Multi-task cooperative training system and method based on gradient weight dynamic adjustment
By adopting a system based on gradient weight dynamic adjustment in multitask learning, using graph networks and graph neural networks to learn task similarity, the problem that traditional methods are difficult to adapt to different task difficulties is solved, and better model generalization and adaptability are achieved.
Patent Information
- Application Number
- CN202411783522.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-12-05
- Publication Date
- 2025-05-06
AI Technical Summary
Traditional multi-task learning methods are difficult to flexibly adapt to the changes in difficulty and importance of different tasks in complex application scenarios, resulting in the model being overfitted or underfitted.
A multi-task collaborative training system based on gradient weight dynamic adjustment is adopted. The similarity relationship between task samples is learned through the graph network model and the graph neural network model, and the gradient weight is dynamically adjusted for multi-task collaborative training.
The generalization ability and adaptability of the model are improved, overfitting or underfitting problems are avoided, and better model performance is achieved.
Smart Images
Figure CN119940400A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of machine learning, and in particular to a multi-task collaborative training system and method based on dynamic adjustment of gradient weights. Background Art
[0002] With the rapid development of artificial intelligence and big data technology, multi-task learning has been widely used in the fields of machine learning and deep learning. Multi-task learning allows the model to train multiple related tasks at the same time, and improves the learning efficiency and effect of each task by sharing representation.
[0003] In multi-task learning, there are often certain correlations and dependencies between tasks. However, traditional multi-task learning methods often adopt fixed weight distribution strategies, such as equal weight distribution or weight distribution based on task difficulty. In complex application scenarios, the difficulty and importance of different tasks may vary significantly, and the fixed weight distribution method cannot flexibly adapt to these changes, resulting in the model overfitting on some tasks and underfitting on other tasks. Summary of the invention
[0004] The purpose of the present invention is to provide a multi-task collaborative training system and method based on dynamic adjustment of gradient weights to solve the problems raised in the above background technology.
[0005] To achieve the above-mentioned purpose, the present invention provides the following technical solutions: a multi-task collaborative training system and method based on dynamic adjustment of gradient weights, characterized in that it includes a multi-task sample set acquisition module, a data preprocessing module, a graph network model update module, a gradient weight determination module and a collaborative training module, wherein the output end of the multi-task sample set acquisition module is connected to the input end of the data preprocessing module, the output end of the data preprocessing module is connected to the input end of the graph network model update module, the output end of the graph network model update module is connected to the input end of the gradient weight determination module, and the output end of the gradient weight determination module is connected to the input end of the collaborative training module, wherein:
[0006] The multi-task sample set acquisition module is used to acquire a task sample set, wherein the task sample set includes multiple task samples, and each task sample includes basic information and features of the corresponding sample;
[0007] The data preprocessing module is used to preprocess the task sample set acquired by the multi-task sample set acquisition module;
[0008] The graph network model updating module is used to obtain each task sample preprocessed by the data preprocessing module, and input each task sample into the graph network model constructed based on the task samples in the historical task sample set, and construct each task sample as a new node in the graph network model to update the graph network model;
[0009] The gradient weight determination module is used to obtain the graph network model updated by the graph network model update module, and input the updated graph network model into the graph neural network model pre-trained using the historical task sample set to obtain the gradient weights corresponding to each task sample in the task sample set;
[0010] The collaborative training module is used to perform multi-task collaborative training corresponding to each task sample based on the gradient weight determined by the gradient weight determination module.
[0011] The present invention provides a multi-task collaborative training system based on dynamic adjustment of gradient weights. The multi-task sample set acquisition module is used to collect task sample sets, and then the collected task sample sets are preprocessed by the data preprocessing module. Next, the graph network model update module uses the preprocessed task samples as new nodes to update the graph network model. Specifically, the preprocessed task samples are input into the graph network model constructed based on the task samples in the historical task sample set, and each task sample is constructed as a new node in the graph network model to update the graph network model. Then, the gradient weight determination module inputs the graph network model into the graph neural network model pre-trained using the historical task sample set to learn the similarity relationship between task samples and generate a graph neural network model for gradient weight determination. Finally, the collaborative training module performs multi-task collaborative training according to the determined gradient weights to achieve better model performance and generalization ability.
