Event detection model federated training system and method based on aggregation strategy

By using aggregation strategy in the event detection model federated training system, the data quality and model performance contribution of each client are calculated and model parameters are aggregated, which solves the problem of low event detection performance caused by data dispersion and improves the accuracy of event detection and robustness.

CN119067197AInactive Publication Date: 2024-12-03BEIJING UNIV OF POSTS & TELECOMM

Patent Information

Application Number
CN202410975858.3
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-19
Publication Date
2024-12-03
Estimated Expiration
Not applicable · inactive patent

AI Technical Summary

Technical Problem

The traditional deep learning-based event detection method is unable to share data due to data dispersion and privacy protection, resulting in low event detection performance.

Method used

The event detection model federated training system is adopted based on the aggregation strategy, and the model is trained locally by the client and sending parameters to the server. The server calculates the aggregation weight based on the data quality and model performance of each client to aggregate the model parameters to form a global event detection model, and then It is updated to each client.

Benefits of technology

Without sharing data, using high-quality training information from multiple clients improves the performance of the global model, reduces the negative impact of low-quality data on the model, and improves the accuracy and robustness of event detection.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119067197A_ABST
    Figure CN119067197A_ABST
Patent Text Reader

Abstract

The invention provides an event detection model federation training system and method based on an aggregation strategy, a client uses a local training set to train a local event detection model to obtain a trained event detection model, and sends model parameters corresponding to the event detection model to a server; after receiving the model parameters sent by the clients, the server performs model parameter aggregation on the model parameters based on the aggregation weights corresponding to the clients to obtain global event detection model parameters, and sends the global event detection model parameters to the clients; each client receives the global event detection model parameters, updates the event detection model by using the global event detection model parameters and stores the event detection model; under the condition that the preset training round is not completed, each client executes the model training of the next round; the problem that one mechanism can only master a small amount of data, and different mechanisms cannot share the data, so that training data is lacked, and the performance of event detection is low can be solved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of computer technology, and in particular to a system and method for federal training of event detection models based on aggregation strategy. Background Art

[0002] Public safety emergencies refer to dangerous and urgent events that require early detection, rapid response, and rapid recovery. If public safety emergencies cannot be detected in a timely and accurate manner, they will have a serious impact on social order, the national economy, and people's lives and health. Therefore, it is of great significance to collect and analyze data related to public safety emergencies and detect public safety emergencies quickly and accurately.

[0003] Traditional event detection methods include: event detection methods based on deep learning, which realize the automation of event detection through the representation ability of neural network models and effective learning frameworks.

[0004] However, deep learning-based event detection methods rely on a large amount of high-quality data, while data on public safety emergencies are usually scattered among different agencies. One agency can only have a small amount of data. Due to the limitations of network privacy protection and data security management, different agencies cannot share data, resulting in a lack of training data, which leads to the problem of low event detection performance. Summary of the invention

[0005] In view of this, an embodiment of the present invention provides an event detection model federation training system and method based on an aggregation strategy to eliminate or improve one or more defects existing in the prior art.

[0006] One aspect of the present invention provides an event detection model federation training system based on aggregation strategy, the system includes a server and several clients;

[0007] The client is used to: use the local training set to perform the t-th round of training on the event detection model stored locally, obtain the event detection model after the t-th training round, and send the model parameters corresponding to the event detection model to the server; wherein the local training set includes the sample message graph, the sample label corresponding to the sample message graph, and the verification message graph; t is a positive integer ranging from 1 to T in sequence, and T is a preset number of training times;

[0008] The server is further configured to: upon receiving the model parameters sent by all the clients, aggregate the model parameters based on the aggregation weights corresponding to the respective clients to obtain global event detection model parameters; and send the global event detection model parameters to the respective clients;

[0009] The client is also used to: upon receiving the global event detection model parameters, use the global event detection model parameters to update the event detection model, obtain and store the updated event detection model; when t is less than T, set t=t+1, and execute the step of using the local training set to perform the tth round of training on the locally stored event detection model.

