Data processing method, system and equipment based on biological information data and medium

Abstract

The invention discloses a data processing method, system and device based on biological information data and a medium, and relates to the technical field of computers. The method comprises the following steps: under the condition that each user in a plurality of users carries out training on an initial biological information model through local data to generate a first biological information training model, obtaining first data feature information of the local data of each user and the first biological information training model of each user, the first data feature information is used for indicating user node feature information corresponding to each user; and according to the first data feature information, determining a first contribution degree ratio of a first biological information training model corresponding to each user in the plurality of users through a preset deep network model, and aggregating the plurality of first biological information training models according to the first contribution degree ratio to generate a first biological information aggregation model. The training efficiency of model training can be improved by improving data contribution and strength to a certain extent.

Description

Technical Field

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

[0002] With the continuous development of communication technology and computer technology, helping users process data by training models has become a very common technical means and has been applied in various fields, such as finance and bioinformatics.

[0003] There are many ways to learn and train models, among which multi-node federated learning is a very common way to learn and train models. In the relevant technology, if in the field of bioinformatics, it is assumed that the bioinformatics data that needs to be learned and trained is distributed in multiple nodes and involves privacy protection, the initial model can be sent from the sender (such as the central server) to each learning node through the learning and training method of federated learning. Each learning node uses local bioinformatics data to train the initial model. After the training is completed, the training model is obtained and sent back to the sender. The sender then converges the training models of each node through the converger. In this way, each node can complete the learning and training without outputting its own data.

[0004] However, in the above-mentioned learning and training process, when the training models of each node are converged to form an aggregate model, the model contribution will be considered according to the amount of bioinformatics data used for the training model of each node, and then the weight ratio of each training model will be divided in the aggregation process. However, this division method is often relatively simple and crude, and cannot reflect the actual contribution of the data. Summary of the invention

[0005] In view of the above-mentioned defects or deficiencies in the relevant technology, it is hoped to provide a data processing method, system, device and medium based on bioinformatics data, which can solve the problem of dividing the weight proportion of each training model during the aggregation process. However, this division method is often relatively simple and crude, and cannot reflect the actual contribution of the data. To a certain extent, the training efficiency of the model training can be improved by improving the data contribution and strength.

[0006] In a first aspect, a data processing method is provided, the method comprising: In the case where each user among a plurality of users trains an initial biometric information model through local data to generate a first biometric information training model, obtaining first data feature information of the local data of each user and the first biometric information training model of each user, wherein the first data feature information is used to indicate user node feature information corresponding to each user, and the local data is data including biometric information; According to the first data feature information, a first contribution ratio of a first bio-information training model corresponding to each of the multiple users is determined through a preset deep network model, and the multiple first bio-information training models are aggregated according to the first contribution ratio to generate a first bio-information aggregation model.

[0007] In the present application, when each user among multiple users trains the initial bioinformation model through local data to generate a first bioinformation training model, the first data feature information of each user's local data and the first bioinformation training model of each user are obtained (the first data feature information is used to indicate the user node feature information corresponding to each user); then, according to the first data feature information, the first contribution ratio of the first bioinformation training model corresponding to each user among multiple users is determined through a preset deep network model, and multiple first bioinformation training models are aggregated according to the first contribution ratio to generate a first bioinformation aggregation model. In this way, by obtaining the data features of each user's user node and the contribution ratio between multiple user nodes, in the process of the final aggregation model, a suitable aggregation ratio is set for each user's first bioinformation training model according to the contribution ratio between multiple user nodes, so that the rationality of the final aggregated first bioinformation aggregation model is greatly improved compared with the previous aggregation model with uncertain contribution, and the training efficiency of the required model put into use by aggregating multiple user node training models is also greatly improved.

[0008] In a second aspect, a data processing system based on bioinformatics data is provided, the system comprising: an acquisition unit, configured to acquire first data feature information of the local data of each user and the first biometric information training model of each user when each user among a plurality of users trains the initial biometric information model through local data to generate a first biometric information training model, wherein the first data feature information is used to indicate user node feature information corresponding to each user, and the local data is data including biometric information; An execution unit is used to determine, according to the first data feature information, a first contribution ratio of a first bio-information training model corresponding to each of the multiple users through a preset deep network model, and aggregate the multiple first bio-information training models according to the first contribution ratio to generate a first bio-information aggregation model.

[0009] 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.

[0010] 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.

[0011] 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.

[0012] 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

[0013] 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: Figure 1 A schematic diagram of a data processing method based on bioinformatics data provided in an embodiment of the present application; Figure 2 A schematic diagram of the structure of a data processing system based on bioinformatics data provided in an embodiment of the present application; Figure 3 A schematic diagram of the structure of a computer device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0014] The present application is 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 invention, rather than to limit the invention. It is also necessary to explain that, for ease of description, only the parts related to the invention are shown in the accompanying drawings.

[0015] 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 implementation scenario of the present application will be described below.

[0016] With the continuous development of communication technology and computer technology, helping users process data by training models has become a very common technical means.

[0017] There are many ways to learn and train models, among which multi-node federated learning is a very common way to learn and train models. In related technologies, it is assumed that the data to be learned and trained is distributed in multiple nodes and involves privacy protection. That is, the initial model can be sent from the sender (such as the central server) to each learning node through the learning and training method of federated learning. Each learning node uses local data to train the initial model. After the training is completed, the training model is obtained and sent back to the sender. The sender then converges the training models of each node through the converger. In this way, each node can complete the learning and training without outputting its own data.

