Bioinformatics model training method and device based on bioinformatics data

Abstract

The present application discloses a method and device for training a bioinformatics model based on bioinformatics data, and generally relates to the field of computer technology. The method includes: training an initial bioinformatics model through a first data set of a first user in a user set to generate a first bioinformatics model, wherein the first data set is the local data of the first user; screening a second user in the user set, wherein the first user and the second user are different users in the user set; training the first bioinformatics model through a second data set of the second user to generate a second bioinformatics model, and obtaining a convergence value of the second bioinformatics model, wherein the second data set is the local data of the second user; when the convergence value of the second bioinformatics model is less than a first preset threshold, confirming that the second bioinformatics model is a target bioinformatics model, wherein the target bioinformatics model is a trained model. The training efficiency can be greatly improved, and the communication resources required for the training process can be reduced.

Description

Technical Field

[0001] The present disclosure generally relates to the field of computer technology, and more particularly to a method and device for training a bioinformatics model based on bioinformatics data. Background Art

[0002] With the continuous advancement and development of Internet technology, various fields (such as bioinformatics, finance, etc.) have achieved significant improvements in efficiency through Internet technology, and the efficiency of data analysis has also been greatly improved accordingly.

[0003] Data analysis technology can often be analyzed through network-based learning models, and different fields use different learning models. Taking the field of bioinformatics as an example, the bioinformatics learning model is obtained through training with a large amount of bioinformatics data. After training with a large amount of bioinformatics data, the converged bioinformatics learning model can be used for data analysis in practical applications. In related technologies, the training method of bioinformatics learning models is often to send them to multiple users with original data through a central server. After multiple users train with their own original data, they send them to the central server, and finally generate a usable bioinformatics learning model.

[0004] However, in this process, since the above-mentioned bioinformatics learning model training process has a high demand for network communication, when communication resources are limited, the training process will generate a training bottleneck, and it is impossible to obtain the best training effect and a bioinformatics learning model with strong data parsing capabilities. Summary of the invention

[0005] In view of the above-mentioned defects or deficiencies in the related art, it is desired to provide a bioinformatics model training method and device based on bioinformatics data, which can solve the problem in the related art that due to the high network communication requirements of the above-mentioned learning model training process, the training process will generate a training bottleneck when communication resources are relatively limited, and it is impossible to obtain the best training effect and a learning model with strong data parsing capabilities. The method can greatly improve the training efficiency and reduce the communication resources required for the training process.

[0006] In a first aspect, a method for training a bioinformatics model based on bioinformatics data is provided, the method comprising:

[0007] Training the initial biometric information model by using a first data set of a first user in the user set to generate a first biometric information model, wherein the first data set is local data of the first user, and the local data of the first user includes first biometric information data;

[0008] screening a second user in the user set, wherein the first user and the second user are different users in the user set;

[0009] Training the first biological information model through a second data set of the second user to generate a second biological information model, and obtaining a convergence value of the second biological information model, wherein the second data set is local data of the second user, and the local data of the second user includes second biological information data;

[0010] When the convergence value of the second biological information model is less than the first preset threshold, the second biological information model is confirmed to be a target biological information model, and the target biological information model is a trained model.

[0011] In the present application, the initial bio-information model is trained by the local data of the first user in the user set, that is, the first data set, to generate the first bio-information model; then, a second user different from the first user in the user set is screened, and the first bio-information model is trained by the local data of the second user, that is, the second data set, to produce the second bio-information model, and the convergence value of the second bio-information model is obtained; finally, when the convergence value of the second bio-information model is less than the first preset threshold value, the second bio-information model is confirmed to be the target bio-information model that has been trained. In this way, in the process of training the model of the present application, the model is trained in sequence for the screened users. On the one hand, it avoids the difficulty in collecting the initial bioinformatics model below uniformly after waiting for the training of each node user to be completed, and the communication bandwidth is insufficient in the aggregation process to collect and aggregate the model. On the other hand, it also avoids the problem of low efficiency caused by waiting for all node users to be trained before feeding back the model in the unified collection process. In the method of the present application, the model is trained by randomly screening node users and training different node users in sequence. After training to any node user, if the training model of the node user can converge, the model of the node user can be determined as the target bioinformatics model, thereby saving the demand for communication bandwidth in the process of sending the model and greatly improving the training efficiency.

[0012] In a second aspect, a bioinformatics model training device based on bioinformatics data is provided, the device comprising:

[0013] A training module, configured to train an initial biometric information model using a first data set of a first user in a user set to generate a first biometric information model, wherein the first data set is local data of the first user, and the local data of the first user includes first biometric information data;

[0014] A screening module, used for screening a second user in the user set, wherein the first user and the second user are different users in the user set;

[0015] The training module is further used to train the first biological information model through a second data set of the second user to generate a second biological information model and obtain a convergence value of the second biological information model, wherein the second data set is local data of the second user, and the local data of the second user includes second biological information data;

[0016] The execution module is used to confirm that the second biological information model is a target biological information model when the convergence value of the second biological information model is less than a first preset threshold value, and the target biological information model is a trained model.

[0017] In a third aspect, a computer device is provided, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the method described in the first aspect is implemented.

[0018] In a fourth aspect, a computer-readable storage medium is provided, on which a computer program is stored, characterized in that when the program is executed by a processor, the method described in the first aspect is implemented.

[0019] In a fifth aspect, a computer program product is provided. The computer program product includes instructions, and the instructions are executed by a processor to implement the method described in the first aspect.

[0020] Additional aspects and advantages of the present invention will be given in part in the following description and in part will be obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Other features, objects and advantages of the present application will become more apparent by reading the detailed description of non-limiting embodiments made with reference to the following drawings:

[0022] Figure 1 A schematic diagram of an implementation environment in the related technology;

[0023] Figure 2 A schematic diagram of a process for training a bioinformatics model based on bioinformatics data provided in an embodiment of the present application;

[0024] Figure 3 A schematic diagram of the composition of a user node for training a biological information model provided in an embodiment of the present application;

[0025] Figure 4 A schematic diagram of the structure of a biological information model training device based on biological information data provided in an embodiment of the present application;

[0026] Figure 5 A schematic diagram of the structure of a computer device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0027] The present application will be further described in detail below in conjunction with the accompanying drawings and embodiments. It is to be understood that the specific embodiments described herein are only used to explain the relevant inventions, rather than to limit the inventions. It should also be noted that, for ease of description, only the parts related to the invention are shown in the accompanying drawings.

[0028] It should be noted that, in the absence of conflict, the embodiments and features in the embodiments of the present application can be combined with each other. The following is an explanation of the usage scenarios of the embodiments of the present application.

[0029] With the continuous development and advancement of data analysis technology, the accuracy and speed of data analysis are constantly improving.