[0010] In some embodiments of the present invention, the aggregation weight is calculated based on the data quality contribution corresponding to the local training set and the model performance contribution corresponding to the event detection model; wherein the data quality contribution is positively correlated with the number of nodes, the number of edges and the number of sample labels in the sample message graph; the model performance contribution is negatively correlated with the loss value corresponding to the event detection model, and positively correlated with the standardized mutual information index corresponding to the event detection model.

[0011] In some embodiments of the present invention, the client is also used to: during the tth round of training, input the sample message graph into the event detection model to obtain training results; input the training results and sample labels into a preset loss function to obtain a loss value; based on the mutual information between the sample labels and the training results, obtain a standardized mutual information index; wherein the preset loss function is a loss function pre-constructed based on contrastive learning.

[0012] In some embodiments of the present invention, the client is further used to: send the data quality contribution, loss value and normalized mutual information index to the server;

[0013] The server is also used to: calculate the model performance contribution based on the loss value and the standardized mutual information index when receiving the loss value and the standardized mutual information index; calculate the aggregation weight corresponding to each client based on the model performance contribution and the data quality contribution.

[0014] In some embodiments of the present invention, the client is further used to: calculate the model performance contribution based on the loss value and the normalized mutual information index; send the data quality contribution and the model performance contribution to the server;

[0015] The server is further used to: upon receiving the data quality contribution and the model performance contribution, calculate the aggregation weight corresponding to each client based on the model performance contribution and the data quality contribution.

[0016] In some embodiments of the present invention, the client is further used to: when the sample training set is updated, calculate the current data quality contribution based on the updated sample training set; and send the current data quality contribution to the server;

[0017] The server is further configured to: upon receiving the current data quality contribution, update the data quality contribution based on the current data quality contribution.

[0018] In some embodiments of the present invention, before executing the step of performing the t-th round of training on the event detection model stored locally using the local training set, the client is also used to: determine whether the event detection model has converged, and if the event detection model has not converged, execute the step of performing the t-th round of training on the event detection model stored locally using the local training set.

[0019] In some embodiments of the present invention, the server is further used to: send the initial event detection model to each client;

[0020] The client is further configured to: upon receiving the initial event detection model, store the initial event detection model as the event detection model.

[0021] In some embodiments of the present invention, the event detection model includes a feature extractor and an event detector; the feature extractor is constructed based on a graph attention network and contrastive learning, and the event detector detects events based on a clustering method.

[0022] Another aspect of the present invention provides a method for federated training of an event detection model based on an aggregation strategy, which is applied to the above-mentioned event detection model federated training system based on an aggregation strategy, comprising:

[0023] The client uses the local training set to perform the t-th round of training on the event detection model stored locally, obtains the event detection model after the t-th training round, and sends the model parameters corresponding to the event detection model to the server; wherein the local training set includes a sample message graph, a sample label corresponding to the sample message graph, and a verification message graph; t is a positive integer ranging from 1 to T in sequence, and T is a preset number of training times;

[0024] The server receives the model parameters sent by the client, aggregates the model parameters based on the aggregation weights corresponding to each client, obtains the global event detection model parameters, and sends the global event detection model parameters to each client;

[0025] Receive global event detection model parameters through the client, use the global event detection model parameters to update the event detection model and store it; when t is less than T, set t=t+1, and execute the step of using the local training set to perform the tth round of training on the locally stored event detection model.

[0026] The present invention provides an event detection model federation training system and method based on an aggregation strategy. The client uses a local training set to train the event detection model stored locally to obtain the trained event detection model, and sends the model parameters corresponding to the event detection model to the server. After receiving the model parameters sent by the client, the server aggregates the model parameters based on the aggregation weights corresponding to each client to obtain the global event detection model parameters, and sends the global event detection model parameters to each client. Each client receives the global event detection model parameters, uses the global event detection model parameters to update the event detection model and stores it. In the case of not completing the preset training rounds, each client executes The next round of model training can be carried out; it can solve the problem that one organization can only master a small amount of data, and different organizations cannot share data, resulting in a lack of training data, and thus leading to low event detection performance; through the topological structure and other information of the sample training set of each client, more high-quality training information is utilized, and without sharing data, the global event detection model parameters can be close to the optimal model parameters, thereby improving the performance of the global model. Each client updates the local event detection model through the global event detection model parameters, which can improve the performance of the local event detection model; at the same time, it solves the impact of data heterogeneity of different clients on the performance of the global event detection model, and improves the accuracy of event detection.