[0018] The following is an explanation based on specific industries, taking the medical industry as an example: In the medical industry, medical federated learning is an emerging machine learning method that allows multiple institutions to share models rather than raw data while protecting data privacy. In the medical field, different medical institutions have a large amount of medical data, but due to issues such as privacy and regulations, data is difficult to share directly. Federated learning provides a solution that allows various institutions to jointly train models without exposing raw data, thereby improving the performance and generalization ability of the models.

[0019] Contribution evaluation is an important part of medical federated learning, which aims to quantify and measure the contributions of various participants or stakeholders in data sharing and collaborative training. It aims to quantify and measure the contributions of various participants or stakeholders in data sharing and collaborative training. Accurate contribution evaluation not only helps to understand the role played by each participant in the entire model training process, but also provides a basis for decisions such as reward mechanisms, resource allocation, and partner selection.

[0020] This evaluation not only helps us understand the role of each participant in the entire model training process, but also provides a basis for decisions such as reward mechanisms, resource allocation, and partner selection.

[0021] In order to fully evaluate the contribution of medical federated learning, we need to focus on multiple indicators, such as the quantity and quality of data provided by participants, the degree of participation in training tasks, and the improvement of model performance. In addition, factors such as resource investment, participation time, and degree of collaboration should also be taken into consideration. By using these evaluation techniques, individuals can be effectively motivated to participate in data sharing and collaborative training, enhance cooperation among all parties, and further promote the development of machine learning in medical research and applications.

[0022] The contribution of medical federated learning can be evaluated from three dimensions: data, model performance, and participant behavior: 1) Data-oriented contribution: Before participating in medical federated learning, participants must evaluate and check the quality of their local data. This can be done through self-assessment or by following standard procedures to collect statistical information on the quality of medical data. The assessment of data quality involves multiple factors, including data completeness, accuracy, and label consistency. Before each training, indicators such as data completeness, accuracy, and label consistency can be used to assess the quality of local data.

[0023] 2) Behavior-oriented contributions: In medical federated learning, the behavior and degree of cooperation of participants have an important impact on the performance of the overall model. In particular, this includes the timeliness of participants' submission of model parameters, the degree of compliance with the protocol, and the degree of their participation in model evaluation.

[0024] 3) Performance-oriented contribution: Model performance can be used as an important indicator when the central server evaluates the contribution of participants in the training process. The efforts of participants can be measured by the performance of their trained models on the validation set. This includes indicators such as accuracy, recall, and F1 score for evaluating model performance.

[0025] Taking the above factors into account can more accurately measure the contribution of each participant to the MFL. This helps to adjust the global model to better reflect the data quality and cooperation level of each participant, as well as the performance of the model. These considerations are critical to ensuring the effectiveness of federated medical learning and the credibility of the results. Despite the importance of contribution assessment, there are still many shortcomings in research in this field due to the heterogeneity and incompleteness of medical data. Different medical institutions have different data collection methods and equipment, which affects the consistency and completeness of the data. This makes data integration and standardization difficult, leading to more complex contribution assessment studies.

[0026] Specifically, there have been some relevant research and practices in the evaluation of the contribution of medical federated learning.

[0027] Some methods evaluate mainly from the perspective of data. For example, the contribution is measured by counting the amount of data provided by the participants. This method is simple and intuitive, but it ignores the differences in data quality. High-quality data may have a much greater effect on model training than a large amount of low-quality data. Other methods consider data quality indicators such as data completeness, accuracy, and label consistency. The quality score of the data is determined by evaluating these indicators, and then the contribution of the participants is measured. However, this method often only focuses on the data itself, without considering other behaviors of the participants in the model training process and their impact on model performance.

[0028] Other methods start from the perspective of model performance. The contribution of participants is measured by comparing the performance of the models trained by participants on the validation set, such as accuracy, recall, and F1 score. This method can reflect the contribution of participants to the improvement of model performance, but it cannot reflect the efforts of participants in other aspects of data provision and collaboration.

[0029] Some studies have attempted to consider multiple factors in a comprehensive manner. For example, they have considered the quantity and quality of data, as well as the degree of participants' participation in the training task. However, these methods often lack systematicity and comprehensiveness, and there is no unified framework to comprehensively evaluate all relevant factors. Moreover, there is no reasonable basis for the weight allocation between different factors, which may lead to biased evaluation results.

[0030] However, these existing solutions have certain limitations. They do not comprehensively and systematically consider multi-dimensional indicators such as medical data quality, communication efficiency, participation and local performance, nor do they use advanced deep learning and reinforcement learning techniques to build an adaptive evaluation framework.

[0031] Based on this, this application proposes a data processing method, system, device and medium, which can solve the problem that existing solutions have certain limitations, do not comprehensively and systematically consider multi-dimensional indicators such as medical data quality, communication efficiency, participation and local performance, and do not use advanced deep learning and reinforcement learning technologies to build an adaptive evaluation framework.

[0032] Figure 1 1 is a flow chart of a data processing method based on bioinformatics data provided by an embodiment of the present application. The execution subject of the method may be a server, which is used to aggregate training models of multiple user nodes. Figure 1 As shown, the method includes the following steps 301 to 302: Step 301: When each user among a plurality of users trains an initial biometric information model using local data to generate a first biometric information training model, first data feature information of the local data of each user and the first biometric information training model of each user are obtained.