[0030] Data analysis is often performed through data models. In related technologies, the federated learning model is a method of training the model through data from multiple nodes and multiple users. Figure 1 As shown, the initial model can be sent to the central server 11 through the model aggregation node 10; thereafter, the central server 11 synchronously sends the initial model 13 to each user 12a-12n. After multiple users have completed the training of the initial model 13 using their own original data, the trained model is synchronously sent to the central server 11 again. The central server 11 sends the collected trained models to the model aggregation node 10, and the model aggregation node 10 converges and aggregates the obtained several trained models to generate a data analysis model that can eventually be used to actually analyze data.

[0031] It can be seen from the above training process that the above-mentioned federated learning training model method does not require each user or each node to upload original data. It uses its own original data for local training, which is more friendly to many industries that require data analysis and have strong requirements for privacy protection (for example, the medical industry or the financial industry, etc.).

[0032] However, this data model also has certain problems in practical applications, as described in the following book:

[0033] ① Heterogeneous data distribution: The statistical distribution of data stored in each node participating in federated learning is different. For example, in medical imaging diagnosis applications, the data of different medical centers are significantly different. For example, the proportion of benign and malignant lung nodules in lung CT images of general hospitals and tumor specialty hospitals is different. This makes the update direction of the federated model of each node very different, and it is difficult to effectively aggregate the model based on these directions, which ultimately makes it difficult for the learning models of each node to converge in the model aggregation stage.

[0034] ② Heterogeneous communication network: In the process of joint model training, the nodes participating in the training communicate with the central server through communication links, and the communication network of each node may not be able to ensure a continuous good state. When the communication links are heterogeneous, the number of participating nodes is large, and the data transmission load is heavy, the network bandwidth resources are scarce, the network is seriously congested, and the data transmission time is long, which greatly hinders efficient model training.

[0035] ③ Heterogeneous operation status: The nodes participating in the training have different network infrastructure, different software and hardware, and different computing and storage resources and capabilities. When training the model, the differences in resources and capabilities lead to different operation status. A few nodes with few resources and weak capabilities will become the bottleneck of federated learning, which is likely to seriously affect the overall training efficiency of the model.

[0036] ④ Limitations of traditional methods: The above traditional federated learning training process usually uses synchronous or semi-asynchronous methods to optimize the solution model. However, due to problems such as heterogeneous computing power, different nodes have different states when updating the model locally, making it difficult to achieve high learning efficiency.

[0037] As a result, the existing federated learning model may have problems such as poor training results.

[0038] Furthermore, in the related art, the model update and training in the federated learning training process usually adopts a large-scale random optimization method. The reason is that this large-scale random optimization method has a low computational cost (such as gradient descent method, mirror descent method and its randomized method). Specifically, this random optimization method can extract node samples iteratively and serially, and use the node samples to update the model.

[0039] Specifically, assuming that the node samples are in an independent and identically distributed (IID) state, these methods usually achieve sublinear convergence speed in the final model convergence stage. Specifically, among the algorithms currently assuming that the instances are IID sampling, there are two algorithms: 1) Asynchronous decentralized stochastic gradient descent algorithm: This algorithm is robust to communication in heterogeneous environments and achieves almost optimal convergence speed. 2) Decentralized optimization algorithm for random communication graphs: After each node parameter is updated, its neighbor nodes are selected to update the model.

[0040] However, in actual application scenarios, the IID assumption may be too extreme and unrealistic. Instead, a relatively mild assumption is that the instances are sampled from a Markov chain. Under this assumption, a large number of studies have developed Markov chain versions of stochastic optimization methods and further studied their convergence speed. Under the Markov chain sampling strategy, more and more federated optimization methods have been proposed and analyzed. Among the related technologies, they include: 1) Finite time convergence analysis of stochastic gradient descent, and the realization of Markov noise and time-varying communication network assumptions. Convergence speed; 2) Random walk stochastic gradient descent method in decentralized networks, and established , The convergence speed.

[0041] In addition, they also focus on alleviating communication bottlenecks and data heterogeneity during model training, and designing a near-optimal node sampling strategy for the model random walk. In the study of the convergence of Markov chain stochastic gradient descent, by choosing a step size of And 0.5<q<1 obtains the convex loss and non-convex loss The convergence speed.

[0042] In response to the above-mentioned convergence speed problem, the stability of Markov chain stochastic gradient descent is studied in related technologies, and a comprehensive analysis of minimization and maximum and minimum problems is conducted; 2) A single time scale algorithm based on moving average is used to find -Fixed point method, and obtained the non-convex loss 3) Adaptive methods that use Markov data for random optimization and do not require knowledge of mixing time; 4) Related techniques also achieve convex and non-convex losses The convergence rate of the least squares loss in related techniques is improved, and the information-theoretic minimax lower bound is established when sampling from Markov chains. The analytical results show that optimization using Markov data is generally more difficult than optimization using independent data. The related techniques provide a unified analytical framework for accelerated versions of first-order gradient methods with Markov noise and extend them to variational inequalities.

[0043] However, existing studies have mainly conducted rigorous analysis on the convergence rate of stochastic gradient descent. However, these analyses cannot be directly extended to the mirror descent algorithm. We note that Duchi et al. analyzed the convergence of Markov chain mirror descent and found that the convergence rate is as fast as In the related art, a random walk strategy for node selection is proposed, and the ADMM method is used to update the local model on the node. Compared with the existing research, this method has achieved certain improvement in convergence speed.

[0044] The pre-federation optimization method generally adopts a synchronous or semi-asynchronous method to optimize and solve the model. Due to problems such as heterogeneous computing power, different nodes have different states when updating the model locally, making it difficult to achieve high learning efficiency.

[0045] Furthermore, in the related art, there are also methods for random optimization through Markov chains. However, although there have been extensive studies analyzing the convergence of Markov chain random optimization methods, Markov chains currently usually have the following problems: the analysis of convergence in related technologies is mostly focused on stochastic gradient descent, and there are few results on mirror descent. Therefore, there is always the problem that the trained model cannot converge.

[0046] Based on this, the present application proposes a model training method, device, equipment and medium, which can solve the problem in the related technology that due to the high network communication requirements of the above-mentioned learning model training process, the training process will generate a training bottleneck when the communication resources are relatively limited, and it is impossible to obtain the best training effect and a learning model with strong data analysis capabilities. It can greatly improve the training efficiency and reduce the communication resources required for the training process.

[0047] Figure 1 1 is a flow chart of a method for training a biological information model based on biological information data provided by an embodiment of the present application. The execution subject of the method may be a server for training a biological information model. Figure 1 As shown, the method includes the following steps 301-304:

[0048] Step 301: Train an initial biometric information model using a first data set of a first user in a user set to generate a first biometric information model.

[0049] In an embodiment of the present application, the above user set includes all user nodes that can be used to train the model.

[0050] It is understandable that in the embodiments of the present application, not all of the user nodes in the user set will actually execute the process of training the model during the model training process. In fact, the user nodes in the user set in the present application are sufficient to ensure that the initial bioinformatics model is trained to become a model that meets the requirements and can be used to process data information. However, in the process of training the model in turn by each user in the user set until a usable model is finally formed, it may not be necessary to use all the users in the user set.

[0051] In the embodiment of the present application, the first data set is the local data of the first user, and the local data of the first user includes first biometric information data.