[0012] Furthermore, before the multi-task sample acquisition module, it includes a historical task sample set acquisition module, a graph network model construction module and a graph neural network training module, wherein:
[0013] The historical task sample set acquisition module is used to acquire a historical task sample set, wherein the historical task sample set includes a plurality of historical task samples, each of which includes basic information and features of the corresponding sample and label data representing a gradient weight corresponding to the task sample;
[0014] The graph network model construction module is used to obtain the historical task sample set in the historical task sample set acquisition module and construct an initial graph network model, wherein the nodes of the graph network model are constructed with the historical task samples, and the nodes are connected based on the similarity of the historical task samples;
[0015] The graph neural network training module is used to take the initial graph network model constructed by the graph network model construction module as input, train the graph neural network model to learn the similarity relationship between historical task samples, and generate a graph neural network model for gradient weight determination.
[0016] By introducing the historical task sample set acquisition module, graph network model construction module and graph neural network training module, we can better understand the similarities and differences between tasks, thereby improving the generalization ability of the model. The graph network model built based on historical task samples can capture the complex relationship between tasks and provide richer information for subsequent gradient weight determination. By training the graph neural network model, we can learn the similarity relationship between historical task samples and generate a model for gradient weight determination. This method of dynamically adjusting gradient weights makes multi-task collaborative training more flexible and efficient.
[0017] Furthermore, the graph network model construction module includes a node construction unit and an edge construction unit, wherein:
[0018] The node construction unit is used to construct each historical task sample in the historical task sample set as a node in the graph network model;
[0019] The edge construction unit is used to construct edge connections between two nodes according to the similarities between the historical task samples, so as to form an association relationship between the historical task samples.
[0020] By introducing node construction units and edge construction units, the graph network model construction module can construct each historical task sample in the historical task sample set as a node in the graph network model, and build edge connections between two nodes based on the similarity between historical task samples, thereby forming an association relationship between historical task samples. This design can better capture the complex relationship between tasks and provide richer information for subsequent gradient weight determination, thereby improving the generalization ability of the model and the effect of multi-task collaborative training.
[0021] Furthermore, the graph neural network training module includes a model initialization unit, a training data input unit, a model training unit and a model output unit, wherein:
[0022] The model initialization unit is used to initialize the parameters of the graph neural network model;
[0023] The training data input unit is used to obtain the graph network model constructed by the graph network model construction module, and input the graph network model into the graph neural network model initialized in the model initialization unit;
[0024] The model training unit is used to train the graph neural network model to a convergence state by using a back propagation algorithm and a gradient descent method, taking the label data of the gradient weights corresponding to the characterization task samples as a supervision standard;
[0025] The model output unit is used to output a trained graph neural network model, which is used for subsequent gradient weight determination.
[0026] By introducing a model initialization unit, a training data input unit, a model training unit, a model output unit, and a model output unit, the graph neural network training module can initialize the parameters of the graph neural network model, obtain the graph network model constructed by the graph network model construction module, and input it into the initialized graph neural network model. Through the back propagation algorithm and the gradient descent method, the label data representing the gradient weights corresponding to the task samples is used as the supervision standard to train the graph neural network model to a convergence state, and output the trained graph neural network model for subsequent gradient weight determination. This can learn the similarity relationship between historical task samples and generate a model for gradient weight determination.
[0027] Furthermore, the preprocessing in the data preprocessing module includes data cleaning and feature engineering, which encodes the features of each ad space into a vector representation to provide input for subsequent model training. Through data cleaning and feature engineering in the data preprocessing module, the task sample set can be preprocessed with high quality. Data cleaning removes noise and redundant information, and feature engineering encodes the features of each task into a vector representation to provide clear, consistent and high-quality input data for subsequent model training. This preprocessing method significantly improves the efficiency and accuracy of model training, ensuring that the model can better understand and process task samples, thereby improving the overall effect of multi-task collaborative training.
