Confidence-guided adaptive graph representation reinforcement learning method, device and medium

Abstract

The application provides a confidence-guided adaptive graph representation enhanced learning method, device and medium, and relates to the technical fields of label prediction and electric data processing.The application obtains a crowdsourcing data set composed of multi-source heterogeneous electronic data, and inputs an adaptive graph representation learning model;the crowdsourcing data set is mapped to construct an original bipartite graph;based on the original bipartite graph, a pruning view and an enhanced view are generated through dynamic confidence pruning and conservativeness edge increase;then, after feature extraction, adaptive information enhancement and learning of worker nodes and task nodes are respectively performed through a double-channel heterogeneous adaptive network, an initial prediction label is obtained through a predictor;and the initial prediction label is corrected in combination with the original bipartite graph to obtain a final prediction label.The application can effectively suppress noise interference, capture local features and adapt to a dynamic environment, is significantly superior to the prior art in terms of accuracy, robustness and computational efficiency, and is suitable for fields such as multi-source heterogeneous intelligent labeling.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of graph representation learning and electronic data processing, and particularly relates to a confidence-guided adaptive graph representation reinforcement learning method, device and medium. BACKGROUND

[0002] In today's big data era, electronic data presents the characteristics of massive, multi-source and heterogeneous. These data may come from different sensors, business systems, network logs or user-generated content. However, different sources of electronic data often have inconsistent or even contradictory information, such as price description differences of the same commodity on different platforms or abnormal readings of different sensors of the same device. How to infer and predict the most reliable true value from these conflicting and noisy electronic data is the key of the present application.

[0003] Existing graph representation prediction methods mainly include direct statistical methods, probabilistic graph models and graph representation learning-based methods, but these methods have significant limitations in dealing with noisy data and sparse associations. Direct statistical model-based methods, such as majority voting algorithm, simply take the answers of the majority of workers as the task truth, ignoring the differences in worker ability, resulting in low accuracy under noisy data; weighted voting algorithm, although it improves the aggregation accuracy by optimizing worker weights, its effect is highly dependent on the accurate estimation of worker ability, but the actual scene is that the quality of workers is uneven, making this assumption difficult to establish; confidence algorithm, which introduces confidence interval estimation to optimize the reliability of weighted voting, still assumes that worker labeling behavior is consistent, and has limited effect when dealing with sparse data. Probabilistic graph model-based methods, such as GLAD algorithm, improve labeling accuracy by jointly modeling task difficulty and worker quality, but the number of parameters increases dramatically, significantly prolonging the training time; DS algorithm uses EM algorithm for iterative optimization, but assumes that workers are independent in labeling, which is seriously inconsistent with the actual situation where workers may exchange or reference; EBCC algorithm introduces a sub-type mixture model to capture worker correlation under the Bayesian framework, but relies on data distribution assumptions; BWA algorithm combines task difficulty and annotator reliability to infer consensus labels, but ignores the mutual influence between workers. This type of model has advantages in expressing uncertainty, but is sensitive to prior assumptions, computationally complex and not suitable for dynamic scenarios. Graph representation learning-based methods, such as TiReMGE algorithm, integrate multiple types of electronic data relationships and introduce a genetic reliability update mechanism to solve the problem of conflicting annotations, but rely on pre-set electronic data relationships, making it difficult to adapt to dynamic environments; GOVERN algorithm uses random perturbation enhancement strategy, which is easy to introduce noisy edges, resulting in insufficient robustness of the model to noisy data.

[0004] The prior art has obvious deficiencies in dealing with the complexity, noise problem and dynamic environment adaptability of multi-source heterogeneous electronic data, and there is still a need to develop a more robust, adaptive and effective electronic digital data inference method that can effectively capture local features.

[0005] In view of this, the present application is proposed. SUMMARY

[0006] The present application aims to provide a confidence-guided adaptive graph representation reinforcement learning method, device and medium to solve the deficiencies of existing graph representation learning models in noise suppression, local feature capture and dynamic environment adaptability when processing multi-source heterogeneous data.

[0007] To solve the above technical problems, the present application realizes the following technical solutions:

[0008] A confidence-guided adaptive graph representation reinforcement learning method, comprising:

[0009] Obtaining a crowdsourcing data set composed of multi-source heterogeneous electronic data of different features;

[0010] Inputting the crowdsourcing data set into an adaptive graph representation learning model, using a multi-objective collaborative optimization loss function for learning and training, to obtain a trained adaptive graph representation learning model for subsequent classification prediction tasks; wherein the adaptive graph representation learning model includes a node mapping module, a view generation module, a dual-channel adaptive learning module and a label correction module;

[0011] The crowdsourcing data set is mapped and constructed into an original bipartite graph containing worker nodes and task nodes through the node mapping module;

[0012] Based on the original bipartite graph, the view generation module is used to generate pruned views and enhanced views through dynamic confidence pruning and conservativeness edge addition;

[0013] According to the pruned views and the enhanced views, the dual-channel heterogeneous adaptive network of the dual-channel adaptive learning module is used to extract features, enhance and learn adaptive information of worker nodes and task nodes respectively, and then the initial prediction labels are obtained through a predictor;

[0014] In combination with the original bipartite graph, the label correction module is used to correct the initial prediction labels to obtain the final prediction labels.