[0052] It is understandable that the learning method of the present application can be a federated learning method, that is, the training model is iteratively trained between the user nodes of different users in the user set in turn to finally form a usable model. Therefore, in the process of the first user training the initial bioinformatics model, its local data is also used.

[0053] Furthermore, since the data sets used to train the initial bioinformatics model are all local data, this training model method can be used in industries involving sensitive information, such as the medical industry and the financial industry. The model generated by the final training can be used to process medical data information or financial data information.

[0054] In one example, the model training method adopted in the embodiment of the present application is a Markov chain training mode, which eventually makes the Markov chain training model converge and generates a target bioinformatics model.

[0055] Step 302: Filter the second user in the user set.

[0056] In the embodiment of the present application, the first user and the second user are different users in the user set.

[0057] Exemplarily, the purpose of the above-mentioned screening of the second user is to transmit the above-mentioned first biological information model to the user node of the second user to perform model training of the first biological information model.

[0058] It can be understood that the model is trained by federated learning in this application, that is, the local data of user nodes of multiple users are needed to train the model in order to complete the model training process.

[0059] Furthermore, different from the federated learning model in the related art, that is, the initial model is sent to each user node through the central server, and all user nodes are sent back to the server after training. The central server then transmits the training model of each user node to the converger, and the converger converges uniformly. After setting the user set, this application trains multiple users in the user set in sequence. When a certain user is used as a user node, the learning model reaches the convergence condition, and the training is considered to be completed. Therefore, this application needs to screen the second user for the next training after obtaining the first bioinformatics model.

[0060] Exemplarily, the above-mentioned method of screening the second user may include any of the following: 1) randomly screening after determining the adjacent users corresponding to the first user; 2) screening according to relevant parameter conditions of the adjacent users after determining the adjacent users corresponding to the first user, which is not limited in this application.

[0061] The second screening method, that is, "screening according to relevant parameter conditions of the neighboring users after determining the neighboring users corresponding to the first user" is described below.

[0062] Optionally, in an embodiment of the present application, for the second user in the above-mentioned screened user set, the model training method provided in the embodiment of the present application includes: determining the adjacent user set of the first user in the above-mentioned user set, and sending a first request message to each adjacent user in the adjacent user set; receiving a first response message sent by an adjacent user in the above-mentioned adjacent user set, and confirming the sending duration of the first response message sent by each of the above-mentioned adjacent users based on the time point when the first response message sent by the above-mentioned adjacent user is received; and confirming the target user in the adjacent user set as the second user based on the sending duration of the first response message sent by each of the above-mentioned adjacent users and the current user state parameters of the above-mentioned adjacent users.

[0063] Exemplarily, for a user set, each user is a user node, which is equipped with a server and can be used to train a model. Further, for each user node, it has a node position, and the neighboring users around it constitute a neighboring user set, that is, a neighboring user node. In the process of selecting the second user, the user is generally selected from the neighboring user set.

[0064] In one example, the neighboring user set may be an edge set.

[0065] Exemplarily, the first request information is used to request each neighboring user to send a current user state parameter.

[0066] Exemplarily, each neighboring user in the neighboring user set of the first user has a different relevant state parameter relative to the first user.

[0067] Exemplarily, the above-mentioned related status parameters include distance parameters, communication quality parameters, computing power status parameters, idle state parameters, etc.

[0068] Exemplarily, by requesting each neighboring user in the neighboring user set to send its respective current user state parameter to the first user, the most suitable next node can be screened out.

[0069] Exemplarily, the first response information is used to respond to the first request information and send the current user state parameter to the first user.

[0070] For example, after sending a first request message to each neighboring user in the neighboring user set, it cannot be guaranteed that each neighboring user will send a first response request to the first user. Based on this, the first response message received is the first response message sent by the neighboring users in the neighboring user set, and not necessarily the first response message sent by each neighboring user.

[0071] Exemplarily, the above-mentioned sending duration means: after receiving the first response information sent by the above-mentioned adjacent user, the time interval between the sending time and the receiving time of the first response information can be calculated, thereby obtaining the interval duration between receiving information and sending information for each adjacent user sending the first response information.

[0072] It can be understood that the interval duration can be used to determine the transmission communication interval duration between the user node of the current first user and the adjacent user who may become the next training node in the future, which can help determine the communication quality during the transmission of the first biometric information model, for example, the communication duration.

[0073] Exemplarily, after receiving the first response information including the current user state parameters and the sending duration, the second user in the adjacent user set who can serve as the second user to receive the first biometric information model and train the first biometric information model can be determined based on these parameters.

[0074] Optionally, in an embodiment of the present application, the specific method of determining the second user based on the above-mentioned sending duration and the above-mentioned current user status parameters is as follows. Exemplarily, the above-mentioned method of confirming the target user among the adjacent users as the second user based on the sending duration of the first response information sent by the above-mentioned adjacent user and the current user status parameters of the above-mentioned adjacent user, includes in the model training method provided in the embodiment of the present application: according to a preset weighted value, the sending duration of the first response information sent by the above-mentioned adjacent user and the current user status parameters of the above-mentioned adjacent user are weighted averaged to confirm the weighted average value; according to the above-mentioned weighted average value, confirm the target user in the adjacent user set as the second user.

[0075] Exemplarily, the above-mentioned current user state parameters include: computing power state parameters and idle state parameters.

[0076] In one example, the computing power status parameter is used to indicate the computing power status of the server of the neighboring user. It is understandable that the computing power of the servers of different neighboring users is not the same, and the computing power status can be used to indicate the speed at which the training model is acquired. For example, the better the computing power status, the stronger the computing power, and the faster the training model; the worse the computing power status, the weaker the computing power, and the slower the training model.

[0077] In one example, the idle state parameter is used to indicate whether the server of the neighboring user is in a working state or an idle state. It is understandable that if the server of the neighboring user is in a working state, the server of the neighboring user may temporarily be unable to start the model training of the first bio-information model immediately after receiving the first bio-information model; correspondingly, if the server of the neighboring user is in an idle state, the server of the neighboring user may start the model training of the first bio-information model immediately after receiving the first bio-information model.

[0078] Exemplarily, as described above, the current user status parameter may also include a distance parameter, which is used to characterize the distance between the first user and the adjacent user. Combined with the transmission duration, the communication quality between the first user and the adjacent user can be determined.

[0079] Exemplarily, the above-mentioned preset weighted value may be preset or user-defined, which is not limited in the embodiments of the present application.

[0080] Exemplarily, by assigning a preset weighted value to each feature of the above-mentioned current user parameters and sending duration, the calculated weighted average can indicate the training effect predicted after receiving the first bio-information model and receiving the first bio-information model, such as the best adjacent user in terms of training efficiency, and is determined as the target user and the second user.

