to 13 are used only to distinguish Fig. 9 The different clusters in cannot be used to represent the names of cells in the clusters. Figure 7 , Figure 8 and Fig. 9 For any two schematic diagrams in the figure, the same label in the two schematic diagrams may point to clusters containing the same cell name or clusters containing different cell names.

[0176] The embodiment of the present application also uses two clustering evaluation indicators to evaluate the accuracy of the clustering results obtained by clustering different data sets using the DESC algorithm and the clustering model of the present method. The two clustering evaluation indicators are the Adjusted Random Index (ARI) and the Normalized Mutual Information (NMI), and the results are shown in Table 2 below.

[0177] Table 2

[0178]

[0179] As can be seen from Table 2, the ARI corresponding to this method is higher than the ARI corresponding to the DESC algorithm. Through further data processing, it can be determined that in terms of ARI, this method is about 13% higher than the DESC algorithm. The NMI corresponding to this method is generally higher than the NMI corresponding to the DESC algorithm. Through further data processing, it can be determined that in terms of NMI, this method is about 1% higher than the DESC algorithm. Since the larger the ARI, the more accurate the clustering result, and the larger the NMI, the more accurate the clustering result, therefore, the clustering result corresponding to this method is more accurate, that is, the clustering model of the application embodiment can more accurately cluster the cell data.

[0180] Fig.10 FIG. 1 is a schematic diagram of a clustering model training device provided in an embodiment of the present application. Fig.10 As shown, the device comprises:

[0181] An acquisition module 1001 is used to acquire a plurality of first object data, a plurality of second object data, and a labeled clustering result of each first object data;

[0182] The acquisition module 1001 is further used to acquire the first prediction clustering result of each first object data and the first prediction clustering result of each second object data according to the first network model;

[0183] The adjustment module 1002 is used to adjust the first network model based on the labeled clustering results of each first object data, the first predicted clustering results and the first predicted clustering results of each second object data to obtain a clustering model.

[0184] In a possible implementation, the acquisition module 1001 is used to obtain the cluster to which each sample object data belongs; based on the cluster to which each sample object data belongs, each sample object data is divided into multiple first object data and multiple second object data; and the cluster to which each first object data belongs is determined as the labeled clustering result of each first object data.

[0185] In a possible implementation, the acquisition module 1001 is used to determine, for any sample object data, a reference cluster of any sample object data from the clusters to which the various sample object data belong; in response to the cluster to which any sample object data belongs being consistent with the reference cluster, taking any sample object data as first object data; in response to the cluster to which any sample object data belongs being inconsistent with the reference cluster, taking any sample object data as second object data.

[0186] In a possible implementation, the acquisition module 1001 is used to determine the distance between any sample object data and each other object data, where the other object data is the sample object data other than any sample object data in the sample object data; based on the distance between any sample object data and each other object data, determine each other object data whose distance is not greater than a reference distance from each other object data; and from the clusters to which each other object data whose distance is not greater than the reference distance belongs, determine the cluster with the largest number of other object data whose distance is not greater than the reference distance as the reference cluster for any sample object data.

[0187] In a possible implementation, the acquisition module 1001 is configured to acquire the probability of each sample object data belonging to a cluster; and divide each sample object data into a plurality of first object data and a plurality of second object data based on the probability of each sample object data belonging to a cluster.

[0188] In one possible implementation, the acquisition module 1001 is used to determine, for any sample object data, cluster center data based on each sample object data in the cluster to which any sample object data belongs, where the cluster center data is used to characterize the center of the cluster to which any sample object data belongs; determine the distance between any sample object data and the cluster center data; and normalize the distance between any sample object data and the cluster center data to obtain the probability of the cluster to which any sample object data belongs.

[0189] In a possible implementation, the acquisition module 1001 is used to, for any sample object data, in response to the probability that any sample object data belongs to a cluster is not less than a probability threshold, use any sample object data as first object data; in response to the probability that any sample object data belongs to a cluster is less than the probability threshold, use any sample object data as second object data.