[0015] Preferably, the crowdsourcing data set is obtained by preprocessing multi-source heterogeneous electronic data; the preprocessing includes data cleaning, data conversion, data desensitization and data standardization.

[0016] Preferably, the generation process of the pruned view is as follows:

[0017] First, based on the original bipartite graph, the maximum class confidence of each edge is calculated using the edge predictor, and the expression of the class confidence is:

[0018] = ;

[0019] wherein, denotes the class confidence of the edge composed of the i-th worker node and the j-th task node; is an edge predictor; is a weight matrix of worker node features; denotes the i-th worker node feature vector; is a weight matrix of task node features; denotes the j-th task node feature vector; is a weight matrix of edge features; is a linear transformation function used to convert the output of the edge predictor into a probability distribution; is a feature vector of the edge ; is a bias term of the edge predictor; is a feature concatenation operation;

[0020] Then, the maximum class confidence of each edge is taken as the confidence, and the edges of the higher probability class are selected to retain high-quality labeled edges, and the expression is:

[0021] ;

[0022] wherein, is the confidence of the edge ; C denotes a set of all possible label classes; c denotes a class; denotes a class label; denotes the maximum value for all possible label classes c; denotes a class confidence function;

[0023] Next, dynamic edge pruning is performed, and a confidence-sensitive factor is used to adaptively adjust the edge pruning probability to control the sensitivity of the confidence edge, and the expression of the dynamic edge pruning probability is:

[0024] ;

[0025] wherein, denotes the dynamic pruning probability of the edge ; is a minimum constant; is a confidence-sensitive factor; ​trimming weight coefficients; denotes L2 norm, used to measure the difference of features; denotes the i-th worker node feature vector; denotes the j-th task node feature vector;

[0026] trimming the edges of the original bipartite graph based on the dynamic edge trimming probability to obtain the trimmed view.

[0027] Preferably, the generation process of the enhanced view is:

[0028] The maximum class confidence of each edge calculated by the edge predictor is taken as the edge increase probability, and the formula is:

[0029] ;

[0030] wherein, denotes the edge increase probability of ; is the edge increase weight coefficient; is a nonlinear activation function; is the weight matrix of edge increase; T is the transpose symbol; is the confidence of edge ; denotes the i-th worker node feature vector; denotes the j-th task node feature vector;

[0031] based on the calculated edge increase probability, the original bipartite graph is edge-increased to obtain the enhanced view.

[0032] Preferably, according to the trimmed view and the enhanced view, the worker nodes and the task nodes are respectively extracted, adaptively information enhanced and learned by a dual-channel heterogeneous adaptive network, specifically:

[0033] First, the worker feature embedding, task feature embedding and edge feature embedding are extracted according to the trimmed view and the enhanced view respectively;

[0034] Then, the transformed weight matrix is used for embedding initialization operation to generate the corresponding initial embedding, and the expression is:

[0035] ;

[0036] ;

[0037] ;

[0038] wherein, , , These represent the initial worker feature embedding, the initial task feature embedding, and the initial edge feature embedding, respectively. , , These represent the extracted worker feature embedding, task feature embedding, and edge feature embedding, respectively. , , They represent , , The weight matrix; , , They represent , , Learnable parameters; Indicates feature index; Indicates the total number of features; Presentation layer normalization operation;

[0039] Then, based on the initial embedding, worker nodes for each iteration are obtained through hierarchical information passing and aggregation. and task nodes Hidden embedding representation and Then, in the i-th iteration, the first... The formula for the embedding of a layer is:

[0040] ;

[0041] ;

[0042] in, , The i-th iteration and the i-th iteration are respectively Layer worker node and task nodes Embedded representation; It is a non-linear activation function; , They represent , The neighboring nodes; , They represent , The set of neighboring nodes; , They represent , Another neighboring node; This represents the original attention score calculated for each pair of nodes (i, m); This represents the attention score calculated for each pair of nodes (i,j) and the normalized result; This represents the temperature coefficient, used to control the smoothness of attention distribution; , They represent , The weight matrix; Indicates residual connection; , The i-th iteration and the i-th iteration are respectively Layer worker node and task nodes Embedded representation; For the i-th iteration, the th Embedded representation of layer edges; It is an exponential function;

[0043] A linear auxiliary layer combined with a multilayer perceptron is used to capture the semantic features of the hidden embedding representation, generating the final worker embedding. Representation and final task embedding The expression is:

[0044] ;

[0045] ;

[0046] in, , Let represent the final worker embedding and the final task embedding at the i-th iteration, respectively; Represents the auxiliary linear transformation function; For multilayer perceptron functions; , These are the weight matrices for the i-th iteration; This indicates a batch normalization operation; This represents the hidden worker embedding at the i-th iteration; This represents the hidden task embedding at the i-th iteration; This represents the learning parameters in the multilayer perceptron function; This represents the embedding bias at the i-th iteration;

[0047] Based on the final task embedding, the task projection node is obtained by mapping using the basic projection head;

[0048] An adaptive information enhancer and a learnable scale parameter are used to locally learn the task projection nodes. During training, backpropagation is used to adaptively adjust the contribution of task node features to obtain higher-order task features, thereby reducing noise information generated by the original task features. The expression of the higher-order task features is as follows:

[0049] ;

[0050] ;

[0051] wherein, is a high-order task feature; is a concatenation operation; is an element-wise multiplication; is a final task embedding; , , denotes a multi-head attention weight matrix; is a normalized scaling factor; is a matrix variation parameter; denotes a regularization operation; denotes a weight matrix; denotes a bias; denotes a total number of high-order task features; denotes an adaptive information enhancer;

[0052] generating an updated final task embedding based on the high-order task feature, and obtaining an initial predicted label by using a predictor.