[0033] In the embodiment of the present application, the above-mentioned local data is data including biometric information.

[0034] In an embodiment of the present application, the first data characteristic information is used to indicate the user node characteristic information corresponding to each user.

[0035] 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.

[0036] Furthermore, since the learning method of the present application can be a federated learning method, that is, the initial bioinformatics model is sent to the user node of each user among multiple users for training to form a first bioinformatics training model, therefore, in the process of each user among multiple users training the initial bioinformatics model, the local data of each user is also used.

[0037] At the same time, the first bio-information training model generated after one training and the first bio-information training models of multiple users are aggregated to generate the first bio-information aggregation model, which is not necessarily the final model that can be put into use. Therefore, after the first bio-information aggregation model is generated, it may be necessary to continue training the first bio-information aggregation model through local data of each of the multiple users. Please refer to the subsequent content for details and will not be repeated here.

[0038] In the embodiment of the present application, the process of obtaining the first data feature information of each user's local data is determined by the user node of each user itself and sent to the server for aggregating the first biometric information training model for acquisition. That is, the first data feature information is extracted from the local data by the user node of each user among the multiple users.

[0039] Optionally, in an embodiment of the present application, when each user among multiple users trains an initial bio-information model through local data to generate a first bio-information training model, first data feature information of the local data of each user is obtained, and the data processing method includes: sending the initial bio-information model and the first information to each user among the multiple users; receiving the first data feature information of the local data obtained by each user using the autoencoder and the first bio-information training model of each user.

[0040] Exemplarily, the first information is used to instruct each of the users to perform data feature extraction on local data.

[0041] It can be understood that the characteristic of federated learning is that each user's user node is trained locally using local data, that is, its data will not be transmitted to other servers through the network. Correspondingly, the first data feature information of each user's local data also needs to be extracted by the user node server itself, and then transmitted to the aggregation server used to aggregate multiple user nodes, so that the aggregation server obtains the first data feature information.

[0042] In the process of each user training and extracting the first data feature information locally, the process of extracting the first data feature information of the user can be performed according to the above first information.

[0043] Exemplarily, for multiple users, the process of extracting their own first data feature information can be extracted by all local autoencoders.

[0044] It is understandable that this method of extracting the first data feature information through an autoencoder belongs to an adaptive evaluation framework. The adaptive evaluation framework integrating deep learning and reinforcement learning proposed in the embodiment of the present application has strong adaptability. By extracting different features through a variety of neural networks and constructing a deep contribution model, it can adapt to different data distributions and learning environments. For example, in different medical centers, the characteristics of the data may be different. The framework of the present invention can effectively extract and process these features through networks such as autoencoders and CNNs to ensure that the evaluation model can operate normally in various situations. For different model training tasks and participant behavior patterns, the adaptive evaluation framework can also be adjusted through a dynamic weight adjustment mechanism. For example, if the behavior of a participant changes during the training process, such as an increase in participation frequency or an improvement in model performance, the framework can adjust the weight configuration in time to adapt to this change and improve the accuracy of the evaluation.

[0045] Optionally, in an embodiment of the present application, the above-mentioned first data characteristic information includes at least one of the following data characteristic information: first data quality characteristic information, first data communication characteristic information, first participation frequency characteristic information, and first user node local performance characteristic information.

[0046] The following describes the process of extracting the above four different data feature information through the autoencoder in the user node of each user among multiple users.

[0047] 1) First data quality characteristic information Exemplarily, the above data quality feature information belongs to the complex features of the data, specifically, it includes data integrity, label correctness, etc. Specifically, the feature vector for indicating the first data quality feature information is obtained by the following formula 1: Formula 1 Among them, in the above formula 1, is the data quality feature vector of user node i of each user. represents the data input of node i, and is the weight matrix and bias vector of the autoencoder, obtained through the initial training phase; ReLU is a commonly used activation function used to introduce nonlinear characteristics.

[0048] It can be understood that in addition to the autoencoder, the principal component analysis (PCA) can also be used to extract medical data quality features in the above-mentioned first data quality feature information extraction process. PCA can project high-dimensional data into a low-dimensional space by performing a linear transformation on the original data while retaining the main variance information of the data. By analyzing the projected feature vectors, information related to data integrity, accuracy, and label consistency can be obtained. Although the extracted features may not be as complex as the autoencoder, they can also reflect the data quality to a certain extent. In addition, independent component analysis (ICA) can also be used for this purpose. ICA assumes that the data is composed of a number of independent components. By finding these independent components, hidden features in the data can be mined, which may be related to certain aspects of data quality.

[0049] 2) First data communication characteristic information Exemplarily, the first data communication characteristic information can be obtained through a convolutional neural network (CNN). Specifically, the first data communication characteristic information includes characteristics of communication bandwidth and communication delay. Specifically, the convolutional neural network (CNN) can be represented by the following formula 2, that is, formula 2 can be used to represent the first data communication characteristic information Formula 2 Among them, in the above formula 2, is the communication eigenvector, and They are the bandwidth and latency of node i, which can be calculated by the network connection status monitoring system between the node and the server.