[0190] In one possible implementation, the adjustment module 1002 is used to obtain the second predicted clustering result of each first object data according to the second network model; adjust the second network model based on the labeled clustering result of each first object data, the first predicted clustering result, the second predicted clustering result and the first predicted clustering result of each second object data to obtain an adjusted second network model; adjust the first network model based on the adjusted second network model to obtain a clustering model.

[0191] In a possible implementation, the adjustment module 1002 is used to perform data enhancement processing on each first object data to obtain each data-enhanced first object data; input each data-enhanced first object data into the second network model, and the second network model outputs the second predicted clustering result of each first object data.

[0192] In one possible implementation, the adjustment module 1002 is used to determine a first loss value based on the labeled clustering results of each first object data and the second predicted clustering results; determine a second loss value based on the first predicted clustering results of each first object data, the second predicted clustering results and the first predicted clustering results of each second object data; and adjust the second network model based on the first loss value and the second loss value to obtain an adjusted second network model.

[0193] In one possible implementation, the adjustment module 1002 is used to obtain a second predicted clustering result of each second object data according to the second network model; based on the labeled clustering results, the first predicted clustering results, the second predicted clustering results of each first object data and the first predicted clustering results, the second predicted clustering results of each second object data, the second network model is adjusted to obtain an adjusted second network model.

[0194] In one possible implementation, the adjustment module 1002 is used to perform data enhancement processing on each second object data to obtain each data-enhanced second object data; input each data-enhanced second object data into the second network model, and the second network model outputs the second predicted clustering result of each second object data.

[0195] In one possible implementation, the acquisition module 1001 is used to obtain, for any first object data, respective object features of any first object data according to respective feature extraction networks in the first network model, any two feature extraction networks in the first network model are neural networks with different structures, and any object feature of any first object data is used to describe any first object data; the respective object features of any first object data are fused to obtain fused features of any first object data; and the first predicted clustering results of each first object data are determined based on the fused features of each first object data.

[0196] The above device adjusts the first network model based on the labeled clustering results of each first object data, the first predicted clustering results and the first predicted clustering results of each second object data to obtain a clustering model. The labeled clustering results of a small amount of data can be used to guide the model to perform accurate clustering, thereby improving the accuracy of the clustering model and thus improving the accuracy of the clustering results.

[0197] Fig.11 FIG. 1 is a schematic diagram of a clustering device provided in an embodiment of the present application. Fig.11 As shown, the device comprises:

[0198] An acquisition module 1101 is used to acquire multiple target object data;

[0199] The determination module 1102 is used to determine the clustering result of each target object data according to the clustering model, where the clustering model is trained according to any of the above clustering model training methods.

[0200] In one possible implementation, the determination module 1102 is used to obtain, for any target object data, each object feature of any target object data according to each feature extraction network in the clustering model, wherein any two feature extraction networks in the clustering model are neural networks with different structures, and any object feature of any target object data is used to describe any target object data; the various object features of any target object data are fused to obtain a fused feature of any target object data; and the clustering results of each target object data are determined based on the fused features of each target object data.

[0201] The clustering model in the above device is obtained by adjusting the first network model based on the labeled clustering results of each first object data, the first predicted clustering results and the first predicted clustering results of each second object data. The labeled clustering results of a small amount of data can be used to guide the model to perform accurate clustering, thereby improving the accuracy of the clustering model and thus improving the accuracy of the clustering results.

[0202] It should be understood that the above Fig.10 and Fig.11 When the device provided realizes its functions, only the division of the above-mentioned functional modules is used as an example for illustration. In practical applications, the above-mentioned functions can be assigned to different functional modules as needed, that is, the internal structure of the device is divided into different functional modules to complete all or part of the functions described above. In addition, the device and method embodiments provided in the above embodiments belong to the same concept, and their specific implementation process is detailed in the method embodiment, which will not be repeated here.

[0203] Fig.12 The structural block diagram of a terminal device 1200 provided by an exemplary embodiment of the present application is shown. The terminal device 1200 may be a portable mobile terminal, such as a smart phone, a tablet computer, an MP3 (Moving Picture Experts Group Audio Layer III, dynamic image expert compression standard audio layer 3) player, an MP4 (Moving Picture Experts Group Audio Layer IV, dynamic image expert compression standard audio layer 4) player, a notebook computer or a desktop computer. The terminal device 1200 may also be called a user device, a portable terminal, a laptop terminal, a desktop terminal, or other names.