[0053] Preferably, the predictor comprises a linear transformation layer and a normalization layer, the final task embedding is mapped to a low dimension by the linear transformation layer, and the initial predicted label is obtained by normalization, and the formula is:

[0054] ;

[0055] wherein, is an initial predicted label; is a linear transformation function; is a layer normalization operation; , denote normalized weight matrices, respectively; , is a normalized bias; is an activation function; is an updated final task embedding; , b denote a weight matrix and a bias of the predictor, respectively.

[0056] Preferably, in combination with the original bipartite graph, the initial predicted label is corrected, specifically:

[0057] First, a similarity weight matrix is calculated according to the original bipartite graph, then the initial predicted label is weighted and summed using the similarity weight matrix to obtain a weighted label, and the final predicted label is obtained by normalizing the weighted label, and the expression is:

[0058] ;

[0059] ;

[0060] wherein, is the initial predicted label is the corresponding final predicted label. is the similarity weight matrix of label i to sample j; denotes the initial label vector of label j; is the normalization of the label vector; is the total number of samples; denotes the multi-layer perception; is the layer normalization operation; is the task projection vector of label i, j , the cosine similarity of , is the similarity weight coefficient; is the transpose symbol; is the task projection vector of label ; , are the query matrix weight and key matrix weight of the attention network, respectively; is the standardization scaling factor; is the temperature parameter, used to control the degree of weight distribution between labels.

[0061] Preferably, learning training is performed using a multi-objective collaborative optimization loss function to maximize the similarity between positive sample pairs while minimizing the similarity between negative sample pairs;

[0062] wherein the multi-objective collaborative optimization loss function includes the optimization loss function between worker nodes and task nodes , the adaptive information enhancement loss function and the label loss function , the expression is:

[0063] ;

[0064] The expression of the optimization loss between worker nodes and task nodes is:

[0065] ;

[0066] ;

[0067] wherein, , are the worker node set and the task node set, respectively. , respectively represent the i-th worker node embedding under the q-th and k-th view; respectively represent the j-th worker node embedding under the q-th and k-th view; respectively represent the i-th task node embedding under the q-th and k-th view; is a temperature parameter; represents a similarity matrix of two vectors; respectively represent two projection vectors; is a weight matrix; is the j-th query vector under the k-th view; , is a similarity weight coefficient;

[0068] Adaptive information enhancement loss function is an Infonce loss used to adaptively adjust the weight of the task feature, and the expression is as follows:

[0069] ;

[0070] wherein, respectively represent the high-order task feature of the i-th task node under the q-th and k-th view respectively represent the high-order task feature of the j-th task node under the q-th and k-th view;

[0071] The label loss function is used to calculate the difference between the original label and the predicted label of different views through cross entropy, and the expression is as follows:

[0072] ;

[0073] wherein, is a label loss function weight; is a task node set; , represents the i-th label based on the pruning view and the enhancement view, respectively; represents an original label.

[0074] The application further provides a confidence-guided adaptive graph representation enhanced learning device, comprising a processor and a memory, wherein the memory stores a computer program which can be executed by the processor to implement the confidence-guided adaptive graph representation enhanced learning method.

[0075] The application further provides a computer-readable storage medium, wherein computer-readable instructions are stored on the computer-readable storage medium, and the computer-readable instructions are executed by a processor of a device where the computer-readable storage medium is located to implement the confidence-guided adaptive graph representation enhanced learning method.

[0076] In summary, compared with the prior art, the application has the following beneficial effects:

[0077] The adaptive graph representation learning model optimizes the graph structure dynamically through the confidence-guided view generation mechanism, realizes feature extraction and information enhancement in combination with the dual-channel heterogeneous adaptive network, dynamically weights the task node features by using the learnable scale parameter, improves the local semantic capturing precision of the graph representation learning model, effectively solves the prediction precision problem of the model under the interference of noise data and the sparse labeling scene, enhances the adaptive ability of the model to the dynamic environment, and improves the local feature capturing capability.

[0078] The application adopts the dynamic weighting correction mechanism based on feature similarity, adjusts the smoothness of the label weight distribution through the temperature parameter, effectively suppresses the influence of abnormal label labeling, and improves the prediction precision of the model.

[0079] The application generates two complementary views of confidence pruning views and confidence enhancement views, and cooperatively optimizes based on a multi-target cooperative optimization loss function, realizes cross-view feature alignment and feature invariance modeling, and improves the generalization ability and stability of the model on unseen data.