[0027] In addition, the server calculates weights based on the data quality contribution and model performance contribution of each client, and aggregates the parameters of the global event detection model. By considering the data quality and model performance of each client, the weight of each client in the global event detection model can be adjusted. In this way, the negative impact of clients with low-quality data or poor performance on the global model can be reduced, thereby improving the robustness and performance of the model. At the same time, the aggregation weight of clients with high-quality data and new models is enhanced, so that the global event detection model can more effectively absorb and reflect these high-quality data, which can further improve the accuracy and generalization ability of the event detection model.

[0028] Additional advantages, purposes, and features of the present invention will be described in part in the following description, and will become apparent to those skilled in the art after studying the following, or may be learned from the practice of the present invention. The purposes and other advantages of the present invention may be achieved and obtained by the structures specifically indicated in the specification and the accompanying drawings.

[0029] Those skilled in the art will appreciate that the objectives and advantages that can be achieved with the present invention are not limited to the above specific description, and the above and other objectives that can be achieved by the present invention will be more clearly understood from the following detailed description. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] The drawings described herein are used to provide a further understanding of the present invention, constitute a part of the present application, and do not constitute a limitation of the present invention. In the drawings:

[0031] Figure 1 A schematic diagram of the structure of an event detection model federation training system based on aggregation strategy provided by an embodiment of the present invention.

[0032] Figure 2 A schematic diagram of the structure of an event detection model provided by an embodiment of the present invention.

[0033] Figure 3 A flowchart of a method for federated training of an event detection model based on an aggregation strategy provided in one embodiment of the present invention. DETAILED DESCRIPTION

[0034] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the embodiments and the accompanying drawings. Here, the illustrative embodiments of the present invention and their descriptions are used to explain the present invention, but are not intended to limit the present invention.

[0035] It should also be noted that, in order to avoid obscuring the present invention due to unnecessary details, only structures and / or processing steps closely related to the solutions according to the present invention are shown in the accompanying drawings, while other details that are not closely related to the present invention are omitted.

[0036] It should be emphasized that the term “include / comprises” when used herein refers to the presence of features, elements, steps or components, but does not exclude the presence or addition of one or more other features, elements, steps or components.

[0037] It should also be noted that, unless otherwise specified, the term “connection” herein may refer not only to a direct connection but also to an indirect connection involving an intermediate.

[0038] Hereinafter, embodiments of the present invention will be described with reference to the accompanying drawings. In the accompanying drawings, the same reference numerals represent the same or similar components, or the same or similar steps.

[0039] This embodiment provides a schematic diagram of the structure of an event detection model federation training system based on an aggregation strategy.

[0040] like Figure 1 As shown, the event detection model federation training system based on the aggregation strategy includes: a server 110 and a plurality of clients 120. A wireless communication connection or a wired communication connection is established between the server 110 and each client 120.

[0041] The server 110 refers to an electronic device for receiving model parameters of event detection models trained from multiple clients 120 and aggregating and distributing global model parameters. The server 110 may be an electronic device such as a cloud server or a central server, and this embodiment does not limit the implementation of the server 110.

[0042] The client 120 refers to an electronic device capable of performing local event detection model training and model parameter transmission, including but not limited to a mobile phone or a tablet computer, etc. This embodiment does not limit the implementation of the client 120 .

[0043] In some embodiments of the present invention, the number of servers 110 is at least one, and this embodiment does not limit the number of servers 110; the number of clients 120 is at least one, and this embodiment does not limit the number of clients 120.