[0204] Typically, the terminal device 1200 includes: a processor 1201 and a memory 1202 .

[0205] The processor 1201 may include one or more processing cores, such as a 4-core processor, an 8-core processor, etc. The processor 1201 may be implemented in at least one hardware form of DSP (Digital Signal Processing), FPGA (Field-Programmable Gate Array), and PLA (Programmable Logic Array). The processor 1201 may also include a main processor and a coprocessor. The main processor is a processor for processing data in an awake state, also known as a CPU (Central Processing Unit); the coprocessor is a low-power processor for processing data in a standby state. In some embodiments, the processor 1201 may be integrated with a GPU (Graphics Processing Unit), which is responsible for rendering and drawing the content to be displayed on the display screen. In some embodiments, the processor 1201 may also include an AI (Artificial Intelligence) processor, which is used to process computing operations related to machine learning.

[0206] The memory 1202 may include one or more computer-readable storage media, which may be non-transitory. The memory 1202 may also include a high-speed random access memory, and a non-volatile memory, such as one or more disk storage devices, flash memory storage devices. In some embodiments, the non-transitory computer-readable storage medium in the memory 1202 is used to store at least one instruction, which is used to be executed by the processor 1201 to implement the training method or clustering method of the clustering model provided in the method embodiment of the present application.

[0207] In some embodiments, the terminal device 1200 may further optionally include: a peripheral device interface 1203 and at least one peripheral device. The processor 1201, the memory 1202 and the peripheral device interface 1203 may be connected via a bus or a signal line. Each peripheral device may be connected to the peripheral device interface 1203 via a bus, a signal line or a circuit board. Specifically, the peripheral device includes: at least one of a radio frequency circuit 1204, a display screen 1205, a camera assembly 1206, an audio circuit 1207 and a power supply 1208.

[0208] The peripheral device interface 1203 may be used to connect at least one peripheral device related to I / O (Input / Output) to the processor 1201 and the memory 1202. In some embodiments, the processor 1201, the memory 1202, and the peripheral device interface 1203 are integrated on the same chip or circuit board; in some other embodiments, any one or two of the processor 1201, the memory 1202, and the peripheral device interface 1203 may be implemented on a separate chip or circuit board, which is not limited in this embodiment.

[0209] The radio frequency circuit 1204 is used to receive and transmit RF (Radio Frequency) signals, also known as electromagnetic signals. The radio frequency circuit 1204 communicates with the communication network and other communication devices through electromagnetic signals. The radio frequency circuit 1204 converts electrical signals into electromagnetic signals for transmission, or converts received electromagnetic signals into electrical signals. Optionally, the radio frequency circuit 1204 includes: an antenna system, an RF transceiver, one or more amplifiers, a tuner, an oscillator, a digital signal processor, a codec chipset, a user identity module card, and the like. The radio frequency circuit 1204 can communicate with other terminals through at least one wireless communication protocol. The wireless communication protocol includes, but is not limited to: the World Wide Web, a metropolitan area network, an intranet, various generations of mobile communication networks (2G, 3G, 4G and 5G), a wireless local area network and / or a WiFi (Wireless Fidelity) network. In some embodiments, the radio frequency circuit 1204 may also include circuits related to NFC (Near Field Communication), which is not limited in this application.

[0210] The display screen 1205 is used to display a UI (User Interface). The UI may include graphics, text, icons, videos, and any combination thereof. When the display screen 1205 is a touch display screen, the display screen 1205 also has the ability to collect touch signals on the surface or above the surface of the display screen 1205. The touch signal can be input to the processor 1201 as a control signal for processing. At this time, the display screen 1205 can also be used to provide virtual buttons and / or virtual keyboards, also known as soft buttons and / or soft keyboards. In some embodiments, the display screen 1205 can be one, set on the front panel of the terminal device 1200; in other embodiments, the display screen 1205 can be at least two, respectively set on different surfaces of the terminal device 1200 or in a folding design; in other embodiments, the display screen 1205 can be a flexible display screen, set on a curved surface or a folding surface of the terminal device 1200. Even, the display screen 1205 can also be set to a non-rectangular irregular shape, that is, a special-shaped screen. The display screen 1205 can be made of materials such as LCD (Liquid Crystal Display) and OLED (Organic Light-Emitting Diode).