[0080] The application overcomes the shortcomings of traditional graph representation methods in noise suppression and heterogeneous information fusion, is widely applicable to the fields of multi-source heterogeneous data cleaning, data credible inference and intelligent labeling, and has significant practical value and popularization potential. BRIEF DESCRIPTION OF DRAWINGS

[0081] In order to more clearly illustrate the technical solutions of the embodiments of the application, the following will briefly introduce the drawings needed to be used in the embodiments. It should be understood that the following drawings only show some embodiments of the application, and therefore should not be regarded as a limitation to the scope, and for those skilled in the art, other related drawings can also be obtained without creative labor on the basis of these drawings.

[0082] Figure 1 The framework schematic diagram of the existing multi-source heterogeneous data inference and prediction method provided for the first embodiment.

[0083] Figure 2 The framework schematic diagram of the confidence-guided adaptive graph representation enhanced learning method provided for the first embodiment.

[0084] Figure 3A technical framework schematic diagram of a confidence-guided adaptive graph representation reinforcement learning method provided for Embodiment One.

[0085] The present application will be further described below in conjunction with the accompanying drawings and specific embodiments. DETAILED DESCRIPTION

[0086] For the purpose, technical solutions and advantages of the embodiments of the present application to be clearer, the technical solutions in the embodiments of the present application will be described clearly and completely below in conjunction with the accompanying drawings of the embodiments of the present application. Obviously, the described embodiments are part of, rather than all of, the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative work, fall within the protection scope of the present application. Therefore, the following detailed description of the embodiments of the present application provided in the accompanying drawings is not intended to limit the scope of the claimed present application, but merely represents selected embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative work, fall within the protection scope of the present application.

[0087] Embodiment One

[0088] Embodiment One of the present application provides a confidence-guided adaptive graph representation reinforcement learning method, which can be implemented by a confidence-guided adaptive graph representation reinforcement learning device (hereinafter referred to as a reinforcement learning device), in particular, by one or more processors in the reinforcement learning device.

[0089] In the present embodiment, the reinforcement learning device can be an electronic device equipped with a processor, the processor being provided with a computer program of the confidence-guided adaptive graph representation reinforcement learning method and the computer program being executable, such as a computer, a smart phone, a smart tablet, a workstation, etc., which are not limited herein.

[0090] As Figure 1As shown, the existing graph representation learning model inference prediction method mainly includes direct statistical method, probabilistic graph model and graph representation learning based method. The existing method often depends on the majority vote result at initialization, which may contain a lot of noise information, resulting in systematic deviation of the final inference result. Although the traditional graph neural network can capture global feature information, it performs poorly in processing local feature information, making it difficult for the model to capture subtle differences in data instances. The noise and sparsity problems in the data further exacerbate the difficulty of training the graph neural network, making it difficult for the model to extract real and effective information from noisy data. In addition, the inherent sparsity and noise problems of electronic digital data pose a great challenge to existing methods. When the annotation density is low, methods based on confusion matrix such as DS are difficult to generate accurate estimates. In high-density annotation scenarios, the proportion of conflicting answers increases, and traditional methods are difficult to effectively handle. Especially when there are malicious workers or workers with extremely poor labeling ability, noisy labels will seriously interfere with the model training process, reducing the accuracy of electronic digital data inference.

[0091] In view of the limitations of the existing multi-source heterogeneous electronic digital data inference technology, the present application aims to solve the following core technical problems:

[0092] (1) Graph structure noise suppression and local feature protection problem: the existing graph representation learning method (such as GOVERN) adopts a random graph enhancement strategy, which is easy to introduce noise edges and destroy key local structures. The present application needs to solve the technical problem of how to effectively suppress the noise interference in the electronic digital data graph without destroying the local semantics of the task. Specifically, a mechanism is needed to distinguish between important connections and noise connections, avoid the destruction of the graph structure by traditional random disturbance strategies, and ensure that the model can accurately capture the local task features that are crucial to electronic digital data.

[0093] (2) Heterogeneous information fusion and representation learning problem: existing methods are difficult to effectively fuse multiple sources of heterogeneous information such as worker attributes, task features and annotation relationships. The present application needs to solve the technical challenge of how to learn fine-grained representations of workers and tasks through a dual-channel heterogeneous adaptive network. Especially in the sparse annotation scenario, a representation learning mechanism is needed that can make full use of limited annotation information while mining deep associations, overcoming the lack of representation ability of traditional single-channel networks in processing heterogeneous information.

[0094] (3) Dynamic environment adaptation and relationship modeling problem: In view of the problem that TiReMGE and other algorithms depend on preset electronic digital data relationship and are difficult to adapt to dynamic change environment, the technical problem of how to realize self-adaptive modeling of electronic digital data relationship needs to be solved. A graph structure learning mechanism independent of artificial preset relationship is needed to be designed, so that the model can automatically capture the dynamic association between workers and tasks and their internal changes with the environment, and improve the applicability of the model in real complex electronic digital data scenarios.