[0050] It can be understood that the process of extracting the communication feature information of the first data, that is, the process of extracting the communication bandwidth and delay features, can use wavelet analysis in addition to convolutional neural network (CNN). Wavelet analysis can decompose the signal into sub-bands of different frequencies. By analyzing the characteristics of these sub-bands, relevant information about the communication bandwidth and delay can be obtained. For example, the energy distribution of different frequency sub-bands may be related to the communication bandwidth, and the phase relationship between the sub-bands may reflect the delay. Hilbert-Huang transform (HHT) is also an option. HHT can decompose non-stationary signals into intrinsic mode functions (IMFs), and communication features can be obtained by analyzing the characteristics of IMFs. It has unique advantages in processing nonlinear and non-stationary signals and is suitable for extracting communication features.

[0051] 3) First participation frequency characteristic information Exemplarily, the first participation frequency feature information can be analyzed and extracted through a long short-term memory network (LSTM). Specifically, the long short-term memory network (LSTM) can analyze the time series of participation frequency (referring to the frequency of each user's user node participating in the federated model aggregation within a certain period of time, how often they participate, the more frequent the more active, and the higher the intention to participate), and generate a participation feature vector for characterizing the first participation frequency feature information. Specifically, the long short-term memory network (LSTM) can be characterized by the following formula 3, that is, formula 3 can be used to extract the first participation frequency feature information: Formula 3 Among them, in formula 3, is the participating frequency eigenvector, Representation Node Participation Frequency, which is the participation frequency of the node in the past period of time, can be obtained through the statistics of the node's activity record data.

[0052] It is understandable that in the above-mentioned process of extracting the first participation frequency feature information, if you want to replace the long short-term memory network (LSTM) to analyze the time series of participation frequency, you can use the Markov chain. The Markov chain can construct a participation feature vector by analyzing the transition probability between the current participation state and the past participation state. It is based on a probability model and can reflect the dynamic change law of the participation frequency. Although it is different from the principle of LSTM, it can also achieve the purpose of extracting participation features. Hidden Markov model (HMM) is also feasible. HMM is a statistical model that adds hidden states to the Markov chain. Through the joint probability analysis of hidden states and observed states, participation features can be extracted more accurately.

[0053] 4) Local performance characteristic information of the first user node Exemplarily, the first participation frequency feature information can be extracted and evaluated by a multi-layer perceptron (MLP), and the first user node local performance feature information extracted by the multi-layer perceptron (MLP) can be represented by the following formula 4: Formula 4 Among them, in formula 3, Refer to the node The local performance eigenvector of Is a node LocalPerformance indicators, such as accuracy and loss value, which can usually be obtained from the historical training logs cached in the user node of each user.

[0054] Step 302: According to the first data feature information, a first contribution ratio of the first biometric information training model corresponding to each of the multiple users is determined through a preset deep network model, and the multiple first biometric information training models are aggregated according to the first contribution ratio to generate a first biometric information aggregation model.

[0055] In the embodiment of the present application, the above-mentioned preset deep network model (DNN) can be preset in the server or customized by the user, which is not limited in the embodiment of the present application.

[0056] In an embodiment of the present application, the preset deep network model may be set based on a deep neural network model (DNN), which may determine the first contribution ratio of each user based on the aforementioned first data feature information.

[0057] It is understandable that the above-mentioned preset deep network model can integrate various types of scores using a weighted synthesis method. This method can reasonably allocate weights according to the importance of indicators of different dimensions to ensure the fairness and reliability of the evaluation results. Compared with the lack of a reasonable basis for weight allocation in the prior art, the weighted biosynthesis method can avoid bias in evaluation results caused by unreasonable weights. For example, if there is no reasonable weight allocation, there may be an overemphasis on data quality and neglect of participation, or vice versa. The embodiment of the present application determines the weights scientifically so that each dimension can affect the final evaluation results in an appropriate proportion.

[0058] Furthermore, in addition to the weighted synthesis method, the median integration method can also be used. Sort the scores of each dimension by size, and take the median as the final contribution score. This method can avoid the influence of extreme values ​​and may be more suitable for integrated scoring in some cases. The mode integration method is also an option. If the scores of each dimension show a clear mode distribution, it may be reasonable to take the mode as the final contribution score. This method may have better effects when processing data with a central tendency. Although these alternatives exist, the technical solution adopted in the embodiment of the present application has unique advantages in comprehensively considering multi-dimensional indicators, dynamically adapting to the environment, and ensuring the fairness and accuracy of the evaluation results, and can better achieve the purpose of the invention.

[0059] Optionally, in an embodiment of the present application, when the above-mentioned first data feature information includes at least two types of data feature information, the above-mentioned first contribution ratio of each of the above-mentioned multiple users is determined based on the above-mentioned data feature information. The data processing method provided in the embodiment of the present application includes: inputting the first data feature information of each of the above-mentioned users into a preset deep network model to generate a first contribution parameter of each user; and determining the above-mentioned first contribution ratio between the above-mentioned multiple users through the above-mentioned preset deep network model based on the first contribution parameter of each of the above-mentioned users.

[0060] Exemplarily, the above-mentioned preset deep network model includes weight ratios of different data feature information.