[0211] The camera assembly 1206 is used to capture images or videos. Optionally, the camera assembly 1206 includes a front camera and a rear camera. Typically, the front camera is arranged on the front panel of the terminal, and the rear camera is arranged on the back of the terminal. In some embodiments, there are at least two rear cameras, which are any one of a main camera, a depth of field camera, a wide-angle camera, and a telephoto camera, so as to realize the fusion of the main camera and the depth of field camera to realize the background blur function, the fusion of the main camera and the wide-angle camera to realize the panoramic shooting and VR (Virtual Reality) shooting function or other fusion shooting functions. In some embodiments, the camera assembly 1206 may also include a flash. The flash can be a monochrome temperature flash or a dual-color temperature flash. A dual-color temperature flash refers to a combination of a warm light flash and a cold light flash, which can be used for light compensation at different color temperatures.

[0212] The audio circuit 1207 may include a microphone and a speaker. The microphone is used to collect sound waves from the user and the environment, and convert the sound waves into electrical signals and input them into the processor 1201 for processing, or input them into the radio frequency circuit 1204 to achieve voice communication. For the purpose of stereo acquisition or noise reduction, there may be multiple microphones, which are respectively arranged at different parts of the terminal device 1200. The microphone may also be an array microphone or an omnidirectional acquisition microphone. The speaker is used to convert the electrical signal from the processor 1201 or the radio frequency circuit 1204 into sound waves. The speaker may be a traditional film speaker or a piezoelectric ceramic speaker. When the speaker is a piezoelectric ceramic speaker, it can not only convert the electrical signal into sound waves audible to humans, but also convert the electrical signal into sound waves inaudible to humans for purposes such as ranging. In some embodiments, the audio circuit 1207 may also include a headphone jack.

[0213] The power supply 1208 is used to power various components in the terminal device 1200. The power supply 1208 can be an alternating current, a direct current, a disposable battery, or a rechargeable battery. When the power supply 1208 includes a rechargeable battery, the rechargeable battery can be a wired rechargeable battery or a wireless rechargeable battery. A wired rechargeable battery is a battery charged through a wired line, and a wireless rechargeable battery is a battery charged through a wireless coil. The rechargeable battery can also be used to support fast charging technology.

[0214] In some embodiments, the terminal device 1200 further includes one or more sensors 1210 , including but not limited to: an acceleration sensor 1211 , a gyroscope sensor 1212 , a pressure sensor 1213 , an optical sensor 1214 , and a proximity sensor 1215 .

[0215] The acceleration sensor 1211 can detect the magnitude of acceleration on the three coordinate axes of the coordinate system established by the terminal device 1200. For example, the acceleration sensor 1211 can be used to detect the components of gravity acceleration on the three coordinate axes. The processor 1201 can control the display screen 1205 to display the user interface in a horizontal view or a vertical view according to the gravity acceleration signal collected by the acceleration sensor 1211. The acceleration sensor 1211 can also be used to collect game or user motion data.

[0216] The gyro sensor 1212 can detect the body direction and rotation angle of the terminal device 1200, and the gyro sensor 1212 can cooperate with the acceleration sensor 1211 to collect the user's 3D actions on the terminal device 1200. The processor 1201 can implement the following functions based on the data collected by the gyro sensor 1212: motion sensing (such as changing the UI according to the user's tilt operation), image stabilization during shooting, game control, and inertial navigation.