[0095] (4) Robust reasoning problem under noisy electronic digital data: The performance of existing methods significantly decreases when facing high proportion of noise labels. The key problem of how to realize accurate electronic digital data results in a highly noisy environment needs to be solved. By introducing a label correction mechanism, the defects of traditional methods that are sensitive to noise and prone to error propagation need to be overcome, and the robustness and stability of the model in the presence of malicious workers or low-quality labeling scenarios need to be improved.

[0096] (5) Effective learning problem under sparse data: In view of the inherent sparsity problem of electronic digital data, the technical challenge of how to still effectively realize electronic digital data results under low labeling density needs to be solved. By optimizing the graph representation learning process, the model needs to be able to mine deep patterns from limited electronic digital data information, reduce the dependence on a large amount of electronic digital data, and improve the practical value in data scarce scenarios.

[0097] As shown in Figure 2 , a confidence-guided adaptive graph representation enhanced learning method comprises:

[0098] Obtain a crowdsourcing data set composed of multiple source heterogeneous electronic data with different features;

[0099] Input the crowdsourcing data set into an adaptive graph representation learning model, use a multi-objective collaborative optimization loss function for learning training, obtain a trained adaptive graph representation learning model, and use it for subsequent classification prediction tasks; wherein the adaptive graph representation learning model comprises a node mapping module, a view generation module, a dual-channel adaptive learning module and a label correction module.

[0100] Specifically, as shown in Figure 2 and 3As shown, the crowdsourcing dataset is obtained by preprocessing multi-source heterogeneous electronic data. The preprocessing includes data cleaning, data conversion, data desensitization, and data standardization to convert multi-source heterogeneous data into structured feature representation. For example, the original multi-source heterogeneous electronic data can be classified and arranged by manual annotation, or the data can be batch cleaned and converted by automated scripts to provide a reliable input basis for subsequent graph structure construction, effectively avoiding the problem of relying on data distribution assumptions. In practical applications, electronic digital data often comes from a multi-source heterogeneous environment, such as sensor failure or user input error, which can easily introduce interference. If not handled, it will directly distort the subsequent confidence calculation and edge pruning process.

[0101] Data conversion can convert heterogeneous data into structured representation to achieve feature semantic alignment. Due to the original format incompatibility of multi-source data, if not converted, the feature vector cannot be effectively spliced, affecting the construction of the mapping relationship between the worker nodes and the task nodes in the bipartite graph. In practical applications, data desensitization refers to a technical method for identifying and encrypting sensitive information (such as personal identity data), which can be implemented using encryption algorithms or data masking techniques to retain data usability while avoiding privacy leakage risks. This is a common compliance requirement for electronic digital data. If the desensitization operation is ignored, it may cause legal problems and pollute the crowdsourcing dataset, thereby interfering with the reliability of label correction. In addition, data standardization can be understood as a technical means for normalizing features to eliminate dimensional differences. It can make the feature scales consistent through linear or nonlinear transformation. Because the original value ranges of different features (such as price and timestamp) are vastly different, if not standardized, it will exacerbate the instability of model training, leading to unbalanced weight distribution in the feature extraction stage, ultimately affecting the accuracy of the initial label generated by the predictor.

[0102] Through the above technical solutions, the problems of noise interference, format inconsistency, sensitive information leakage risk, and feature scale difference in the original electronic digital data can be significantly improved, thereby improving the quality of the crowdsourcing dataset and providing a guarantee for the accuracy of subsequent graph construction and the robustness of the prediction model.

[0103] In the node mapping module, the crowdsourcing dataset is mapped to construct an original bipartite graph containing worker nodes and task nodes.

[0104] Worker nodes and task nodes are the core entity abstractions of the crowdsourcing dataset. Worker nodes correspond to data annotators / task executors in the crowdsourcing platform (e.g., human annotators of pictures, users filling out questionnaires, participants completing data classification), which are the subjects that perform tasks and produce annotation results / feedback. Task nodes correspond to specific tasks that need to be processed in the crowdsourcing platform (e.g., pictures to be annotated, texts to be classified, information to be verified), which are the objects that are executed, annotated, and produce original data.

[0105] The purpose of this step is to abstract the data relationship into a graph structure. Specifically, the connection relationship between nodes can be represented in the form of an adjacency matrix, or the nodes and their associated information can be stored using a graph database to model complex interactions in sparse association scenarios. Workers and tasks are modeled as heterogeneous nodes, and annotation relationships are abstracted as edges. This structured expression directly targets sparse association scenarios and solves the limitations of traditional methods that cannot capture complex interactions between nodes.

[0106] In the electronic digital data annotation task, the cognitive differences and annotation ability differences of different workers on the same task lead to significant differences in labels. Specifically, each task will receive annotation results from workers with different backgrounds, and these annotations have systematic differences in reliability and accuracy, resulting in multiple different data views. In the electronic digital data system, the data diversity phenomenon caused by the inconsistent labels annotated by workers. Although these annotation results have significant differences, they all have a common core element, the true label corresponding to the task. As an unchanging core, the true label always runs through all different annotation views.

[0107] On this basis, the view generation unit generates two enhanced views: a pruned view and an augmented view to comprehensively capture graph information, and the two views are used as inputs to the graph neural network module to obtain node embeddings.

[0108] In the view generation module, based on the original bipartite graph, a pruned view and an augmented view are generated by dynamic confidence pruning and conservative edge addition.