[0081] In this way, the second user is screened and determined in the above manner, so that the user node training data of the best training model can be finally screened out, that is, a suitable network topology is generated. According to the factors such as the number, geographical location, and communication capability of each user node, that is, each participant, a suitable network topology is selected. For example, for a situation where the number of participants is small and the communication conditions are good, a fully connected topology can be used to ensure efficient communication between nodes; for a situation where the number of participants is large and widely distributed, a hierarchical topology or a distributed topology can be used to reduce communication overhead and improve the scalability of the system.

[0082] In addition, during the model training process, this progressive and dynamic node screening training method can dynamically adjust the network topology according to the communication status and node operation status. For example, when it is found that the communication congestion between some nodes is serious, the topology can be temporarily adjusted to bypass the congested links or regroup some nodes to improve communication efficiency.

[0083] Furthermore, in order to ensure the communication efficiency between nodes, the communication quality can be ensured by adopting efficient communication protocols, as well as compression and encryption optimization. The specific methods are as follows:

[0084] 1) Use efficient communication protocols: Choose an efficient communication protocol to reduce data transmission overhead and increase transmission speed. For example, you can use some improved protocols based on UDP, or some communication protocols designed specifically for distributed systems, such as gRPC. These protocols can improve communication efficiency while ensuring data reliability.

[0085] 2) Compression and encryption optimization: compress and encrypt the transmitted data. Data compression can reduce the amount of data transmission and increase the transmission speed; data encryption can ensure the security and privacy of data. Some advanced compression algorithms and encryption algorithms can be used, such as LZ77 compression algorithm, AES encryption algorithm, etc., and the algorithm parameters can be optimized according to the actual situation to achieve the best effect.

[0086] Step 303: training the first biological information model using the second data set of the second user to generate a second biological information model, and obtaining a convergence value of the second biological information model.

[0087] In the embodiment of the present application, the second data set is the local data of the second user, and the local data of the second user includes second biometric information data.

[0088] It can be understood that, from the foregoing content, in the embodiment of the present application, the training process of the user node is to iterate the training model multiple times through the user nodes of several users in the user training set. Therefore, after the first user uses his own local data set, that is, the first data set, as the initial bio-information model for training, the generated first bio-information model will be transmitted to the user node of the selected second user to continue training, and the trained model is the second bio-information model.

[0089] In the model in the embodiment of the present application, before the model trained by the user node reaches the convergence condition, the next user will be continuously screened for the untrained model, and the model will be transmitted to the user node of the next user for training. Based on this, after the second bio-information model is generated, the convergence value of the second bio-information model will be generated, and the convergence value can be used to indicate the change between the first bio-information model and the second bio-information model, that is, the change range of the second bio-information model compared with the first bio-information model, and further indicate whether the second bio-information model has reached a stable effect compared with the first bio-information model and does not need to be trained further.

[0090] Step 304: When the convergence value of the second biological information model is less than the first preset threshold, confirming the second biological information model as the target biological information model.

[0091] In the embodiment of the present application, the above-mentioned target biological information model is a trained model.

[0092] It can be understood that from the above content, the convergence value of the second bio-information model can be used to show the changes of the second bio-information model compared with the first bio-information model, and then reflect whether the second bio-information model is already in a stable and usable state, thereby determining that it does not need to continue training.

[0093] In the embodiment of the present application, the above-mentioned first preset threshold value can be a user-defined setting or can be preset in advance, and the embodiment of the present application does not limit this.

[0094] It can be understood that when the convergence value of the second bio-information model is less than the first preset threshold, it means that it has reached the convergence condition and no further training is needed; and when the convergence value of the second bio-information model is still greater than or equal to the first preset threshold, it means that it has not reached the convergence condition and it is necessary to continue looking for the next user, such as the third user, to continue training.

[0095] Furthermore, the method for screening the third user may refer to the method for screening the second user in the above step 302; the method for training the second bio-information model through the third user may refer to the method for training the second bio-information model in the above step 303; the method for determining whether the third model generated after training by the third user is the final target bio-information model may refer to step 304, comparing whether the convergence value of the third model is less than the first preset threshold, if so, terminate the training, otherwise continue the training.

[0096] like Figure 3 As shown, Figure 3 That is, this is a schematic diagram of the application sequentially screening the user nodes of multiple users as the training nodes of the training model. It can be seen that it can screen out user nodes in sequence, from the first user 21 to the fifth user 25, until the fifth training. After the model converges, the subsequent user nodes do not need to continue training, that is, the sixth user 26 and the seventh user 27. The trained users are connected to each other by solid lines, and the non-trained users are connected by dotted lines.

[0097] In the method provided in the embodiment of the present application, the initial bio-information model is trained by the local data of the first user in the user set, that is, the first data set, to generate a first bio-information model; then, a second user different from the first user in the user set is screened, and the first bio-information model is trained by the local data of the second user, that is, the second data set, to produce a second bio-information model, and the convergence value of the second bio-information model is obtained; finally, when the convergence value of the second bio-information model is less than the first preset threshold value, the second bio-information model is confirmed to be the target bio-information model that has been trained. In this way, in the process of training the model of the present application, the model is trained in sequence for the screened users. On the one hand, it avoids the situation in which the initial bioinformatics model below is unified and waits for the training of each node user before unified collection and aggregation, resulting in insufficient communication bandwidth for collection. On the other hand, it also avoids the problem of low efficiency caused by the need to wait for all node users to complete training before feeding back the model in the unified collection process. In the method of the present application, the model is trained by randomly screening node users and training different node users in sequence. After training to any node user, if the training model of the node user can converge, the model of the node user can be determined as the target bioinformatics model, thereby saving the demand for communication bandwidth for sending and retrieving the model and greatly improving the training efficiency.

[0098] It is understandable that the above-mentioned method of training the user node of the first user and then training through the second user in this application is to update the model in an asynchronous manner, so that the user node can continue its own training and update process without waiting for other user nodes to complete the update. For example, the user node can trigger the update operation according to a certain time interval or condition based on its own computing resources and data conditions. This can avoid the situation where the node with fast processing progress in synchronous training during federated learning waits for the slow node, thereby improving the overall training.

[0099] In another embodiment of the present application, a specific implementation method for training the initial bio-information model is also provided. Exemplarily, the specific implementation of "training the initial bio-information model through the first data set of the first user in the user set to generate the first bio-information model" mentioned above includes: training the initial bio-information model in sequence through the m first data subsets in the above first data set to obtain m training results; confirming the training results less than or equal to the second preset threshold among the above m training results as the target training results; and updating the above initial bio-information model according to the above target training results to generate the above first bio-information model.

[0100] Exemplarily, the first data subset corresponds to a training result, and m is a positive integer.

[0101] It can be understood that in the embodiment of the present application, the data of the above-mentioned first data set will be pre-divided into multiple first data subsets, that is, m first data subsets, and then, for example, each data subset in the m first data subsets will train the initial bioinformatics model in turn to generate m training results.

[0102] Exemplarily, the second preset threshold is used to indicate a reliable training result.

[0103] It can be understood that the second preset threshold can indicate whether the training process corresponding to the m training results is a valid training process, that is, it can be used to indicate which training results among the m training results can be used for reference to generate the first bio-information model. Finally, the initial bio-information model is updated using these target training results that are less than or equal to the second preset threshold to generate the first bio-information model.