[0028] Furthermore, the graph network model update module constructs each task sample as a new starting point of the graph network model, calculates the similarity between the newly added nodes and each node in the graph network model based on the basic information and characteristics of the task samples, and establishes edge connections between the newly added nodes and each node in the graph network model to update the graph network model.
[0029] Through the graph network model update module, each task sample can be constructed as a new node of the graph network model, and the similarity between the new node and each node in the graph network model can be calculated based on the basic information and features of the task sample. By establishing edge connections between the new nodes and the existing nodes, the graph network model can be dynamically updated to capture the relationship between the new task samples and the historical task samples.
[0030] See also Figure 2As shown, a multi-task collaborative training method based on dynamic adjustment of gradient weights specifically includes the following steps:
[0031] S01: Obtain a task sample set, and perform a preprocessing operation on the task sample set;
[0032] S02: inputting the task sample set into a graph network model constructed based on task samples in the historical task sample set, constructing each task sample as a newly added node in the graph network model, so as to update the graph network model;
[0033] S03: Inputting the graph network model updated in step S02 into the graph neural network model pre-trained using the historical task sample set to obtain the gradient weights corresponding to each task sample in the task sample set;
[0034] S04: Perform multi-task collaborative training corresponding to each task sample according to the gradient weight determined in step S03.
[0035] Furthermore, before step S01, the following steps are specifically included:
[0036] S05: Acquire a historical task sample set, and perform a preprocessing operation on the historical task sample set;
[0037] S06: constructing an initial graph network model based on the historical task sample set preprocessed in step S05;
[0038] S07: Use the graph network model constructed in step S06 as input to train the graph neural network model to learn the similarity relationship between historical task samples and generate a graph neural network model for gradient weight determination.
[0039] Furthermore, the task sample set in step S01 includes multiple task samples, and each task sample includes basic information and features of the corresponding sample.
[0040] Furthermore, when training the graph neural network model in step S07, label data representing the gradient weights corresponding to the task samples are used as supervisory labels for training, and the graph neural network model is trained to a convergence state.
[0041] It is to be understood that the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the present disclosure.
[0042] Compared with the prior art, the beneficial effects of the present invention are: by introducing a historical task sample set acquisition module, a graph network model construction module and a graph neural network training module, the present invention can better understand the similarities and differences between tasks, capture the complex relationship between tasks, and provide richer information for subsequent gradient weight determination, thereby improving the generalization ability of the model. The multi-task collaborative training system and method with dynamic adjustment of gradient weights can dynamically adjust the gradient weights according to the similarity relationship between task samples. Compared with the fixed weight allocation strategy, it can more flexibly adapt to the difficulty and importance changes of different tasks, thereby improving the adaptability and training effect of the model, and avoiding the overfitting or underfitting problems that may occur in certain tasks. BRIEF DESCRIPTION OF THE DRAWINGS
[0043] In order to more clearly illustrate the implementation methods of the present invention or the technical solutions in the prior art, the drawings required for the implementation methods or the prior art descriptions are briefly introduced below. Obviously, the drawings in the following description are only exemplary, and for ordinary technicians in this field, other implementation drawings can be derived from the provided drawings without creative work.
[0044] Figure 1 A system block diagram of a multi-task collaborative training system based on dynamic adjustment of gradient weights according to the present invention;
[0045] Figure 2 A system block diagram for training a graph neural network model of the present invention;
[0046] Figure 3 A workflow diagram of a multi-task collaborative training method based on dynamic adjustment of gradient weights according to the present invention;
[0047] Figure 4 This is a workflow diagram for training the graph neural network model of the present invention.