[0217] The pressure sensor 1213 can be set on the side frame of the terminal device 1200 and / or the lower layer of the display screen 1205. When the pressure sensor 1213 is set on the side frame of the terminal device 1200, it can detect the user's holding signal of the terminal device 1200, and the processor 1201 performs left and right hand recognition or shortcut operation according to the holding signal collected by the pressure sensor 1213. When the pressure sensor 1213 is set on the lower layer of the display screen 1205, the processor 1201 controls the operability controls on the UI interface according to the user's pressure operation on the display screen 1205. The operability controls include at least one of a button control, a scroll bar control, an icon control, and a menu control.

[0218] The optical sensor 1214 is used to collect the ambient light intensity. In one embodiment, the processor 1201 can control the display brightness of the display screen 1205 according to the ambient light intensity collected by the optical sensor 1214. Specifically, when the ambient light intensity is high, the display brightness of the display screen 1205 is increased; when the ambient light intensity is low, the display brightness of the display screen 1205 is reduced. In another embodiment, the processor 1201 can also dynamically adjust the shooting parameters of the camera assembly 1206 according to the ambient light intensity collected by the optical sensor 1214.

[0219] The proximity sensor 1215, also called a distance sensor, is usually arranged on the front panel of the terminal device 1200. The proximity sensor 1215 is used to collect the distance between the user and the front of the terminal device 1200. In one embodiment, when the proximity sensor 1215 detects that the distance between the user and the front of the terminal device 1200 is gradually decreasing, the processor 1201 controls the display screen 1205 to switch from the screen-on state to the screen-off state; when the proximity sensor 1215 detects that the distance between the user and the front of the terminal device 1200 is gradually increasing, the processor 1201 controls the display screen 1205 to switch from the screen-off state to the screen-on state.

[0220] Those skilled in the art will understand that Fig.12 The structure shown in the figure does not constitute a limitation on the terminal device 1200, and may include more or fewer components than shown in the figure, or combine certain components, or adopt a different component arrangement.

[0221] Fig.13 This is a schematic diagram of the structure of the server provided in the embodiment of the present application. The server 1300 may have relatively large differences due to different configurations or performances, and may include one or more processors 1301 and one or more memories 1302, wherein the one or more memories 1302 store at least one program code, and the at least one program code is loaded and executed by the one or more processors 1301 to implement the training method or clustering method of the clustering model provided in the above-mentioned various method embodiments. Exemplarily, the processor 1301 is a CPU. Of course, the server 1300 may also have components such as a wired or wireless network interface, a keyboard, and an input and output interface for input and output. The server 1300 may also include other components for implementing device functions, which will not be described in detail here.

[0222] In an exemplary embodiment, a computer-readable storage medium is also provided, in which at least one program code is stored. The at least one program code is loaded and executed by a processor to enable an electronic device to implement any of the above-mentioned clustering model training methods or clustering methods.

[0223] Optionally, the computer-readable storage medium may be a read-only memory (ROM), a random access memory (RAM), a compact disc (CD-ROM), a magnetic tape, a floppy disk, an optical data storage device, etc.

[0224] In an exemplary embodiment, a computer program or a computer program product is also provided, in which at least one computer instruction is stored, and the at least one computer instruction is loaded and executed by a processor to enable a computer to implement any of the above-mentioned clustering model training methods or clustering methods.

[0225] It should be understood that the "plurality" mentioned in this article refers to two or more. "And / or" describes the association relationship of the associated objects, indicating that there can be three relationships. For example, A and / or B can mean: A exists alone, A and B exist at the same time, and B exists alone. The character " / " generally indicates that the associated objects are in an "or" relationship.

[0226] The serial numbers of the above-mentioned embodiments of the present application are for description only and do not represent the advantages or disadvantages of the embodiments.

[0227] The above description is only an exemplary embodiment of the present application and is not intended to limit the present application. Any modifications, equivalent substitutions, improvements, etc. made within the principles of the present application shall be included in the protection scope of the present application.