[0044] Before each client 120 trains a local event detection model, each client 120 needs to first obtain an initial event detection model, and the initial event detection models obtained by different clients 120 are the same.

[0045] In some embodiments of the present invention, the server 110 sends a unified initial event detection model to each client 120. After receiving the initial event detection model uniformly sent by the server 110, each client 120 stores the initial event detection model in a local storage medium as a local event detection model, so that each client 120 uses a local training set to perform T rounds of model training on its own event detection model. Wherein, T is a preset number of training times, which is a positive integer greater than 1.

[0046] Specifically, the server 110 is used to send the initial event detection model to each client. The client 120 is used to store the initial event detection model as the event detection model when receiving the initial event detection model.

[0047] In actual implementation, the event detection model in each client 120 can also be pre-stored in the local storage medium of each client 120, or obtained by each client 120 from an external storage medium (such as a USB flash drive or a CD, etc.). This embodiment does not limit the way in which each client 120 obtains the event detection model.

[0048] In some embodiments of the present invention, any type of event is input into the event detection model in the form of a graph structure, and any type of event is detected by the event detection model, such as public safety emergencies or natural disaster events. Based on this, in order to better utilize the topological structure information and semantic information of the event to achieve more accurate event detection, an initial event detection model is constructed based on the Graph Attention Network (GAT) and contrastive learning as the local event detection model of each client 120. Among them, the graph attention network is a deep learning model for processing graph data, which is suitable for tasks such as node classification, graph classification and graph representation learning. Contrastive learning is a self-supervised learning method that learns feature representations by letting the model compare different parts of the data. The core idea of ​​contrastive learning is to compare positive samples from the same data sample with negative samples from different data samples to prompt the model to learn meaningful features to distinguish these samples. The network structure of the event detection model is as follows: Figure 2 As shown, the event detection model includes a feature extractor and an event detector.

[0049] Among them, the feature extractor is constructed based on the graph attention network and contrastive learning to fully extract the information in the message graph corresponding to the event. In some embodiments of the present invention, the encoder in the graph attention network generates an updated node feature representation by processing the input features and the graph structure. The encoder based on the graph attention network can be expressed by the following formula:

[0050] E:H N×d ×{0,1} N×N →H N×d′

[0051] Where E represents the encoder based on the graph attention network; H N×d Represents the feature matrix of the input graph structure, N represents the number of nodes in the message graph, and d represents the dimension of each node feature; {0,1} N×N The adjacency matrix representing the graph structure is a matrix with N rows and N columns, where an element value of 1 indicates that there is a connection between nodes, and a value of 0 indicates that there is no connection; H N×d′ represents the output feature matrix of the encoder, which has a similar structure to the input feature matrix, where N is the number of nodes and d′ is the new feature dimension of each node.

[0052] In the graph attention network, the graph structure is processed through a multi-head attention mechanism at each layer, and each head can focus on different local structures in the graph structure, thereby capturing the complex relationships in the graph structure.

[0053] In some embodiments of the present invention, the propagation formula for each layer is as follows:

[0054]

[0055] In the formula, represents the representation of node i at layer l; || represents the connection operation; K represents the number of heads; is the attention score of node i and node i in layer l-1; W l-1 is the trainable parameter of the l-1 layer; represents the representation of node i at layer l-1; is the attention score of node i and node j in layer l-1.

[0056] In some embodiments of the present invention, in order to better distinguish events, contrastive learning is used to construct a preset loss function. i , sample a positive message (with message m i news from the same type of event) and a negative news (with message m i messages from different types of events), forming a triple

[0057] The contrastive learning loss function converts positive information Push message m i , negative news Push message i , specifically, the preset loss function is expressed by the following formula:

[0058]

[0059] Where d(·) represents the Euclidean distance between two vectors; m i Indicates the message corresponding to the event; Represents and message m i There is news about the same type of events; Represents and message m i messages from different types of events; ReLU represents the ReLU activation function; T is a set of online sampling that satisfies Such triplets can converge faster.