[0048] Among them, 1. Multi-task sample set acquisition module; 2. Data preprocessing module; 3. Graph network model update module; 4. Gradient weight determination module; 5. Collaborative training module; 01. Historical task sample set acquisition module; 10. Graph network model construction module; 101. Node construction unit; 102. Edge construction unit; 11. Graph neural network training module; 111. Model initialization unit; 112. Training data input unit; 113. Model training unit; 114. Model output unit. DETAILED DESCRIPTION
[0049] Exemplary embodiments will be described in detail herein, examples of which are shown in the accompanying drawings. When the following description refers to the drawings, the same numbers in different drawings represent the same or similar elements unless otherwise indicated. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present disclosure. Instead, they are merely examples of devices consistent with some aspects of the present disclosure as detailed in the appended claims.
[0050] The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the accompanying drawings in the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0051] See also Figure 1 As shown, a multi-task collaborative training system and method based on dynamic adjustment of gradient weights according to an embodiment of the present invention comprises a multi-task sample set acquisition module 1, a data preprocessing module 2, a graph network model update module 3, a gradient weight determination module 4 and a collaborative training module 5, wherein the output end of the multi-task sample set acquisition module 1 is connected to the input end of the data preprocessing module 2, the output end of the data preprocessing module 2 is connected to the input end of the graph network model update module 3, the output end of the graph network model update module 3 is connected to the input end of the gradient weight determination module 4, the output end of the gradient weight determination module 4 is connected to the input end of the collaborative training module 5, wherein:
[0052] The multi-task sample set acquisition module 1 is used to acquire a task sample set, where the task sample set includes multiple task samples, and each task sample includes basic information and features of the corresponding sample;
[0053] The data preprocessing module 2 is used to preprocess the task sample set acquired by the multi-task sample set acquisition module 1;
[0054] The graph network model updating module 3 is used to obtain each task sample preprocessed by the data preprocessing module 2, and input each task sample into the graph network model constructed based on the task samples in the historical task sample set, and construct each task sample as a new node in the graph network model to update the graph network model;
[0055] The gradient weight determination module 4 is used to obtain the graph network model updated by the graph network model update module 3, and input the updated graph network model into the graph neural network model pre-trained using the historical task sample set to obtain the gradient weights corresponding to each task sample in the task sample set;
[0056] The collaborative training module 5 is used to perform multi-task collaborative training corresponding to each task sample based on the gradient weight determined by the gradient weight determination module 4.
[0057] The present invention provides a multi-task collaborative training system based on dynamic adjustment of gradient weights. The multi-task sample set acquisition module 1 is used to collect task sample sets, and then the collected task sample sets are preprocessed by the data preprocessing module 2. Next, the graph network model update module 3 uses the preprocessed task samples as new nodes to update the graph network model. Specifically, the preprocessed task samples are input into the graph network model constructed based on the task samples in the historical task sample set, and each task sample is constructed as a new node in the graph network model to update the graph network model. Then, the gradient weight determination module 4 inputs the graph network model into the graph neural network model pre-trained using the historical task sample set to learn the similarity relationship between the task samples and generate a graph neural network model for gradient weight determination. Finally, the collaborative training module 5 performs multi-task collaborative training according to the determined gradient weights to achieve better model performance and generalization ability.
[0058] Further, see Figure 2 As shown, before the multi-task sample acquisition module 1, it includes a historical task sample set acquisition module 01, a graph network model construction module 10 and a graph neural network training module 11, wherein:
[0059] The historical task sample set acquisition module 01 is used to acquire a historical task sample set, which includes a plurality of historical task samples, each of which includes basic information and features of the corresponding sample and label data representing the gradient weight corresponding to the task sample;
[0060] The graph network model construction module 10 is used to obtain the historical task sample set in the historical task sample set acquisition module 01 and construct an initial graph network model, wherein the nodes of the graph network model are constructed with the historical task samples, and the nodes are connected based on the similarity of the historical task samples;
[0061] The graph neural network training module 11 is used to take the initial graph network model constructed by the graph network model construction module 10 as input, train the graph neural network model to learn the similarity relationship between historical task samples, and generate a graph neural network model for gradient weight determination.