Claims

1. A clustering model training method, characterized in that: The method comprises: Performing clustering processing on each sample object data to obtain a cluster to which each sample object data belongs, wherein the sample object data is gene expression data of cells or image data of cells; Based on the clusters to which the respective sample object data belong, the respective sample object data are divided into a plurality of first object data and a plurality of second object data, wherein the confidence level of the clusters to which the first object data belongs is higher than the confidence level of the clusters to which the second object data belongs; Determine the cluster to which each first object data belongs as a labeled clustering result of the each first object data; Acquire, according to the first network model, a first prediction clustering result of each first object data and a first prediction clustering result of each second object data; The first network model is adjusted based on the labeled clustering results of each first object data, the first predicted clustering results, and the first predicted clustering results of each second object data to obtain a clustering model.

2. The method according to claim 1, characterized in that The dividing each sample object data into a plurality of first object data and a plurality of second object data based on the cluster to which each sample object data belongs comprises: For any sample object data, determining a reference cluster of the sample object data from the clusters to which the respective sample object data belong; In response to the cluster to which any one of the sample object data belongs being consistent with the reference cluster, taking any one of the sample object data as the first object data; In response to the cluster to which any one of the sample object data belongs being inconsistent with the reference cluster, the any one of the sample object data is used as the second object data.

3. The method according to claim 2, characterized in that The step of determining a reference cluster of any one of the sample object data from the clusters to which the various sample object data belong comprises: determining a distance between any one of the sample object data and each other object data, wherein the other object data is sample object data other than the any one of the sample object data among the each of the sample object data; Based on the distance between any one of the sample object data and the respective other object data, determining respective other object data having a distance not greater than a reference distance from the respective other object data; From the clusters to which the other object data within the reference distance belong, a cluster containing the largest number of other object data within the reference distance is determined as a reference cluster for any one of the sample object data.

4. The method according to claim 1, characterized in that: The dividing each sample object data into a plurality of first object data and a plurality of second object data based on the cluster to which each sample object data belongs comprises: Obtaining the probability of each sample object data belonging to a cluster; Based on the probability of the cluster to which the respective sample object data belongs, the respective sample object data are divided into the plurality of first object data and the plurality of second object data.

5. The method according to claim 4, characterized in that The obtaining the probability of the cluster to which each sample object data belongs includes: For any sample object data, based on each sample object data in the cluster to which the any sample object data belongs, determine cluster center data, wherein the cluster center data is used to represent the center of the cluster to which the any sample object data belongs; Determine the distance between any one of the sample object data and the cluster center data; The distance between any one of the sample object data and the cluster center data is normalized to obtain the probability of the cluster to which any one of the sample object data belongs.

6. The method according to claim 4, characterized in that The dividing each sample object data into the plurality of first object data and the plurality of second object data based on the probability of the cluster to which each sample object data belongs comprises: For any sample object data, in response to the probability that the cluster to which the any sample object data belongs is not less than a probability threshold, taking the any sample object data as the first object data; In response to the probability that the cluster to which any one of the sample object data belongs is less than the probability threshold, the any one of the sample object data is used as the second object data.

7. The method according to any one of claims 1 to 6, characterized in that: The adjusting the first network model based on the labeled clustering results of each first object data, the first predicted clustering results, and the first predicted clustering results of each second object data to obtain a clustering model includes: Acquire a second predicted clustering result of each first object data according to the second network model; Based on the labeled clustering results of each first object data, the first predicted clustering results, the second predicted clustering results, and the first predicted clustering results of each second object data, adjusting the second network model to obtain an adjusted second network model; The first network model is adjusted based on the adjusted second network model to obtain a clustering model.

8. The method according to claim 7, characterized in that The obtaining of the second prediction clustering result of each first object data according to the second network model includes: Performing data enhancement processing on each of the first object data to obtain each of the first object data after data enhancement; The first object data after the data enhancement are input into the second network model, and the second network model outputs the second predicted clustering result of the first object data.

9. The method according to claim 7, characterized in that: The adjusting the second network model based on the labeled clustering results of each first object data, the first predicted clustering results, the second predicted clustering results and the first predicted clustering results of each second object data to obtain the adjusted second network model includes: Determine a first loss value based on the labeled clustering results of each first object data and the second predicted clustering results; Determine a second loss value based on the first predicted clustering result of each first object data, the second predicted clustering result, and the first predicted clustering result of each second object data; Based on the first loss value and the second loss value, the second network model is adjusted to obtain an adjusted second network model.