[0109] Existing methods typically rely solely on generating perturbed subgraphs to capture graph information, which can potentially corrupt the graph's information. In this embodiment, a pruned view and an enhanced view are generated through dynamic confidence pruning and conservative edge augmentation. Dynamic confidence pruning is an operation that removes low-quality edges based on their confidence scores, filtering noisy data while preserving key topological structures. Conservative edge augmentation is an operation that supplements important structures with high-confidence connections, strengthening important relationships within the graph structure. By generating two views, diverse enhanced views can be constructed, enabling the model to extract consistent true label information from diverse labeled views, significantly reducing the impact of label discrepancies on model performance.

[0110] like Figure 2 and Figure 3 As shown, the process of generating the trimmed view is as follows:

[0111] First, based on the original bipartite graph, the maximum class confidence of each edge is calculated using an edge predictor. The expression for the class confidence is:

[0112] = ;

[0113] in, This represents the edge formed by the i-th worker node and the j-th task node. Category confidence; For edge predictors; The weight matrix for worker node features; This represents the feature vector of the i-th worker node; This is the weight matrix for the features of the task nodes; This represents the feature vector of the j-th task node; is the weight matrix for edge features; This is a linear transformation function used to convert the output of the edge predictor into a probability distribution; For the edge eigenvectors; This is the bias term for the edge predictor; For feature splicing operations;

[0114] Then, using the maximum class confidence score for each edge as the confidence score, edges belonging to classes with higher probabilities are selected to retain high-quality labeled edges. The expression is as follows:

[0115] ;

[0116] in, For the edge The confidence level; C represents the set of all possible label categories; c represents the category; Indicates category label; represents taking the maximum value for all possible label classes c; represents the class confidence function;

[0117] Then, dynamic edge pruning is performed, and a confidence sensitivity factor is used to adaptively adjust the edge pruning probability to control the sensitivity of the confidence edge. The expression of the dynamic edge pruning probability is:

[0118]

[0119] wherein, represents the dynamic pruning probability of edge ; is a very small constant; is a confidence sensitivity factor; is a pruning weight coefficient; represents the L2 norm, which is used to measure the difference between features; represents the feature vector of the i-th worker node; represents the feature vector of the j-th task node;

[0120] The edges of the original bipartite graph are pruned based on the dynamic edge pruning probability to obtain the pruned view.

[0121] In practical applications, the edge predictor refers to a model component that can predict edge attributes based on input features. It can be implemented using structures such as multi-layer perceptron, convolutional neural network, or recurrent neural network. Specifically, the softmax function is a commonly used normalization exponential function, which aims to convert model output into a probability distribution, facilitating subsequent confidence evaluation. The feature concatenation operation refers to the operation of connecting multiple feature vectors by dimension, which aims to fuse feature information from different sources and improve the accuracy of confidence evaluation. In addition, the confidence sensitivity factor is a parameter used to adjust the sensitivity of the pruning probability, which can be determined through experimental parameter tuning or adaptive learning, aiming to balance the pruning intensity and the need to retain high-value edges.

[0122] ​The scheme solves the problem of insufficient adaptability caused by fixed threshold pruning through a dynamic confidence pruning mechanism. First, the maximum class confidence is calculated based on the original bipartite graph using an edge predictor. This process fuses worker node features, task node features, and edge features through a feature concatenation operation, so that the confidence evaluation considers both node attributes and edge relationships, avoiding evaluation bias caused by relying on a single feature. Then, the maximum class confidence is used as the edge confidence, and the highest quality label is focused by taking the maximum value of all class probabilities, ensuring that the retained edges correspond to the most reliable labeling results, effectively filtering out high-confidence labeling edges. Through the above technical solution, adaptive pruning of the original bipartite graph is achieved, effectively retaining high-value edges while removing low-quality labeling edges, improving the accuracy of graph structure representation and the robustness of subsequent prediction.

[0123] Specifically, the generation process of the enhanced view is:

[0124] The maximum class confidence calculated by the edge predictor for each edge is used as the probability of edge addition, and the formula is:

[0125] ;

[0126] Wherein, represents the edge addition probability of ; is the edge addition weight coefficient; is a nonlinear activation function; is the weight matrix of edge addition; T is the transpose symbol; is the confidence of edge ; represents the i-th worker node feature vector; represents the j-th task node feature vector.

[0127] Based on the calculated edge addition probability, the original bipartite graph is edge-added to obtain the enhanced view.

[0128] The purpose of introducing the edge addition probability is to ensure that the newly added edges have high reliability, thereby avoiding the influence of noise data on the graph structure. The generation process of the enhanced view effectively solves the problem of useful edge loss caused by dynamic confidence pruning. While maintaining the purity of the graph structure, by conservatively adding high-probability edges, it not only enriches the information flow, but also provides more complete topological support for subsequent double-channel feature extraction.

[0129] The embodiment adjusts the pruning probability adaptively according to the maximum class confidence of the edge by dynamic confidence pruning, and controls the noise edge removal strength by a confidence sensitive factor, so as to avoid information loss caused by excessive pruning; the edge increase is calculated by the edge confidence and feature similarity, and high-confidence connections are supplemented around the key structure. This double-view generation mechanism not only filters low-quality annotation noise, but also strengthens important topological relationships, overcoming the defect of introducing noise by random disturbance.