[0062] By introducing the historical task sample set acquisition module 01, the graph network model construction module 10 and the graph neural network training module 11, we can better understand the similarities and differences between tasks, thereby improving the generalization ability of the model. The graph network model constructed based on historical task samples can capture the complex relationship between tasks and provide richer information for subsequent gradient weight determination. By training the graph neural network model, we can learn the similarity relationship between historical task samples and generate a model for gradient weight determination. This method of dynamically adjusting gradient weights makes multi-task collaborative training more flexible and efficient.
[0063] Furthermore, the graph network model construction module 10 includes a node construction unit 101 and an edge construction unit 102, wherein:
[0064] The node construction unit 101 is used to construct each historical task sample in the historical task sample set as a node in the graph network model;
[0065] The edge construction unit 102 is used to construct edge connections between two nodes according to the similarity between the historical task samples, so as to form an association relationship between the historical task samples.
[0066] By introducing the node construction unit 101 and the edge construction unit 102, the graph network model construction module 10 can construct each historical task sample in the historical task sample set as a node in the graph network model, and construct edge connections between two nodes according to the similarity between the historical task samples, thereby forming an association relationship between the historical task samples. This design can better capture the complex relationship between tasks and provide richer information for subsequent gradient weight determination, thereby improving the generalization ability of the model and the effect of multi-task collaborative training.
[0067] Furthermore, the graph neural network training module 11 includes a model initialization unit 111, a training data input unit 112, a model training unit 113 and a model output unit 114, wherein:
[0068] The model initialization unit 111 is used to initialize the parameters of the graph neural network model;
[0069] The training data input unit 112 is used to obtain the graph network model constructed by the graph network model construction module 10, and input the graph network model into the graph neural network model initialized in the model initialization unit 111;
[0070] The model training unit 113 is used to train the graph neural network model to a convergence state by using a back propagation algorithm and a gradient descent method, taking the label data representing the gradient weight corresponding to the task sample as a supervision standard;
[0071] The model output unit 114 is used to output the trained graph neural network model, which is used for subsequent gradient weight determination.
[0072] By introducing the model initialization unit 111, the training data input unit 112, the model training unit 113, and the model output unit 114, the graph neural network training module 11 can initialize the parameters of the graph neural network model, obtain the graph network model constructed by the graph network model construction module 10, and input it into the initialized graph neural network model. Through the back propagation algorithm and the gradient descent method, the label data representing the gradient weights corresponding to the task samples is used as the supervision standard to train the graph neural network model to a convergence state, and output the trained graph neural network model for subsequent gradient weight determination. This can learn the similarity relationship between historical task samples and generate a model for gradient weight determination.
[0073] Furthermore, the preprocessing in the data preprocessing module 2 includes data cleaning and feature engineering, which encodes the features of each ad space into a vector representation to provide input for subsequent model training. Through the data cleaning and feature engineering in the data preprocessing module 2, the task sample set can be preprocessed with high quality. Data cleaning removes noise and redundant information, and feature engineering encodes the features of each task into a vector representation to provide clear, consistent and high-quality input data for subsequent model training. This preprocessing method significantly improves the efficiency and accuracy of model training, ensuring that the model can better understand and process task samples, thereby improving the overall effect of multi-task collaborative training.
[0074] Furthermore, the graph network model updating module 3 constructs each task sample as a new starting point of the graph network model, calculates the similarity between the newly added nodes and each node in the graph network model based on the basic information and characteristics of the task samples, and establishes edge connections between the newly added nodes and each node in the graph network model to update the graph network model.
[0075] Through the graph network model update module 3, each task sample can be constructed as a new node of the graph network model, and the similarity between the new node and each node in the graph network model can be calculated based on the basic information and characteristics of the task sample. By establishing edge connections between the new nodes and the existing nodes, the graph network model can be dynamically updated to capture the relationship between the new task samples and the historical task samples.