[0060] In some embodiments of the present invention, the event detector detects events based on a clustering method. The clustering method includes but is not limited to K-means clustering, hierarchical clustering, or density clustering. Figure 2As shown, in the event detection stage, clustering is performed based on the learned message representations.

[0061] In some embodiments of the present invention, after acquiring the initial event detection model, each client 120 performs T rounds of training on the event detection model. In each round of training, each client 120 uses a local training set stored locally to train the event detection model, and after the training is completed, the parameters of the event detection model obtained by the training are sent to the server 110. The local training set includes a sample message graph, a sample label corresponding to the sample message graph, and a verification message graph.

[0062] Specifically, the client 120 is used to: use the local training set to perform the tth round of training on the event detection model stored locally, obtain the event detection model after the tth training round, and send the model parameters corresponding to the event detection model to the server; wherein t is a positive integer ranging from 1 to T in sequence.

[0063] The server 110 is also used to: upon receiving the model parameters sent by all the clients 120 , aggregate the model parameters based on the aggregation weights corresponding to each client 120 to obtain global event detection model parameters; and send the global event detection model parameters to each client 120 .

[0064] Specifically, the aggregation formula of server 110 is as follows:

[0065]

[0066] Among them, W t represents the global event detection model parameter; M represents the number of clients 120; represents the model parameters sent by the mth client 120; W t represents the global event detection model parameters, represents the aggregation weight corresponding to the m-th client 120.

[0067] In some embodiments of the present invention, the aggregation weight is calculated based on the data quality contribution corresponding to the local training set and the model performance contribution corresponding to the event detection model, and the aggregation weights corresponding to different clients 120 are the same or different.

[0068] The data quality contribution is positively correlated with the number of nodes, edges, and sample labels in the sample message graph. In the sample message graph corresponding to the client 120, the more nodes, the more edges between nodes, and the more sample label categories, the richer the topological information of the sample message graph, and the higher the data quality. Specifically, the data quality contribution is expressed by the following formula:

[0069]

[0070] Where D m represents the data quality contribution corresponding to the mth client 120; M represents the number of clients 120; c m represents the number of nodes corresponding to the message graph of the mth client 120 data sample, e m represents the number of edges corresponding to the sample message graph of the mth client 120, l m Indicates the number of sample labels corresponding to the sample message graph of the mth client 120.

[0071] The model performance contribution is negatively correlated with the loss value corresponding to the event detection model, and positively correlated with the Normalized Mutual Information (NMI) indicator corresponding to the event detection model.

[0072] Specifically, the model performance contribution is expressed by the following formula:

[0073]

[0074] In the formula, represents the model performance contribution corresponding to the m-th client 120; represents the loss value corresponding to the mth client 120 during the tth round of training; represents the normalized mutual information index corresponding to the m-th client 120 in the t-th round of federated training.

[0075] In some embodiments of the present invention, during each round of training, the client 120 inputs the sample message graph into the event detection model obtained by training to obtain the training results corresponding to the current training round, and inputs the training results and sample labels into a preset loss function to obtain a loss value; the client 120 calculates the mutual information (Mutual Information) between the sample label and the predicted label of the training result, as well as the entropy (Entropy) of the sample label and the predicted label, and uses the mutual information divided by the product of the entropy to normalize and obtain a standardized mutual information index.

[0076] Specifically, the client 120 is also used to: during the tth round of training, input the sample message graph into the event detection model to obtain the training results; input the training results and the sample labels into the preset loss function to obtain the loss value; based on the mutual information between the sample labels and the training results, obtain the standardized mutual information index; wherein the preset loss function is a loss function pre-constructed based on contrastive learning.

[0077] In some embodiments of the present invention, after the data quality contribution, loss value and standardized mutual information index are calculated by the client 120, the data quality contribution, loss value and standardized mutual information index are sent to the server 110; the server 110 calculates the model performance contribution based on the received loss value and standardized mutual information index, and then calculates the aggregation weight corresponding to each client 120 based on the model performance contribution and the data quality contribution.