[0061] In one example, combining the four data features (i.e., the at least two data feature information) obtained by the above formulas 1 to 4, the model structure of the preset deep network model can be as shown in the following formula 5: Formula 5 For example, after obtaining the above formula 5, the calculation of the first contribution ratio needs to go through the hidden layer The calculation formula of hidden layer l is as described in the following formula 6, and the process includes (input layer → hidden layer 1 → hidden layer 2 → … → hidden layer L → output layer):

[0062] ...... ...... Formula 6 In Formula 6, It is The output of the layer, and are the weight and bias of the layer respectively. Specifically, the number of measurement layers L of the hidden layer is actually determined according to the actual situation of the specific task. For example, large data sets usually require deeper networks (i.e., more hidden layers) for effective learning, and L can be determined by default.

[0065] The final model output contribution, that is, the first contribution ratio of each user's user node, can be represented by a contribution score, as shown in the following formula 7: Formula 7 in, Is a node Contribution score of and are the weights and biases of the output layer.

[0066] Furthermore, after obtaining the above contribution score, the contribution score can be further corrected by the model loss function to improve the accuracy of the actual contribution score. Specifically, the model loss function uses the mean square error (MSE) to measure the accuracy of the contribution prediction, as described in the following formula 8: Formula 8 in, It's a loss. is the number of nodes, It is the actual contribution score of the node.

[0067] It is understandable that, in addition to the method described in the above formulas 5-8, the evaluation method of the above-mentioned first contribution ratio can also be constructed using a Bayesian network. A Bayesian network is a graphical model based on probabilistic reasoning, which represents variables and the relationship between them through nodes and edges. Taking the various extracted features as nodes, and by learning the probabilistic relationship between them, a Bayesian network model can be constructed to evaluate the contribution. For example, data quality characteristics may affect model performance characteristics. This influence relationship can be clarified through a Bayesian network, and the contribution can be calculated accordingly. Bayesian networks can learn and update based on prior knowledge and data, and have certain adaptability and flexibility.

[0068] Furthermore, in addition to the above-mentioned preset deep network model, the above-mentioned aggregation model process can also use capsule networks or generative adversarial networks. Capsule Network is a new type of neural network structure that processes information through capsule units and dynamic routing algorithms. The extracted features can be input into the capsule network, and the contribution score can be output through the learning and calculation of the capsule network. The conditional generative adversarial network (cGAN) can generate outputs that meet the requirements under given conditions. The feature vector can be used as a condition to generate a contribution score or related evaluation indicators through cGAN.

[0069] In the method provided in the embodiment of the present application, when each user among multiple users trains the initial bioinformation model through local data to generate a first bioinformation training model, the first data feature information of each user's local data and the first bioinformation training model of each user are obtained (the first data feature information is used to indicate the user node feature information corresponding to each user); then, according to the first data feature information, the first contribution ratio of the first bioinformation training model corresponding to each user among multiple users is determined through a preset deep network model, and multiple first bioinformation training models are aggregated according to the first contribution ratio to generate a first bioinformation aggregation model. In this way, by obtaining the data features of the user node of each user and the contribution ratio between multiple user nodes, in the process of finally aggregating the model, a suitable aggregation ratio is set for the first bioinformation training model of each user according to the contribution ratio between multiple user nodes, so that the rationality of the first bioinformation aggregation model after the final aggregation is greatly improved compared with the previous aggregation model with uncertain contribution, and the training efficiency of the required model put into use by aggregating multiple user node training models is also greatly improved.

[0070] It is understandable that the method in the embodiment of the present application also has the following advantages: (I) Motivating participants. The fair and accurate contribution evaluation method in the embodiment of the present application can motivate participants. When participants know that their contributions can be accurately evaluated and recognized, they will be more motivated to participate in federated learning. For example, for institutions that provide high-quality data and actively participate in training, they can obtain reasonable contribution scores, which will encourage them to continue to maintain good behavior and will also attract more institutions to participate because they see a fair competition environment and a reasonable reward mechanism.

[0071] (II) Optimize resource allocation. By accurately evaluating the contribution of participants, resource allocation in federated learning can be optimized. For example, based on the contribution score, computing resources, data resources, etc. can be reasonably allocated so that resources can flow to institutions or participants who contribute more to model training, improve resource utilization efficiency, and promote better development of medical federated learning.

[0072] (III) Improving model performance. Since the embodiments of the present application can accurately evaluate the contribution of participants, the resources and advantages of all parties can be better integrated during the model training process. For example, based on the local performance indicators of the participants, the model training strategy can be adjusted in a targeted manner to improve the application of the overall model and promote the further development of federated learning in research and application.

[0073] In another embodiment of the present application, a specific implementation method of cyclic iterative training is also provided. Exemplarily, after the aforementioned "aggregating the above-mentioned multiple training models according to the first contribution ratio to generate a first biological information aggregation model", the data processing method provided in the embodiment of the present application also includes: obtaining the convergence value of the above-mentioned first biological information aggregation model; when the above-mentioned convergence value is less than a predetermined threshold, determining the above-mentioned first biological information aggregation model as the final output model.

[0074] Exemplarily, the convergence value of the first biological information aggregation model is a threshold used to verify whether the aggregation model meets the convergence condition.

[0075] It can be understood that after aggregating the first bio-information training models of multiple user nodes, the convergence value of the first bio-information aggregation model can be detected. When it is less than a predetermined threshold, that is, when the convergence condition is met, the first bio-information aggregation model can be determined as the final output, which can be used for actual use without further training.