[0076] See also Figure 3 As shown, a multi-task collaborative training method based on dynamic adjustment of gradient weights specifically includes the following steps:
[0077] S01: Obtain a task sample set and perform preprocessing operations on the task sample set;
[0078] S02: input the task sample set into a graph network model constructed based on task samples in the historical task sample set, and construct each task sample as a new node in the graph network model to update the graph network model;
[0079] S03: Input the graph network model updated in step S02 into the graph neural network model pre-trained using the historical task sample set to obtain the gradient weights corresponding to each task sample in the task sample set;
[0080] S04: Perform multi-task collaborative training corresponding to each task sample according to the gradient weight determined in step S03.
[0081] Further, see Figure 4 As shown, before step S01, the following steps are specifically included:
[0082] S05: Obtain a historical task sample set, and perform preprocessing operations on the historical task sample set;
[0083] S06: constructing an initial graph network model based on the historical task sample set preprocessed in step S05;
[0084] S07: Use the graph network model constructed in step S06 as input to train the graph neural network model to learn the similarity relationship between historical task samples and generate a graph neural network model for gradient weight determination.
[0085] Furthermore, the task sample set in step S01 includes multiple task samples, and each task sample includes basic information and features of the corresponding sample.
[0086] Furthermore, when training the graph neural network model in step S07, the label data representing the gradient weights corresponding to the task samples is used as the supervision label for training, and the graph neural network model is trained to a convergence state.
[0087] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.
[0088] Those skilled in the art will readily appreciate other embodiments of the present disclosure after considering the specification and practicing the disclosure disclosed herein. This application is intended to cover any variations, uses, or adaptations of the present disclosure that follow the general principles of the present disclosure and include common knowledge or customary techniques in the art that are not disclosed in the present disclosure. The description and examples are intended to be exemplary only, and the true scope of the present disclosure is indicated by the following claims.
Claims
1. A multi-task collaborative training system based on dynamic adjustment of gradient weights, characterized in that: The method comprises a multi-task sample set acquisition module (1), a data preprocessing module (2), a graph network model updating module (3), a gradient weight determination module (4) and a collaborative training module (5), wherein the output end of the multi-task sample set acquisition module (1) is connected to the input end of the data preprocessing module (2), the output end of the data preprocessing module (2) is connected to the input end of the graph network model updating module (3), the output end of the graph network model updating module (3) is connected to the input end of the gradient weight determination module (4), and the output end of the gradient weight determination module (4) is connected to the input end of the collaborative training module (5), wherein: The multi-task sample set acquisition module (1) is used to acquire a task sample set, wherein the task sample set includes a plurality of task samples, and each task sample includes basic information and features of the corresponding sample; The data preprocessing module (2) is used to preprocess the task sample set acquired by the multi-task sample set acquisition module (1); The graph network model updating module (3) is used to obtain each task sample preprocessed by the data preprocessing module (2), and input each task sample into a graph network model constructed based on the task samples in the historical task sample set, and construct each task sample as a new node in the graph network model to update the graph network model; The gradient weight determination module (4) is used to obtain the graph network model updated by the graph network model updating module (3), and input the updated graph network model into the graph neural network model pre-trained using the historical task sample set to obtain the gradient weight corresponding to each task sample in the task sample set; The collaborative training module (5) is used to perform multi-task collaborative training corresponding to each task sample based on the gradient weight determined by the gradient weight determination module (4).