10. The method according to claim 7, characterized in that The adjusting the second network model based on the labeled clustering results of each first object data, the first predicted clustering results, the second predicted clustering results and the first predicted clustering results of each second object data to obtain the adjusted second network model includes: Acquire a second predicted clustering result of each second object data according to the second network model; Based on the labeled clustering results, the first predicted clustering results, the second predicted clustering results of each first object data and the first predicted clustering results, the second predicted clustering results of each second object data, the second network model is adjusted to obtain an adjusted second network model.

11. The method according to claim 10, characterized in that The obtaining a second prediction clustering result of each second object data according to the second network model includes: Performing data enhancement processing on each of the second object data to obtain each of the second object data after data enhancement; The second object data after the data enhancement are input into the second network model, and the second network model outputs the second predicted clustering result of the second object data.

12. The method according to any one of claims 1 to 6, characterized in that: The obtaining, according to the first network model, a first prediction clustering result of each first object data includes: For any first object data, obtaining respective object features of the any first object data according to respective feature extraction networks in the first network model, wherein any two feature extraction networks in the first network model are neural networks with different structures, and any object feature of the any first object data is used to describe the any first object data; fusing the object features of any one of the first object data to obtain a fused feature of any one of the first object data; Based on the fusion features of the respective first object data, a first predicted clustering result of the respective first object data is determined.

13. A clustering method, characterized in that: The method comprises: Get multiple target object data; A clustering result of each target object data is determined according to a clustering model, wherein the clustering model is trained according to the clustering model training method according to any one of claims 1 to 12.

14. The method according to claim 13, characterized in that Determining the clustering results of each target object data according to the clustering model includes: For any target object data, according to each feature extraction network in the clustering model, each object feature of the target object data is respectively obtained, any two feature extraction networks in the clustering model are neural networks with different structures, and any object feature of the target object data is used to describe the target object data; fusing the object features of any target object data to obtain a fused feature of any target object data; Based on the fusion features of the respective target object data, clustering results of the respective target object data are determined.

15. A clustering model training device, characterized in that: The device comprises: an acquisition module, configured to perform clustering processing on each sample object data to obtain a cluster to which each sample object data belongs, wherein the sample object data is gene expression data of a cell or image data of a cell; based on the cluster to which each sample object data belongs, divide each sample object data into a plurality of first object data and a plurality of second object data, wherein the confidence of the cluster to which the first object data belongs is higher than the confidence of the cluster to which the second object data belongs; and determine the cluster to which each first object data belongs as a labeled clustering result of each first object data; The acquisition module is further used to acquire the first prediction clustering result of each first object data and the first prediction clustering result of each second object data according to the first network model; The adjustment module is used to adjust the first network model based on the labeled clustering results of each first object data, the first predicted clustering results and the first predicted clustering results of each second object data to obtain a clustering model.

16. The device according to claim 15, characterized in that The acquisition module is used to determine, for any sample object data, a reference cluster of the sample object data from the clusters to which the various sample object data belong; in response to the cluster to which the sample object data belongs being consistent with the reference cluster, use the sample object data as the first object data; In response to the cluster to which any one of the sample object data belongs being inconsistent with the reference cluster, the any one of the sample object data is used as the second object data.

17. The device according to claim 16, characterized in that The acquisition module is used to determine the distance between any one of the sample object data and each other object data, where the other object data is the sample object data other than the any one of the sample object data; based on the distance between any one of the sample object data and the each other object data, determine each other object data whose distance is not greater than a reference distance from the each other object data; From the clusters to which the other object data within the reference distance belong, a cluster containing the largest number of other object data within the reference distance is determined as a reference cluster for any one of the sample object data.

18. The device according to claim 15, characterized in that The acquisition module is used to acquire the probability of each sample object data belonging to a cluster; based on the probability of each sample object data belonging to a cluster, divide each sample object data into the plurality of first object data and the plurality of second object data.