[0078] Specifically, the client 120 is further used to: send the data quality contribution, loss value and normalized mutual information index to the server 110. The server 110 is further used to: calculate the model performance contribution based on the loss value and the normalized mutual information index when receiving the loss value and the normalized mutual information index; calculate the aggregation weight corresponding to each client 120 based on the model performance contribution and the data quality contribution.

[0079] In some embodiments of the present invention, data quality contribution and model performance contribution are calculated by client 120, and the data quality contribution and model performance contribution are sent to server 110; server 110 calculates the aggregation weight corresponding to each client 120 based on the model performance contribution and data quality contribution.

[0080] Specifically, the client 120 is further used to: calculate the model performance contribution based on the loss value and the normalized mutual information index; and send the data quality contribution and the model performance contribution to the server 110. The server 110 is further used to: calculate the aggregation weight corresponding to each client 120 based on the model performance contribution and the data quality contribution when receiving the data quality contribution and the model performance contribution.

[0081] In actual implementation, the aggregation weight corresponding to each client 120 may also be sent to the server 110 after each client 120 completes the calculation. This embodiment does not limit the implementation method of the aggregation weight.

[0082] In addition, during the training process of the event detection model of the client 120, when new data that can reflect the latest trends of the events is collected, the client 120 is also used to update the sample training set according to the collected new data, and calculate the data quality contribution based on the updated sample training set, and send it to the server 110, so that the server 110 can update the aggregation weights corresponding to each client 120.

[0083] Specifically, the client 120 is further used to: when the sample training set is updated, calculate the current data quality contribution based on the updated sample training set; and send the current data quality contribution to the server 110. The server 110 is further used to: when receiving the current data quality contribution, update the data quality contribution based on the current data quality contribution.

[0084] In some embodiments of the present invention, after aggregating the global event detection model parameters, the server 110 sends the global event detection model parameters to each client 120, so that the client 120 updates the corresponding local event detection model according to the received global event detection model parameters and stores them, and performs the next round of training when the current training number has not reached the preset training number T.

[0085] Specifically, the client 120 is also used to: upon receiving the global event detection model parameters, use the global event detection model parameters to update the event detection model, obtain and store the updated event detection model; when t is less than T, set t=t+1, and execute the step of performing the tth round of training on the locally stored event detection model using the local training set.

[0086] When the current number of training times reaches the preset number of training times T, that is, when t is equal to T, the training of the local event detection model by each client 120 is completed.

[0087] In addition, after the event detection model is updated using the global event detection model parameters, it is also possible to judge whether the updated event detection model has converged. If the updated event detection model has converged, the training is completed; if the updated event detection model has not converged and the current number of training times reaches the preset number of training times T, continue with the next round of model training.

[0088] Specifically, before executing the step of using the local training set to perform the tth round of training on the event detection model stored locally, the client 120 is also used to: determine whether the event detection model has converged, and if the event detection model has not converged, execute the step of using the local training set to perform the tth round of training on the event detection model stored locally.

[0089] This embodiment provides a method for federated training of event detection models based on aggregation strategy. Figure 3 The flow chart shown is applied to the above-mentioned event detection model federation training system based on aggregation strategy. The event detection model federation training method based on aggregation strategy at least includes steps S301 to S303:

[0090] Step S301, the client uses the local training set to perform the t-th round of training on the event detection model stored locally, obtains the event detection model after the t-th training round, and sends the model parameters corresponding to the event detection model to the server.

[0091] The local training set includes a sample message graph, a sample label corresponding to the sample message graph, and a verification message graph; t is a positive integer ranging from 1 to T, and T is a preset number of training times.

[0092] Step S302: The server receives the model parameters sent by the client, aggregates the model parameters based on the aggregation weights corresponding to each client, obtains global event detection model parameters, and sends the global event detection model parameters to each client.

[0093] Step S303, receiving global event detection model parameters through the client, using the global event detection model parameters to update the event detection model and store it; when t is less than T, set t=t+1, and perform the step of using the local training set to train the locally stored event detection model for the tth round. When t is equal to T, the training of the event detection model of the client is completed.