[0104] For example, the first data set has 100 first data subsets. Through the training of the 100 first data subsets, 100 training results will be obtained, among which 95 training results are less than or equal to the second preset threshold, and thus 95 training results belong to the target training results.

[0105] It is understandable that in the process of training the initial biometric information model using the local data of the first user, the data in the local data should be preprocessed so that the first data set finally used for training meets the characteristics of standardization and non-abnormality. The data standard of the first data set is introduced as follows:

[0106] 1) Data standardization: Standardize the local data of the first user's user node to make it have similar statistical distribution characteristics. For example, methods such as mean normalization and standard deviation normalization can be used to map the characteristic values ​​of the data to a specific interval. This can reduce the impact of data distribution heterogeneity on model training and improve the convergence performance of the model.

[0107] 2) Data enhancement: By performing some transformation operations on the local data of the first user's user node, such as rotation, flipping, scaling, etc., the diversity of the data is increased. For medical imaging data, some specific enhancement methods can also be used, such as adjusting contrast, adding noise, etc. Data enhancement can enable the model to better learn the characteristics of the data and improve the generalization ability of the model.

[0108] 3) Selection based on feature importance: Analyze the feature importance of the local data of the first user's user node, and select the features that are most helpful for model training. For each node in federated learning, some feature selection algorithms can be used, such as information gain, Gini index, etc., to determine the most important features in the local data of each node, and then only use these important features for model training, which can reduce the dimension of the data and improve training efficiency.

[0109] 4) Abnormal data filtering: Identify and filter out abnormal data in the local data of the first user's user node. Abnormal data may have a negative impact on the training of the model, causing difficulty in model convergence or deviation. Some anomaly detection algorithms, such as statistical methods, clustering methods, etc., can be used to detect and remove abnormal data points to improve data quality.

[0110] Optionally, in an embodiment of the present application, the initial bio-information model is updated according to the target training result to generate a first bio-information model. The model training method in the embodiment of the present application includes: according to the target training result, the initial bio-information model is updated using a preset averaging algorithm to generate a first bio-information model.

[0111] Exemplarily, after obtaining m groups of training results, the initial bio-information model has actually performed m trainings accordingly. In this case, the target training results are further screened out, and the model parameters need to be averaged according to the target training results to finally obtain the first bio-information model.

[0112] Example 1: In this example 1, there are y groups of second data subsets, which are applied to the user node of the second user.

[0113] Step S1: A user node of a first user receives a first biometric information model;

[0114] Step S2: the user node of the second user has y groups of first biometric information models of the second data subsets, the y groups of second data subsets are numbered 0, 1, 2, …, m−1 respectively, and the initial biometric information model is trained from the second data subset No. 0 until the second data subset No. y−1;

[0115] Step S3: After obtaining x training results, the number of feasible solutions that can be used for final training can be determined to be m (target training results) through the constraint formula (i.e., the second preset threshold value mentioned above). The constraint formula is shown as follows:

[0116]

[0117] Indicates that the id is The external variance between user nodes of the first user is expressed as , the internal variance of the user node of the first user is

[0118] The final feasible solutions include the following m:

[0119]

[0120] These feasible solutions include training samples that can be used to calculate the current stochastic gradient, which is .

[0121] Step S3: According to the aforementioned target training results, the first bioinformatics model is obtained by the following Markov chain model generation formula, that is, :

[0122]

[0123] Step S4: Averaging the model parameters using a preset averaging algorithm to finally obtain the first bioinformatics model :

[0124]

[0125] In another embodiment of the present application, a specific implementation method for training the first bio-information model is also provided. Exemplarily, the aforementioned "training the first bio-information model through the second data set of the second user, generating a second bio-information model, and obtaining the convergence value of the second bio-information model" includes: training the first bio-information model in sequence through t second data subsets in the second data set to obtain t training results; confirming the training results less than or equal to the second preset threshold among the t training results as the target training results; updating the first bio-information model according to the above target training results, the current stochastic gradient and the current divergence to generate the second bio-information model.

[0126] Exemplarily, the above-mentioned second data subset corresponds to a training result, and t is a positive integer.

[0127] Exemplarily, the second data set also has the characteristics of standardization, non-abnormality, etc. in the first data set, and please refer to the above content for details.

[0128] It can be understood that in the embodiment of the present application, the data of the above-mentioned second data set will be pre-divided into multiple second data subsets, that is, t second data subsets, and then, for example, each data subset in the t second data subsets will train the initial bioinformatics model in turn to generate t training results.

[0129] Exemplarily, the second preset threshold is used to indicate a reliable training result.

[0130] Exemplarily, the second preset threshold may be as described above and will not be described again here.

[0131] Exemplarily, the current stochastic gradient is used to characterize the gradient of change between the first bio-information model and the target training result, and the current divergence is used to characterize the user divergence relationship between the first user and the second user.

[0132] It can be understood that the current stochastic gradient is used to characterize the change between the first bio-information model and the target training result generated after the user node training of the second user, thereby determining the second bio-information model. Further, the current stochastic gradient can characterize the overall convergence value of the second bio-information model, and when the convergence value is less than the first preset threshold, the target bio-information model can be determined.

[0133] Exemplarily, the current divergence may represent a divergence relationship between the first user node and the second user node, and the second biometric information model may be generated through the divergence relationship.

[0134] Optionally, in an embodiment of the present application, a method for determining a current stochastic gradient is also provided. Exemplarily, before the first bio-information model is updated according to the target training result, the current stochastic gradient and the current divergence and the second bio-information model is generated, the model training method provided in the embodiment of the present application further includes: confirming a target first data subset corresponding to the target training result; and obtaining a current stochastic gradient corresponding to the second user according to the target first data subset and a preset random function.

[0135] Exemplarily, the target first data subset is the second data subset corresponding to the target training result, that is, the second data subset corresponding to the training result that is less than the second preset threshold.

[0136] Exemplarily, the preset random function is a function used to calculate the current stochastic gradient.

[0137] It can be understood that the preset random function may be pre-set in the user node of the second user.

[0138] Optionally, in an embodiment of the present application, a specific method for generating a second bio-information model is also provided. Exemplarily, the above-mentioned updating of the first bio-information model and generating the second bio-information model according to the target training result, the current stochastic gradient and the current divergence, the model training method provided in the embodiment of the present application includes: obtaining the current divergence between the first user and the second user through a preset divergence function; updating the first bio-information model and generating the second bio-information model according to the target training result, the current stochastic gradient and the current divergence.

[0139] Exemplarily, the current divergence may be used to characterize the divergence relationship between the first biometric information model of the current user node, ie, the user node of the second user, and the previous user node, ie, the user node of the first user.

[0140] In one example, the current divergence may be determined by a preset divergence function.