19. The device according to claim 18, characterized in that The acquisition module is used to determine, for any sample object data, cluster center data based on each sample object data in the cluster to which the any sample object data belongs, wherein the cluster center data is used to represent the center of the cluster to which the any sample object data belongs; and determine the distance between the any sample object data and the cluster center data; The distance between any one of the sample object data and the cluster center data is normalized to obtain the probability of the cluster to which any one of the sample object data belongs.

20. The device according to claim 18, characterized in that The acquisition module is used for, for any sample object data, in response to the probability that the cluster to which any sample object data belongs is not less than a probability threshold, taking the any sample object data as the first object data; in response to the probability that the cluster to which any sample object data belongs is less than the probability threshold, taking the any sample object data as the second object data.

21. The device according to any one of claims 15 to 20, characterized in that The adjustment module is used to obtain the second predicted clustering result of each first object data according to the second network model; adjust the second network model based on the labeled clustering results of each first object data, the first predicted clustering result, the second predicted clustering result and the first predicted clustering result of each second object data to obtain an adjusted second network model; adjust the first network model based on the adjusted second network model to obtain a clustering model.

22. The device according to claim 21, characterized in that The adjustment module is used to perform data enhancement processing on each of the first object data to obtain each of the data-enhanced first object data; input the each of the data-enhanced first object data into the second network model, and the second network model outputs the second predicted clustering result of each of the first object data.

23. The device according to claim 21, characterized in that The adjustment module is used to determine a first loss value based on the labeled clustering results of the respective first object data and the second predicted clustering results; determine a second loss value based on the first predicted clustering results of the respective first object data, the second predicted clustering results and the first predicted clustering results of the respective second object data; and adjust the second network model based on the first loss value and the second loss value to obtain an adjusted second network model.

24. The device according to claim 21, characterized in that The adjustment module is used to obtain the second predicted clustering result of each second object data according to the second network model; adjust the second network model based on the labeled clustering results, the first predicted clustering results, the second predicted clustering results of each first object data and the first predicted clustering results, the second predicted clustering results of each second object data to obtain an adjusted second network model.

25. The device according to claim 24, characterized in that The adjustment module is used to perform data enhancement processing on each of the second object data to obtain each of the second object data after data enhancement; input the each of the second object data after data enhancement into the second network model, and the second network model outputs the second predicted clustering result of each of the second object data.

26. The device according to any one of claims 15 to 20, characterized in that The acquisition module is used to acquire, for any first object data, respective object features of the first object data according to respective feature extraction networks in the first network model, wherein any two feature extraction networks in the first network model are neural networks of different structures, and any object feature of the first object data is used to describe the first object data; and to fuse the respective object features of the first object data to obtain a fused feature of the first object data; Based on the fusion features of the respective first object data, a first predicted clustering result of the respective first object data is determined.

27. A clustering device, characterized in that: The device comprises: An acquisition module, used to acquire data of multiple target objects; A determination module is used to determine the clustering result of each target object data according to a clustering model, wherein the clustering model is trained according to the clustering model training method according to any one of claims 1 to 12.

28. The device according to claim 27, characterized in that The determination module is used to obtain, for any target object data, respective object features of the target object data according to respective feature extraction networks in the clustering model, any two feature extraction networks in the clustering model are neural networks with different structures, and any object feature of the target object data is used to describe the target object data; and to fuse the respective object features of the target object data to obtain a fused feature of the target object data; Based on the fusion features of the respective target object data, clustering results of the respective target object data are determined.

29. An electronic device, characterized in that: The electronic device includes a processor and a memory, wherein the memory stores at least one program code, and the at least one program code is loaded and executed by the processor so that the electronic device implements the training method of the clustering model as described in any one of claims 1 to 12 or the clustering method as described in claim 13 or 14.

30. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores at least one program code, and the at least one program code is loaded and executed by the processor so that the computer implements the training method of the clustering model as described in any one of claims 1 to 12 or the clustering method as described in claim 13 or 14.

31. A computer program product, characterized in that The computer program product stores at least one computer instruction, and the at least one computer instruction is loaded and executed by a processor so that a computer implements the training method of a clustering model as described in any one of claims 1 to 12 or the clustering method as described in claim 13 or 14.