[0076] Exemplarily, the predetermined threshold may be preset or user-defined.

[0077] Furthermore, the above-mentioned predetermined threshold value may also indicate: an accuracy threshold value and an F1 score threshold value. Accordingly, the convergence value may also represent the accuracy rate of the first biological information aggregation model and the F1 score of the first biological information aggregation model. In one example, if the accuracy of the first biological information aggregation model reaches or exceeds 0.95 (ie, 95%) (ie, the above-mentioned accuracy threshold), it is considered that the performance of the first biological information aggregation model meets the requirements, and the federated learning process ends.

[0078] In one example, if the F1 score of the first biological information aggregation model reaches or exceeds 0.95 (ie, the above-mentioned F1 score threshold), the model performance is considered to meet the requirements and the federated learning process is terminated.

[0079] In another embodiment of the present application, a specific implementation method for loop iterative training is also provided when the convergence value is greater than or equal to a predetermined threshold. Exemplarily, when the convergence value is greater than or equal to the predetermined threshold, after the aforementioned "aggregating the multiple training models according to the first contribution ratio to generate a first bio-information aggregation model", the data processing method provided in the embodiment of the present application also includes: sending the first bio-information aggregation model to the user nodes of the above-mentioned multiple users; when the user nodes of the above-mentioned multiple users train the above-mentioned first bio-information aggregation model to generate a second bio-information training model, obtaining the second data feature information of the second bio-information training model corresponding to each user through an autoencoder; according to the above-mentioned second data feature information, determining the second contribution ratio of each user in the above-mentioned multiple users through a preset deep network model, and aggregating the above-mentioned multiple second bio-information training models according to the second contribution ratio.

[0080] According to the foregoing content, it can be known that since the federated learning method in the present application sends the initial bioinformatics model to the user node of each user among multiple users for training once, reports its own data feature information and the training model after training the initial bioinformatics model, and performs the first aggregation to generate the first bioinformatics aggregation model, the first bioinformatics aggregation model needs to be trained again if the convergence value is greater than or equal to a predetermined threshold, that is, the first bioinformatics aggregation model is sent to the user nodes of multiple users again, and the user nodes of multiple users train the first bioinformatics aggregation model to generate a second bioinformatics training model and obtain the second data feature information through the autoencoder.

[0081] It can be understood that the process of obtaining the second data characteristic information is the same as the process of obtaining the first data characteristic information.

[0082] Exemplarily, the first data characteristic information and the second data characteristic information may be the same or different.

[0083] It can be understood that, when the user nodes of multiple users are trained using the first bio-information aggregation model, and the training data used is the same as the data for training the aforementioned initial bio-information model, the first data feature information and the above-mentioned second data feature information may be the same; when the user nodes of multiple users are trained using the first bio-information aggregation model, and the training data used is different from the data for training the aforementioned initial bio-information model, the first data feature information and the above-mentioned second data feature information may be different.

[0084] Exemplarily, the above-mentioned preset deep network model can refer to the description of the preset deep network model in the aforementioned content, which will not be repeated here.

[0085] Exemplarily, the method for acquiring the second data characteristic information may refer to the contents of Formula 1 to Formula 4 above, which will not be described in detail here.

[0086] Exemplarily, the process of obtaining the second contribution ratio may refer to the contents of Formula 5 to Formula 8, which will not be described in detail here.

[0087] In another embodiment of the present application, a specific implementation method for optimizing the first contribution ratio is also provided. Exemplarily, after the aforementioned "aggregating the above-mentioned multiple training models according to the first contribution ratio", the above-mentioned method also includes: updating the first contribution ratio of each user for optimization according to the above-mentioned first data feature information and time step through a preset deep Q network model to generate a second contribution ratio; obtaining the second bio-information training model of each user through training the first bio-information aggregation model through local data of each user among the multiple users to generate a second bio-information training model; aggregating the above-mentioned multiple second bio-information training models according to the above-mentioned second contribution ratio to generate a second bio-information aggregation model.

[0088] It can be understood that the above preset deep Q network model can be preset or user-defined.

[0089] Exemplarily, after the first bioinformatics training model is aggregated according to the first contribution ratio, the first contribution ratio may be further optimized so that the first contribution ratio is more in line with the data to be processed in the actual scenario, thereby further improving the training efficiency.

[0090] Furthermore, in order to dynamically adjust the weight of contribution evaluation in combination with each model aggregation, a reinforcement learning strategy (RL) is introduced to optimize the first contribution ratio. Reinforcement learning strategy (RL) is a machine learning method that learns the best strategy by having an agent perform actions in an environment and obtain rewards. Deep Q network (DQN) is an implementation of RL that approximates the Q value function through a neural network to select the best action in each state.

[0091] The dynamic weight adjustment mechanism combined with the preset deep Q network (DQN) is the key to improving adaptability and accuracy. The mechanism rewards according to the improvement of evaluation accuracy and dynamically adjusts the feature weight configuration. This dynamic adjustment can make the weight adjustment direction and amplitude play the best support role for the current contribution evaluation. If it is found that the influence of the features of a certain dimension on the contribution evaluation increases at a certain stage, the dynamic weight adjustment mechanism can increase the weight of this dimension in time to improve the accuracy of the evaluation. At the same time, the Q value update equation further helps optimize the action selection to maximize the overall evaluation benefit and ensure that the evaluation algorithm can remain efficient and accurate in a changing environment.