[0094] The event detection model federation training method based on aggregation strategy provided in this embodiment refers to the above-mentioned event detection model federation training system embodiment based on aggregation strategy, which will not be repeated here.

[0095] In summary, the event detection model federation training system and method based on the aggregation strategy provided in this embodiment is that the client uses the local training set to train the event detection model stored locally to obtain the trained event detection model, and sends the model parameters corresponding to the event detection model to the server; after receiving the model parameters sent by the client, the server aggregates the model parameters based on the aggregation weights corresponding to each client to obtain the global event detection model parameters, and sends the global event detection model parameters to each client; each client receives the global event detection model parameters, uses the global event detection model parameters to update the event detection model and stores it; in the case of not completing the preset training rounds, each client The next round of model training is performed on the client side; it can solve the problem that an organization can only master a small amount of data, and different organizations cannot share data, resulting in a lack of training data, which leads to low event detection performance; through the topological structure and other information of the sample training set of each client, more high-quality training information is utilized. Without sharing data, the global event detection model parameters can be close to the optimal model parameters, thereby improving the performance of the global model. Each client updates the local event detection model through the global event detection model parameters, which can improve the performance of the local event detection model; at the same time, it solves the impact of data heterogeneity of different clients on the performance of the global event detection model and improves the accuracy of event detection.

[0096] In addition, the server calculates weights based on the data quality contribution and model performance contribution of each client, and aggregates the parameters of the global event detection model. By considering the data quality and model performance of each client, the weight of each client in the global event detection model can be adjusted. In this way, the negative impact of clients with low-quality data or poor performance on the global model can be reduced, thereby improving the robustness and performance of the model. At the same time, the aggregation weight of clients with high-quality data and new models is enhanced, so that the global event detection model can more effectively absorb and reflect these high-quality data, which can further improve the accuracy and generalization ability of the event detection model.

[0097] It should be understood by those skilled in the art that the exemplary components, systems and methods described in conjunction with the embodiments disclosed herein can be implemented in hardware, software or a combination of the two. Whether it is performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of the present invention. When implemented in hardware, it can be, for example, an electronic circuit, an application specific integrated circuit (ASIC), appropriate firmware, a plug-in, a function card, etc. When implemented in software, the elements of the present invention are programs or code segments used to perform the required tasks. The program or code segment can be stored in a machine-readable medium, or transmitted on a transmission medium or a communication link via a data signal carried in a carrier.

[0098] It should be clear that the present invention is not limited to the specific configuration and processing described above and shown in the figures. For the sake of simplicity, a detailed description of the known method is omitted here. In the above embodiments, several specific steps are described and shown as examples. However, the method process of the present invention is not limited to the specific steps described and shown, and those skilled in the art can make various changes, modifications and additions, or change the order between the steps after understanding the spirit of the present invention.

[0099] In the present invention, features described and / or illustrated for one embodiment may be used in the same or similar manner in one or more other embodiments, and / or combined with features of other embodiments or replace features of other embodiments.

[0100] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the embodiments of the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A federated training system for event detection models based on aggregation strategy, characterized in that: The system includes a server and several clients; The client is used to: use the local training set to perform the t-th round of training on the event detection model stored locally, obtain the event detection model after the t-th training round, and send the model parameters corresponding to the event detection model to the server; wherein the local training set includes a sample message graph, a sample label corresponding to the sample message graph, and a verification message graph; t is a positive integer ranging from 1 to T in sequence, and T is a preset number of training times; The server is further configured to: upon receiving the model parameters sent by all the clients, aggregate the model parameters based on the aggregation weights corresponding to the respective clients to obtain global event detection model parameters; and send the global event detection model parameters to the respective clients; The client is also used to: upon receiving the global event detection model parameters, use the global event detection model parameters to update the event detection model, obtain and store the updated event detection model; when t is less than T, set t=t+1, and execute the step of using the local training set to perform the tth round of training on the locally stored event detection model.