2. A multi-task collaborative training system based on dynamic adjustment of gradient weights according to claim 1, characterized in that: Before the multi-task sample acquisition module (1), it includes a historical task sample set acquisition module (01), a graph network model construction module (10) and a graph neural network training module (11), wherein: The historical task sample set acquisition module (01) is used to acquire a historical task sample set, wherein the historical task sample set includes a plurality of historical task samples, each of which includes basic information and features of the corresponding sample and label data representing a gradient weight corresponding to the task sample; The graph network model construction module (10) is used to obtain the historical task sample set in the historical task sample set acquisition module (01) and construct an initial graph network model, wherein the nodes of the graph network model are constructed with the historical task samples, and the nodes are connected based on the similarity of the historical task samples; The graph neural network training module (11) is used to take the initial graph network model constructed by the graph network model construction module (10) as input, train the graph neural network model to learn the similarity relationship between historical task samples, and generate a graph neural network model for gradient weight determination.
3. A multi-task collaborative training system based on dynamic adjustment of gradient weights according to claim 2, characterized in that: The graph network model construction module (10) comprises a node construction unit (101) and an edge construction unit (102), wherein: The node construction unit (101) is used to construct each historical task sample in the historical task sample set as a node in a graph network model; The edge construction unit (102) is used to construct edge connections between two nodes according to the similarities between the historical task samples, so as to form an association relationship between the historical task samples.
4. A multi-task collaborative training system based on dynamic adjustment of gradient weights according to claim 2, characterized in that: The graph neural network training module (11) comprises a model initialization unit (111), a training data input unit (112), a model training unit (113) and a model output unit (114), wherein: The model initialization unit (111) is used to initialize the parameters of the graph neural network model; The training data input unit (112) is used to obtain the graph network model constructed by the graph network model construction module (10), and input the graph network model into the graph neural network model initialized in the model initialization unit (111); The model training unit (113) is used to train the graph neural network model to a convergence state by using a back propagation algorithm and a gradient descent method, taking the label data of the gradient weight corresponding to the characterization task sample as a supervision standard; The model output unit (114) is used to output a trained graph neural network model, which is used for subsequent gradient weight determination.
5. The multi-task collaborative training system based on dynamic adjustment of gradient weights according to claim 1, characterized in that: The preprocessing in the data preprocessing module (2) includes data cleaning and feature engineering, wherein the feature engineering encodes the features of each task into a vector representation to provide input for subsequent model training.
6. A multi-task collaborative training system based on dynamic adjustment of gradient weights according to claim 1, characterized in that: The graph network model updating module (3) constructs each task sample as a new starting point of the graph network model, calculates the similarity between the new node and each node in the graph network model based on the basic information and features of the task sample, and establishes edge connections between the new node and each node in the graph network model to update the graph network model.
7. A multi-task collaborative training method based on dynamic adjustment of gradient weights, characterized in that: The specific steps include: S01: Obtain a task sample set, and perform a preprocessing operation on the task sample set; S02: inputting the task sample set into a graph network model constructed based on task samples in the historical task sample set, constructing each task sample as a newly added node in the graph network model, so as to update the graph network model; S03: Inputting the graph network model updated in step S02 into the graph neural network model pre-trained using the historical task sample set to obtain the gradient weights corresponding to each task sample in the task sample set; S04: Perform multi-task collaborative training corresponding to each task sample according to the gradient weight determined in step S03.
8. The multi-task collaborative training method based on dynamic adjustment of gradient weights according to claim 7, characterized in that: Before step S01, the following steps are specifically included: S05: Acquire a historical task sample set, and perform a preprocessing operation on the historical task sample set; S06: constructing an initial graph network model based on the historical task sample set preprocessed in step S05; S07: Use the graph network model constructed in step S06 as input to train the graph neural network model to learn the similarity relationship between historical task samples and generate a graph neural network model for gradient weight determination.
9. The multi-task collaborative training method based on dynamic adjustment of gradient weights according to claim 7, characterized in that: The task sample set in step S01 includes multiple task samples, and each task sample includes basic information and features of the corresponding sample.
10. The multi-task collaborative training method based on dynamic adjustment of gradient weights according to claim 8, characterized in that: When the step S07 trains the graph neural network model, the label data representing the gradient weights corresponding to the task samples is used as the supervisory label for training, and the graph neural network model is trained to a convergence state.