[0141] represents the Bregman divergence between u and v, where is a given distance function. Wherein, V is the model generated by the user node of the second user, and U is the first biometric information model of the user node of the first user. Specifically, the current divergence can be:

[0142]

[0143] Optionally, in an embodiment of the present application, a specific method for generating a second biological information model is also provided. The updating of the first biological information model according to the target training result and the current stochastic gradient to generate the second biological information model includes: updating the initial biological information model according to the target training result and the current stochastic gradient using a preset averaging algorithm to generate the second biological information model.

[0144] Exemplarily, the preset averaging algorithm is as described above and will not be described in detail here.

[0145] Assuming that the final generated target selection results include T of the t training results, the formula for the average preset algorithm is .

[0146] The following describes the method of training the user node model of the second user in the present application through Example 2:

[0147] Example 2: In this Example 2, there are x sets of second data subsets, which are applied to the user node of the second user.

[0148] Step S1: The user node of the second user receives the first biometric information model;

[0149] Step S2: the user node of the second user has x groups of first biometric information models of the second data subsets, the numbers of the x groups of second data subsets are 0, 1, 2, …, x−1 respectively, and the first biometric information model is trained starting from the second data subset No. 0 and training until the second data subset No. x−1;

[0150] Step S3: After obtaining x training results, the number of feasible solutions (target training results) that can be used for final training can be determined through the constraint formula (i.e., the second preset threshold value mentioned above). The constraint formula is as follows:

[0151]

[0152] Indicates that the id is The external variance between user nodes of the second user is expressed as , the internal variance of the user node of the second user is .

[0153] The final feasible solutions include the following t:

[0154]

[0155] These feasible solutions include training samples that can be used to calculate the current stochastic gradient, which is .

[0156] Step S4: Calculate the current stochastic gradient by using a preset stochastic gradient function (i.e., the preset stochastic function mentioned above). The current stochastic gradient is used to characterize the stochastic gradient between the above t target training results and the first bioinformatics model, i.e., the Markov chain stochastic gradient, and the formula is as follows:

[0157]

[0158] Step S5: After obtaining the current stochastic gradient, determine the current divergence. The divergence formula is as described above and will not be repeated here.

[0159] Step S6: According to the current stochastic gradient and the current divergence and the aforementioned target training results, the second bio-information model is obtained by the following Markov chain model generation formula, and according to the comparison between the current stochastic gradient and the first preset threshold, when the current stochastic gradient is less than the first preset threshold, the second bio-information model can be determined as the target bio-information model, that is, For the target bioinformatic model:

[0160] .

[0161] Step S7: Averaging the model parameters using a preset averaging algorithm to finally obtain the first bioinformatics model :

[0162]

[0163] It should be noted that this application is embodied in the technology of mirror descent in the convergence process of Markov chain random optimization carried out for federated learning. In this application, the overall global model of federated learning used is as follows:

[0164] The goal of federated learning is to find a global model for the following optimization problem:

[0165]

[0166] in , is the objective function of the local device (i.e., a user, such as the first user or the second user). In order to simultaneously meet the communication constraints, the problem can be decomposed into multiple sub-problems, and the local objective function of each sub-problem is defined as: , so the global objective function can be written as:

[0167]

[0168] In combination with the above example 1, the federated learning formula corresponding to its global objective function is as follows:

[0169]

[0170] represents a neural network, represents each user node, such as the user node of the first user or the user node of the second user, The adjacent user nodes of their user nodes, that is, the adjacent user nodes of the user node of the first user and the adjacent user nodes of the user node of the second user, can also be called edge sets. Represents a node, yes The number of nodes in .

[0171] The following is a specific explanation and description of the convergence analysis of the Markov chain in the federated learning model in this application, that is, the Markov chain mirror descent method:

[0172] Based on the following assumptions, the convergence analysis of the proposed Markov chain optimization method is carried out.

[0173] Assumption 1 (Markov chain). Assume that Markov chain It is time homogeneous, irreducible, and non-periodic. It has a transfer matrix P and its stationary distribution ,in .

[0174] Assumption 2 (Stationarity). of is L-smooth, that is, for any x and y,

[0175] .

[0176] Assumption 3 (Bounded Variance). Assumption and , and for any , .

[0177] Assumption 4 (Bounded Gradient). Assumption For any and x both satisfy.

[0178] Assumption 5 (bounded domain). Assumption This holds true for any x and y.

[0179] Assumption 6. Assumption yes , that is, for any x and y, .

[0180] Afterwards, we introduce some constants that will be used in regret analysis. These constants were first introduced in Sun et al. and then widely used in existing work Even. Assume that the Jordan normal form of the transfer matrix P is expressed as

[0181]

[0182] where m is the number of blocks, It is Block submatrix The dimension of In addition, assuming Indicates the first The largest eigenvalue, Expressed as

[0183] .

[0184] Specifically, Defined as

[0185]

[0186] Defined as

[0187] .

[0188] constant Defined as:

[0189] ,

[0190] ,

[0191] ,

[0192] ,

[0193] ,

[0194] .

[0195] Recall that for a network N, a Markov chain is express: ,in Given a Markov chain on a network N , the regret of any optimization method A is defined by the following formula

[0196]

[0197] The local loss function Defined as:

[0198] .

[0199] Theorem 1. Given a Markov chain , the sequence of models Produced by any optimization method A, we have

[0200] .

[0201] in , Defined in 6. Note represents the maximum absolute value of element A, that is Theorem 1 shows that the analysis of convergence rate can be simplified to analyzing the optimal regret of model x*. Therefore, we focus on the regret analysis of the proposed method and then directly implement its convergence rate.

[0202] 1. Regret Analysis of Convex Loss

[0203] When the node The loss function When is convex, its regret analysis is shown in Theorem 2.

[0204] Theorem 2. From Algorithm 1, we get t=1 and select a non-increasing positive sequence . Assume that for any , both According to assumptions 1-6, we have

[0205]

[0206] when When setting an appropriate learning rate , we obtain the following regret bound.

[0207] Corollary 1. According to Algorithm 1, let , and select by

[0208]

[0209] According to assumptions 1-6, we have

[0210] .

[0211] at this time, For any They are generally convex.

[0212] Recall Theorem 1 and set the learning rate Setting this to Corollary 1, we get

[0213] ,

[0214] For convex functions Any All are established.

[0215] The present application also provides a computer-readable storage medium on which a computer program is stored. When the program is executed by a processor, the training rule determination method described in the present application is implemented. For example, the following can be executed: Figure 2 The individual steps of the method are shown.

[0216] The present invention provides a computer program product, which includes instructions that are executed by a processor when the processor is running. Figure 2 The individual steps of the method are shown.

[0217] It should be noted that although the operations of the method of the present invention are described in a particular order in the drawings, this does not require or imply that the operations must be performed in this particular order or that all illustrated operations must be performed to achieve desired results.

[0218] Figure 4This is a block diagram of a biological information model training device based on biological information data according to an embodiment of the present application. Figure 4 The device includes a training module 601, a screening module 602 and an execution module 603.