2. The system according to claim 1, characterized in that The aggregation weight is calculated based on the data quality contribution corresponding to the local training set and the model performance contribution corresponding to the event detection model; wherein the data quality contribution is positively correlated with the number of nodes, the number of edges and the number of sample labels in the sample message graph; the model performance contribution is negatively correlated with the loss value corresponding to the event detection model, and positively correlated with the standardized mutual information index corresponding to the event detection model.

3. The system according to claim 2, characterized in that The client is further used to: during the t-th round of training, input the sample message graph into the event detection model to obtain a training result; input the training result and the sample label into a preset loss function to obtain the loss value; Based on the mutual information between the sample label and the training result, the standardized mutual information index is obtained; wherein the preset loss function is a loss function pre-constructed based on contrastive learning.

4. The system according to claim 3, characterized in that The client is further used to: send the data quality contribution, the loss value and the normalized mutual information index to the server; The server is further configured to: upon receiving the loss value and the normalized mutual information index, calculate the model performance contribution based on the loss value and the normalized mutual information index; Based on the model performance contribution and the data quality contribution, the aggregation weight corresponding to each of the clients is calculated.

5. The system according to claim 3, characterized in that The client is further used to: calculate the model performance contribution based on the loss value and the standardized mutual information index; and send the data quality contribution and the model performance contribution to the server; The server is further configured to: upon receiving the data quality contribution and the model performance contribution, calculate the aggregation weight corresponding to each of the clients based on the model performance contribution and the data quality contribution.

6. The system according to claim 2, characterized in that The client is further configured to: when the sample training set is updated, calculate the current data quality contribution based on the updated sample training set; and send the current data quality contribution to the server; The server is further configured to: upon receiving the current data quality contribution, update the data quality contribution based on the current data quality contribution.

7. The system according to claim 1, characterized in that Before executing the step of using the local training set to perform the tth round of training on the event detection model stored locally, the client is also used to: determine whether the event detection model has converged, and if the event detection model has not converged, execute the step of using the local training set to perform the tth round of training on the event detection model stored locally.

8. The system according to claim 1, characterized in that The server is used to: send the initial event detection model to each client; The client is further configured to: upon receiving the initial event detection model, store the initial event detection model as the event detection model.

9. The system according to claim 8, characterized in that The event detection model includes a feature extractor and an event detector; the feature extractor is constructed based on a graph attention network and contrastive learning, and the event detector detects events based on a clustering method.

10. A method for federated training of event detection models based on aggregation strategy, characterized in that: Applied to the event detection model federation training system based on aggregation strategy as described in claims 1 to 9, the method comprises: The client uses the local training set to perform the t-th round of training on the event detection model stored locally, obtains the event detection model after the t-th training round, and sends the model parameters corresponding to the event detection model to the server; wherein the local training set includes a sample message graph, a sample label corresponding to the sample message graph, and a verification message graph; t is a positive integer ranging from 1 to T in sequence, and T is a preset number of training times; The server receives the model parameters sent by the client, aggregates the model parameters based on the aggregation weights corresponding to the respective clients, obtains global event detection model parameters, and sends the global event detection model parameters to the respective clients; The global event detection model parameters are received through the client, and the event detection model is updated and stored using the global event detection model parameters; when t is less than T, t=t+1 is set, and the step of using the local training set to perform the tth round of training on the locally stored event detection model is executed.

Citation Information

Patent Citations

  • Recommendation model training method and device, recommendation method and device, electronic equipment and medium

    CN114817734A

  • Node data classification method and system based on federated learning

    CN116541769A

  • Integrated Sensing and Communications Empowered by Networked Hybrid Quantum-Classical Machine Learning

    US20230368065A1

  • Cross-customer weighted federated domain adaptation for event detection in warehouses

    US20240078409A1

Cited By

  • Automatic sample collection method based on artificial intelligence cloud edge collaboration

    CN120355964A

  • Traffic event prediction method, prediction device and prediction system

    CN121981316A

  • A traffic incident prediction method, prediction device, and prediction system

    CN121981316B