[0092] The initial feature weights are configured as:

[0093] Specifically, the initial feature weight configuration is: . ( ) At each time step, the weight It can be adjusted according to the strategy. Here, "time step" refers to an iterative step in the model training process. The optimization goal is to maximize the evaluation accuracy and reward.

[0094]

[0095] in, Indicates the current state. is the current action, is the learning rate (usually set to 0.0001), is the current reward, is the discount factor, and They are the next state and action respectively. This mechanism ensures the efficiency and stability of the evaluation process under various possible changing situations.

[0096] Specifically, each time step includes the following 4 steps: 1. Status : The weight of the current input feature .

[0097] 2. Action : Choose an action to adjust the current weights. Actions can be small changes to weights (such as increasing or decreasing the value of a feature weight): Specifically, a greedy action is selected with probability 1-ϵ: a=argmaxaQ(s,a), and an action is randomly selected with probability ϵ.

[0098] Set ϵ=0.5 Execute actions and update weights: Assuming that the selected action represents increasing the weight of feature 1, then:

[0099] in, is a small positive number representing a small adjustment in weights 3. Get rewards The reward is calculated based on the difference between the contribution score output by the model and the actual contribution. The smaller the difference, the greater the reward.

[0100] General settings: ∈[0,1] like Close to 1: Usually set to a value close to 1, indicating that the agent is more concerned about the long-term cumulative reward. For example, =0.9 or =0.99. like Close to 0: Indicates that the agent pays more attention to immediate rewards and ignores future rewards. Usually not used, but may be useful in some specific tasks.

[0101] Here is the setting =0.9 4. Next state : The new weight after adjustment .

[0102] For example: Assume that the weight of the current state s is [1,1,1,1], and the Q-value function is as follows: 1) Q(s, increase the weight of feature 1) = 0.8 2) Q(s, reduce the weight of feature 2) = 0.5 3) Q(s, increase the weight of feature 3) = 0.7 4) Q(s, reduce the weight of feature 4) = 0.6 Assuming a greedy strategy is used, the action with the highest Q value will be selected, that is, increasing the weight of feature 1: a=argmaxQ(s,a)=increase the weight of feature 1 Then, the weights are adjusted as follows:

[0103] Assumptions =0.1, then the updated weight is: [1,1,1,1,1].

[0104] In this way, through multiple iterations (that is, multiple time steps), the preset deep Q network model DQN will learn a strategy so that in each state Select Action After that, the cumulative reward can be maximized. This process will continue until the model reaches a satisfactory performance or completes the predetermined number of training rounds. For example, you can train for 50 rounds / 50 time steps.

[0105] It is understandable that in addition to the above-mentioned preset deep Q network model, a heuristic weight adjustment method can also be used. For example, some heuristic rules for weight adjustment are set based on experience, such as when the model performance is greatly improved, the weights of features related to model performance are increased accordingly. This method does not require complex learning algorithms, but relies on experience and understanding of the problem. Another heuristic method is to adjust weights based on the importance ranking of features. For example, by evaluating the importance of features of different dimensions (some statistical methods or expert evaluations can be used), weights are then assigned in order of importance, and fine-tuned according to actual conditions.

[0106] 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 1 The individual steps of the method are shown.

[0107] The present invention provides a computer program product, which includes instructions that are executed by a processor when the processor executes the instructions. Figure 1 The individual steps of the method are shown.

[0108] 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.

[0109] Figure 2 This is a block diagram of a data processing system based on biological information data according to an embodiment of the present application. The device can be deployed in a server for aggregating biological information training models. Figure 2 The device includes an acquisition unit 601 and an execution unit 602.

[0110] An acquisition unit 601 is used to acquire first data feature information of the local data of each user and the first biometric information training model of each user when each user among a plurality of users trains the initial biometric information model through local data to generate a first biometric information training model, wherein the first data feature information is used to indicate user node feature information corresponding to each user, and the local data is data including biometric information; The execution unit 602 is used to determine, according to the first data feature information, a first contribution ratio of the first biometric information training model corresponding to each of the multiple users through a preset deep network model, and aggregate the multiple first biometric information training models according to the first contribution ratio to generate a first biometric information aggregation model.

[0111] In one embodiment, the acquisition unit 601 is specifically used to send an initial biometric information model and first information to each of the multiple users, wherein the first information is used to instruct each user to perform data feature extraction on local data; and receive first data feature information of local data acquired by each user using an autoencoder and a first biometric information training model of each user.

[0112] In one embodiment, the first data characteristic information includes at least one of the following data characteristic information: first data quality characteristic information, first data communication characteristic information, first participation frequency characteristic information, and first user node local performance characteristic information.

[0113] In one embodiment, when the first data feature information includes at least two types of data feature information, the execution unit 602 is specifically used to: input the first data feature information of each user into a preset deep network model to generate a first contribution parameter for each user, wherein the preset deep network model includes a weight ratio of different data feature information; and determine the first contribution ratio between the multiple users through the preset deep network model according to the first contribution parameter of each user.

[0114] In one embodiment, the execution unit 602 is further configured to: obtain a convergence value of the first biological information aggregation model; and determine that the first biological information aggregation model is a final output model when the convergence value is less than a predetermined threshold.