[0219] A training module 601 is used to train an initial biometric information model using a first data set of a first user in a user set to generate a first biometric information model, wherein the first data set is local data of the first user, and the local data of the first user includes first biometric information data;

[0220] A screening module 602, configured to screen a second user in the user set, wherein the first user and the second user are different users in the user set;

[0221] The training module 601 is further used to train the first biological information model through a second data set of the second user to generate a second biological information model and obtain a convergence value of the second biological information model, wherein the second data set is local data of the second user, and the local data of the second user includes second biological information data;

[0222] The execution module 603 is used to confirm that the second biological information model is a target biological information model when the convergence value of the second biological information model is less than a first preset threshold, and the target biological information model is a trained model.

[0223] In one embodiment, the training module 601 is specifically used to: train the initial bioinformatics model in sequence through m first data subsets in the first data set to obtain m training results, where one first data subset corresponds to one training result, and m is a positive integer; confirm the training result of the m training results that is less than or equal to a second preset threshold as the target training result, where the second preset threshold is used to indicate a reliable training result;

[0224] According to the target training result, the initial biological information model is updated to generate the first biological information model.

[0225] In one embodiment, the training module 601 is specifically used to: update the initial biological information model using a preset averaging algorithm according to the target training result to generate a first biological information model.

[0226] In one embodiment, the above-mentioned training module 601 is specifically used to: train the first bio-information model in sequence through t second data subsets in the second data set to obtain t training results, wherein one second data subset corresponds to one training result, and t is a positive integer; confirm the training results less than or equal to a second preset threshold among the t training results as target training results, and the second preset threshold is used to indicate a reliable training result; update the first bio-information model according to the target training result, the current stochastic gradient and the current divergence to generate a second bio-information model, wherein the current stochastic gradient is used to characterize the change gradient between the first bio-information model and the target training result, and the current divergence is used to characterize the user divergence relationship between the first user and the second user.

[0227] In one embodiment, the training module 601 is specifically used to: confirm the target first data subset corresponding to the target training result; and obtain the current random gradient corresponding to the second user according to the target first data subset and a preset random function.

[0228] In one embodiment, the execution module 603 is specifically used to: obtain the current divergence between the first user and the second user through a preset divergence function; update the first bio-information model according to the target training result, the current stochastic gradient and the current divergence, and generate the second bio-information model.

[0229] In one embodiment, the execution module 603 is specifically used to: update the initial biological information model using a preset averaging algorithm according to the target training result and the current stochastic gradient to generate a second biological information model.

[0230] In one embodiment, the above-mentioned screening module 602 is specifically used to: determine a neighboring user set of the first user in the user set, and send a first request message to each neighboring user in the neighboring user set, wherein the first request message is used to request each neighboring user to send a current user state parameter; receive a first response message sent by a neighboring user in the neighboring user set, and confirm a sending duration of the first response message sent by each neighboring user based on a time point when the first response message sent by each neighboring user is received; and confirm a target user in the neighboring user set as the second user based on the sending duration of the first response message sent by the neighboring user and the current user state parameter of the neighboring user.

[0231] In one embodiment, the above-mentioned screening module 602 is specifically used to: perform weighted averaging on the sending time of the first response information sent by each adjacent user and the current user state parameters of each adjacent user according to a preset weighted value, and confirm the weighted average value, wherein the current user state parameters include: computing power state parameters and idle state parameters; and confirm the target user in the adjacent user set as the second user according to the weighted average value.

[0232] It should be understood that the units recorded in the model training device correspond to the various steps in the method described in the accompanying drawings. Therefore, the operations and features described above for the method are also applicable to the model training device and the units contained therein, and will not be repeated here. The model training device can be pre-implemented in a browser or other security application of a computer device, or can be loaded into a browser or its security application of a computer device by downloading or the like. The corresponding units in the model training device can cooperate with the units in the computer device to implement the solution of the embodiment of the present application.

[0233] For the several modules or units mentioned in the above detailed description, this division is not mandatory. In fact, according to the embodiments of the present disclosure, the features and functions of two or more modules or units described above can be embodied in one module or unit. On the contrary, the features and functions of one module or unit described above can be further divided into multiple modules or units to be embodied.

[0234] It should be noted that for details not disclosed in the model training device of the embodiment of the present application, please refer to the details disclosed in the above embodiments of the present application, which will not be repeated here.

[0235] Reference below Figure 5 , Figure 5 FIG. 1 is a schematic diagram showing a structure of a computer device suitable for implementing an embodiment of the present application. Figure 5 As shown, the computer system 1700 includes a central processing unit (CPU) 1701, which can perform various appropriate actions and processes according to the program stored in the read-only memory (ROM) 1702 or the program loaded from the storage part 1708 to the random access memory (RAM) 1703. In the RAM 1703, various programs and data required for the operation instructions of the system are also stored. The CPU 1701, the ROM 1702, and the RAM 1703 are connected to each other through a bus 1704. An input / output (I / O) interface 1705 is also connected to the bus 1704.

[0236] The following components are connected to the I / O interface 1705: an input section 1706 including a keyboard, a mouse, etc.; an output section 1707 including a cathode ray tube (CRT), a liquid crystal display (LCD), etc., and a speaker, etc.; a storage section 1708 including a hard disk, etc.; and a communication section 1709 including a network interface card such as a LAN card, a modem, etc. The communication section 1709 performs communication processing via a network such as the Internet. A drive 1710 is also connected to the I / O interface 1705 as needed. A removable medium 1711, such as a magnetic disk, an optical disk, a magneto-optical disk, a semiconductor memory, etc., is installed on the drive 1710 as needed, so that a computer program read therefrom is installed into the storage section 1708 as needed.

[0237] In particular, according to an embodiment of the present application, the above reference flow chart Figure 2 The described process can be implemented as a computer software program. For example, an embodiment of the present application includes a computer program product, which includes a computer program carried on a computer-readable medium, and the computer program includes a program code for executing the method shown in the flowchart. In such an embodiment, the computer program includes a program code for executing the method shown in the flowchart. In such an embodiment, the computer program can be downloaded and installed from a network through the communication part 1709, and / or installed from a removable medium 1711. When the computer program is executed by the central processing unit (CPU) 1701, the above-mentioned functions defined in the system of the present application are executed.

[0238] It should be noted that the computer-readable medium shown in the present application may be a computer-readable signal medium or a computer-readable storage medium or any combination of the above two. The computer-readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, device or device, or any combination of the above. More specific examples of computer-readable storage media may include, but are not limited to: an electrical connection with one or more wires, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above. In the present application, a computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in combination with an instruction execution system, device or device. In the present application, a computer-readable signal medium may include a data signal propagated in a baseband or as part of a carrier wave, which carries a computer-readable program code. This propagated data signal may take a variety of forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination of the above. Computer-readable signal media may also be any computer-readable medium such as a computer-readable storage medium that can send, propagate or transmit a program for use by or in conjunction with an instruction execution system, device or device. The program code contained on the computer-readable medium may be transmitted using any appropriate medium, including but not limited to: wireless, wire, optical cable, RF, etc., or any suitable combination of the above.