[0130] In the double-channel adaptive learning module, according to the pruning view and the enhanced view, the worker nodes and the task nodes are respectively subjected to feature extraction, adaptive information enhancement and learning by a double-channel heterogeneous adaptive network, and an initial prediction label is obtained by a predictor.

[0131] In order to capture the interaction between workers and tasks at the same time, the traditional graph neural network has limitations when processing heterogeneous graphs, and cannot directly construct the heterogeneous relationship between workers and tasks, resulting in the loss of semantic information of nodes. In order to solve this problem, a new double-channel heterogeneous adaptive network is designed.

[0132] Specifically, first, worker feature embeddings, task feature embeddings and edge feature embeddings are extracted according to the pruning view and the enhanced view respectively;

[0133] Then, the transformed weight matrix is used for embedding initialization operation to generate the corresponding initial embedding, and the expression is:

[0134] ;

[0135] ;

[0136] ;

[0137] Among them, 、 、 respectively represent the initial worker feature embedding, the initial task feature embedding and the initial edge feature embedding; 、 、 respectively represent the extracted worker feature embedding, the task feature embedding and the edge feature embedding; 、 、 respectively represent the weight matrix of 、 、 ; 、 、 respectively represent 、 、 Learnable parameters; Indicates feature index; Indicates the total number of features; Presentation layer normalization operation;

[0138] Then, based on the initial embedding, worker nodes for each iteration are obtained through hierarchical information passing and aggregation. and task nodes Hidden embedding representation and Then, in the i-th iteration, the first... The formula for the embedding of a layer is:

[0139] ;

[0140] ;

[0141] in, , The i-th iteration and the i-th iteration are respectively Layer worker node and task nodes Embedded representation; It is a non-linear activation function; , They represent , The neighboring nodes; , They represent , The set of neighboring nodes; , They represent , Another neighboring node; This represents the original attention score calculated for each pair of nodes (i,m); This represents the attention score calculated for each pair of nodes (i,j) and the normalized result; This represents the temperature coefficient, used to control the smoothness of attention distribution; , They represent , The weight matrix; Indicates residual connection; , The i-th iteration and the i-th iteration are respectively Layer worker node and task nodes Embedded representation; For the i-th iteration, the th Embedded representation of layer edges; It is an exponential function.

[0142] Capturing semantic features of hidden embedding representation using linear auxiliary layer combined with multi-layer perception to generate final worker embedding representation and final task embedding , the expression is:

[0143] ;

[0144] ;

[0145] wherein, , respectively represent the final worker embedding, the final task embedding at the i-th iteration; represents an auxiliary linear transformation function; is a multi-layer perception function; , respectively are the weight matrix at the i-th iteration; represents a batch normalization operation; represents a layer normalization operation; represents the hidden worker embedding at the i-th iteration; represents the hidden task embedding at the i-th iteration; represents a learning parameter in the multi-layer perception function; represents the embedding bias at the i-th iteration.

[0146] Based on the final task embedding, a task projection node is obtained by using a basic projection head for mapping;

[0147] The task projection node is locally learned by using an adaptive information enhancer and a learnable scale parameter, and the contribution of the task node feature is adaptively adjusted through back propagation during the training process, to obtain a high-order task feature, so as to reduce the noise information generated by the original task feature, and the expression of the high-order task feature is:

[0148] ;

[0149] ;

[0150] wherein, is a high-order task feature; is a splicing operation; is an element-wise multiplication; is a final task embedding; , , represents a multi-head attention weight matrix; is a standardization scaling factor; is a matrix change parameter; denotes a regularization operation; denotes denotes a weight matrix; denotes denotes a bias; denotes the total number of high-order task features; denotes an adaptive information enhancer;

[0151] generates an updated final task embedding based on the high-order task features, and obtains an initial predicted label using a predictor.

[0152] The predictor comprises a linear transformation layer and a normalization layer, the linear transformation layer is used to map the final task embedding to a low dimension, and the normalization layer is used to obtain the initial predicted label, and the formula is as follows:

[0153]

[0154] wherein, is the initial predicted label; is a linear transformation function; is a normalization function; , denote a normalized weight matrix respectively; , is a normalized bias; is an activation function; is an updated final task embedding; and b denote a weight matrix and a bias of the predictor respectively.

[0155] The dual-channel heterogeneous adaptive network refers to a neural network architecture designed for the heterogeneous characteristics of worker nodes and task nodes, and each channel focuses on feature modeling of a specific type of node. In actual application, the pruning view and the enhancement view are two complementary views generated by dynamically pruning edges and conservatively adding edges to the original bipartite graph, aiming to filter noise edges and supplement high-quality connections, thereby improving the accuracy of feature extraction. Among them, the hierarchical information transmission aggregation refers to a feature aggregation mechanism based on dynamically weighting neighbor information based on edge confidence, which can use an exponential function and a temperature parameter to adjust the weight distribution, ensuring that high-confidence edges dominate in information transmission. The linear auxiliary layer is designed in combination with a multilayer perceptron to capture deep semantic features, which can realize adaptive enhancement of features through nonlinear transformation and standardization operation. The adaptive information enhancer is a component based on a multi-head attention mechanism, which can dynamically adjust feature contribution through a learnable scale parameter, thereby suppressing noise information and strengthening key semantic features.