[0115] In one embodiment, when the convergence value is greater than or equal to a predetermined threshold, the execution unit 602 is further used to: send the first bio-information aggregation model to the user nodes of the multiple users; when the user nodes of the multiple users train the first bio-information aggregation model to generate a second bio-information training model, obtain second data feature information of the second bio-information training model corresponding to each user through an autoencoder; determine the second contribution ratio of each user among the multiple users through a preset deep network model based on the second data feature information, and aggregate the multiple second bio-information training models based on the second contribution ratio.

[0116] In one embodiment, the execution unit 602 is further used to: update the first contribution ratio of each user for optimization through a preset deep Q network model according to the first data feature information and time step to generate a second contribution ratio; obtain the second bio-information training model of each user when each user among multiple users trains the first bio-information aggregation model through local data to generate a second bio-information training model; aggregate the multiple second bio-information training models according to the second contribution ratio to generate a second bio-information aggregation model.

[0117] It should be understood that the units recorded in the data processing system based on bioinformatics data correspond to the steps in the method described in the accompanying drawings. Therefore, the operations and features described above for the method are also applicable to the data processing system and the units contained therein, and will not be repeated here. The data processing system 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 data processing system can cooperate with the units in the computer device to implement the solution of the embodiment of the present application.

[0118] 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.

[0119] It should be noted that for details not disclosed in the data processing system of the embodiments of the present application, please refer to the details disclosed in the above embodiments of the present application, which will not be repeated here.

[0120] Reference below Figure 3 , Figure 3 A schematic diagram of the structure of a computer device suitable for implementing the embodiments of the present application is shown. Figure 3 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.

[0121] 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.

[0122] In particular, according to an embodiment of the present application, the above reference flow chart Figure 1 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.

[0123] 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.

[0124] 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.

[0125] 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.

[0126] 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 data processing method described in the present application.

[0127] 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 data processing method based on bioinformatics data, characterized in that: include: When each user among a plurality of users trains an initial biometric information model through local data to generate a first biometric information training model, obtaining first data feature information of the local data of each user and the first biometric information training model of each user, wherein the first data feature information is used to indicate user node feature information corresponding to each user, and the local data is data including biometric information; According to the first data feature information, a first contribution ratio of a first bio-information training model corresponding to each of the multiple users is determined through a preset deep network model, and the multiple first bio-information training models are aggregated according to the first contribution ratio to generate a first bio-information aggregation model.

2. The method according to claim 1, characterized in that When each user among a plurality of users trains an initial biometric information model using local data to generate a first biometric information training model, obtaining first data feature information of the local data of each user includes: Sending an initial biometric information model and first information to each of the multiple users, where the first information is used to instruct each user to perform data feature extraction on local data; Receive first data feature information of local data obtained by each user using an autoencoder and a first biometric information training model of each user.

3. The method according to claim 1, characterized in that The first data characteristic information includes at least one of the following data characteristic information: first data quality characteristic information, first data communication characteristic information, first participation frequency characteristic information, and first user node local performance characteristic information.

4. The method according to claim 1, characterized in that: In a case where the first data feature information includes at least two types of data feature information, determining a first contribution ratio of each of the multiple users according to the data feature information includes: Inputting the first data feature information of each user into a preset deep network model to generate a first contribution parameter of each user, wherein the preset deep network model includes a weight ratio of different data feature information; According to the first contribution parameter of each user, the first contribution ratio among the multiple users is determined by the preset deep network model.

5. The method according to claim 1, characterized in that After aggregating the plurality of first biological information training models according to the first contribution ratio to generate a first biological information aggregation model, the method further includes: Obtaining a convergence value of the first biological information aggregation model; When the convergence value is less than a predetermined threshold, the first biological information aggregation model is determined as the final output model.

6. The method according to claim 5, characterized in that When the convergence value is greater than or equal to a predetermined threshold, after aggregating the multiple training models according to the first contribution ratio to generate a first biological information aggregation model, the method further includes: Sending the first biological information aggregation model to the user nodes of the plurality of users; When the user nodes of the plurality of users train the first biological information aggregation model to generate a second biological information training model, obtaining second data feature information of the second biological information training model corresponding to each user through an autoencoder; According to the second data feature information, a second contribution ratio of each of the multiple users is determined by a preset deep network model, and the multiple second biometric information training models are aggregated according to the second contribution ratio.

7. The method according to claim 1, characterized in that After aggregating the multiple training models according to the first contribution ratio, the method further includes: According to the first data feature information and the time step, the first contribution ratio of each user is updated and optimized by using a preset deep Q network model to generate a second contribution ratio; In the case where each user among a plurality of users trains the first bio-information aggregation model using local data to generate a second bio-information training model, obtaining the second bio-information training model of each user; The plurality of second biological information training models are aggregated according to the second contribution ratio to generate a second biological information aggregated model.

8. A data processing system based on biological information data, characterized in that: include: an acquisition unit, configured to acquire first data feature information of the local data of each user and the first biometric information training model of each user when each user among a plurality of users trains the initial biometric information model through local data to generate a first biometric information training model, wherein the first data feature information is used to indicate user node feature information corresponding to each user, and the local data is data including biometric information; An execution unit is used to determine, according to the first data feature information, a first contribution ratio of a first bio-information training model corresponding to each of the multiple users through a preset deep network model, and aggregate the multiple first bio-information training models according to the first contribution ratio to generate a first bio-information aggregation model.

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