[0239] The flow chart and block diagram in the accompanying drawings illustrate the possible architecture, functions and operating instructions of the system, method and computer program product according to various embodiments of the present application. In this regard, each box in the flow chart or block diagram can represent a module, a program segment or a part of a code, and the aforementioned module, program segment or a part of a code contains one or more executable instructions for realizing the specified logical function. It should also be noted that in some implementations as replacements, the functions marked in the box can also occur in a sequence different from that marked in the accompanying drawings. For example, the boxes represented by two connections can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each box in the block diagram and / or flow chart, and the combination of the boxes in the block diagram and / or flow chart can be implemented with a dedicated hardware-based system that performs a specified function or operating instruction, or can be implemented with a combination of dedicated hardware and computer instructions.

[0240] The units or modules involved in the embodiments described in the present application may be implemented by software or hardware. The units or modules described may also be set in a processor. For example, it may be described as: a processor includes a first receiving module, a second receiving module, and a sending module. The names of these units or modules do not, in some cases, constitute limitations on the units or modules themselves.

[0241] As another aspect, the present application further provides a computer-readable storage medium, which may be included in the electronic device described in the above embodiment, or may exist independently without being assembled into the electronic device. The above computer-readable storage medium stores one or more programs, and when the above programs are used by one or more processors to execute the model training method described in the present application.

[0242] The above description is only a preferred embodiment of the present application and an explanation of the technical principles used. Those skilled in the art should understand that the scope of disclosure involved in the present application is not limited to the technical solution formed by a specific combination of the above technical features, but should also cover other technical solutions formed by any combination of the above technical features or their equivalent features without departing from the aforementioned disclosed concept. For example, the above features are replaced with the technical features with similar functions disclosed in the present application (but not limited to) by each other to form a technical solution.

Claims

1. A bioinformatics model training method based on bioinformatics data, characterized in that: include: Training the initial biometric information model by using a first data set of a first user in the user set to generate a first biometric information model, wherein the first data set is local data of the first user, and the local data of the first user includes first biometric information data; screening a second user in the user set, wherein the first user and the second user are different users in the user set; Training the first biological information model through a second data set of the second user to generate a second biological information model, and obtaining a convergence value of the second biological information model, wherein the second data set is local data of the second user, and the local data of the second user includes second biological information data; When the convergence value of the second biological information model is less than the first preset threshold, the second biological information model is confirmed to be the target biological information model, and the target biological information model is a trained model; wherein the convergence value is used to characterize the change range of the second biological information model compared with the first biological information model; The screening of the second user in the user set includes: Determine a neighboring user set of a first user in the user set, and send first request information to each neighboring user in the neighboring user set, where the first request information is used to request each neighboring user to send a current user state parameter; receiving first response information sent by neighboring users in the neighboring user set, and confirming a sending duration of the first response information sent by each neighboring user according to a time point at which the first response information sent by each neighboring user is received; According to a preset weighted value, weighted average is performed on the sending time length of the first response information sent by each adjacent user and the current user state parameter of each adjacent user, and the weighted average value is determined, where the current user state parameter includes: a computing power state parameter and an idle state parameter; According to the weighted average value, a target user in the adjacent user set is confirmed as the second user.

2. The method according to claim 1, characterized in that The step of training the initial biological information model by using the first data set of the first user in the user set to generate the first biological information model includes: The initial bioinformatics model is trained sequentially through the m first data subsets in the first data set to obtain m training results, where one first data subset corresponds to one training result, and m is a positive integer; Confirming a training result that is less than or equal to a second preset threshold among the m training results as a target training result, where the second preset threshold is used to indicate a reliable training result; According to the target training result, the initial biological information model is updated to generate the first biological information model.

3. The method according to claim 2, characterized in that The updating of the initial biological information model according to the target training result to generate a first biological information model includes: According to the target training result, the initial biological information model is updated using a preset averaging algorithm to generate a first biological information model.

4. The method according to claim 1, characterized in that: The first biological information model is trained using the second data set of the second user to generate a second biological information model, and a convergence value of the second biological information model is obtained, including: The first biological information model is trained sequentially using t second data subsets in the second data set to obtain t training results, where one second data subset corresponds to one training result, and t is a positive integer; Confirming a training result that is less than or equal to a second preset threshold among the t training results as a target training result, where the second preset threshold is used to indicate a reliable training result; According to the target training result, the current stochastic gradient and the current divergence, the first bio-information model is updated to generate a second bio-information model, wherein the current stochastic gradient is used to characterize the change gradient between the first bio-information model and the target training result, and the current divergence is used to characterize the user divergence relationship between the first user and the second user.

5. The method according to claim 4, characterized in that Before updating the first biological information model according to the target training result, the current stochastic gradient and the current divergence to generate the second biological information model, the method further includes: Confirming a target first data subset corresponding to the target training result; According to the target first data subset and a preset random function, a current random gradient corresponding to the second user is obtained.

6. The method according to claim 4, characterized in that The updating of the first biological information model according to the target training result, the current stochastic gradient and the current divergence to generate the second biological information model includes: Obtaining a current divergence between the first user and the second user by using a preset divergence function; According to the target training result, the current stochastic gradient and the current divergence, the first biological information model is updated to generate a second biological information model.

7. The method according to claim 4, characterized in that The updating of the first biological information model according to the target training result and the current stochastic gradient to generate a second biological information model includes: According to the target training result and the current stochastic gradient, the initial biological information model is updated using a preset averaging algorithm to generate a second biological information model.

8. A biological information model training device based on biological information data, characterized in that: include: A training module, configured to train an initial biometric information model using a first data set of a first user in a user set to generate a first biometric information model, wherein the first data set is local data of the first user, and the local data of the first user includes first biometric information data; A screening module, used for screening a second user in the user set, wherein the first user and the second user are different users in the user set; The training module is further used to train the first biological information model through a second data set of the second user to generate a second biological information model and obtain a convergence value of the second biological information model, wherein the second data set is local data of the second user, and the local data of the second user includes second biological information data; an execution module, configured to confirm that the second biological information model is a target biological information model when the convergence value of the second biological information model is less than a first preset threshold value, and the target biological information model is a trained model; wherein the convergence value is used to characterize the change range of the second biological information model compared with the first biological information model; The screening module is specifically used for: Determine a neighboring user set of a first user in the user set, and send first request information to each neighboring user in the neighboring user set, where the first request information is used to request each neighboring user to send a current user state parameter; receiving first response information sent by neighboring users in the neighboring user set, and confirming a sending duration of the first response information sent by each neighboring user according to a time point at which the first response information sent by each neighboring user is received; According to a preset weighted value, weighted average is performed on the sending time length of the first response information sent by each adjacent user and the current user state parameter of each adjacent user, and the weighted average value is determined, where the current user state parameter includes: a computing power state parameter and an idle state parameter; According to the weighted average value, a target user in the adjacent user set is confirmed as the second user.

9. A computer device comprising a memory, a processor and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, the method according to any one of claims 1 to 7 is implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the method according to any one of claims 1 to 7 is implemented.

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

Images