[0156] ​Then the task projection node is locally learned by using an adaptive information enhancer, feature contribution is dynamically adjusted based on a multi-head attention mechanism and a learnable scale parameter, and weight distribution is optimized through back propagation, which specifically suppresses noise information in the original task features and strengthens key semantic features, significantly improving the robustness of feature representation. Finally, the updated embedding is generated based on the high-order task features and the label is predicted, and multi-scale information is integrated through splicing operation and element-by-element multiplication, so that the prediction process fully integrates the features after adaptive enhancement, thereby outputting more reliable initial prediction labels to lay a high-quality foundation for subsequent label correction. The scheme systematically overcomes the negative effects of noise and sparsity in electronic digital data on feature learning through double-channel separation processing of heterogeneous nodes, confidence-guided information aggregation, and dynamic noise suppression mechanism.

[0157] In the label correction module, the initial prediction label is corrected in combination with the original bipartite graph to obtain a final prediction label.

[0158] In combination with the original bipartite graph, the initial prediction label is corrected, specifically:

[0159] First, a similarity weight matrix is calculated according to the original bipartite graph, then the initial prediction label is weighted and summed using the similarity weight matrix to obtain a weighted label, and then the weighted label is normalized to obtain a final prediction label, the expression is:

[0160] ;

[0161] ;

[0162] Among them, is the initial prediction label corresponding final prediction label; is the similarity weight matrix of label i to sample j; denotes the initial label vector of label j; is the normalization of the label vector; is the total number of samples; denotes a multi-layer perception; is a layer normalization operation; is the cosine similarity of the task projection vectors of labels i and j , ; , is the similarity weight coefficient; is the transpose symbol; is the task projection vector of label ; , Query matrix weight and key matrix weight of attention network, respectively; is a standardized scaling factor; is a temperature parameter for controlling the degree of weight distribution between labels.

[0163] The iterative training can be divided into two stages. If the current iteration period does not exceed the set threshold weighted_epoch, the initial label is the final label. If the current iteration period exceeds weighted_epoch, the label correction mechanism is triggered.

[0164] This step corrects the initial predicted label in combination with the original bipartite graph. A dynamic weight matrix is generated by calculating the cosine similarity of the task projection vector, and the initial label is weighted and normalized. This iterative correction mechanism based on feature similarity can dynamically optimize the label quality according to the sample relevance, significantly improving the inference robustness in noisy environments.

[0165] This embodiment realizes accurate inference of multi-source conflict data through the synergistic effect of confidence-guided view generation, heterogeneous feature adaptive enhancement, and dynamic label correction, systematically solving the noise interference and dynamic adaptation problems in multi-source heterogeneous data inference, and generating more accurate labels.

[0166] In a preferred embodiment, it also includes: using a multi-objective collaborative optimization loss function for optimization training to maximize the similarity between positive sample pairs and minimize the similarity between negative sample pairs;

[0167] The multi-objective collaborative optimization loss function includes the optimization loss function between worker nodes and task nodes , the adaptive information enhancement loss function , and the label loss function , the expression is:

[0168] ;

[0169] The expression of the optimization loss between worker nodes and task nodes is:

[0170] ;

[0171] ;

[0172] Among them, , are the worker node set and the task node set, respectively; , respectively represent the i-th worker node embedding under the q-th and k-th view; respectively represent the j-th worker node embedding under the q-th and k-th view; respectively represent the i-th task node embedding under the q-th and k-th view; is a temperature parameter; represents a similarity matrix of two vectors; respectively represent two projection vectors; is a weight matrix; is the j-th query vector under the k-th node; , is a similarity weight coefficient;

[0173] Adaptive information enhancement loss function is the weight of the task feature adaptively adjusted by the Infonce loss, and the expression is:

[0174]

[0175] wherein, respectively represent the high-order task feature of the i-th task node under the q-th and k-th view respectively represent the high-order task feature of the j-th task node under the q-th and k-th view.

[0176] The label loss function is a loss function for evaluating the difference between the predicted label and the original label, which is calculated by cross-entropy, and the expression is:

[0177] ;

[0178] wherein, is the label loss function weight; is the task node set; , represents the i-th label based on the pruning view and the enhanced view, respectively. represents the original label.

[0179] The embodiment introduces a multi-objective collaborative optimization loss function, which systematically improves the consistency and noise robustness of node representation in the training process, thereby providing a more reliable supervision signal for label correction. The multi-objective collaborative optimization mechanism solves the problem of lack of collaborative consistency optimization of worker node and task node representation, reduces the sensitivity of the similarity weight matrix based on the task projection vector to noise data, and at the same time suppresses the noise influence under sparse annotation by adaptively adjusting the task feature weight, thereby significantly improving the robustness and label correction effect of the model in the dynamic multi-source heterogeneous data environment.

[0180] Compared with the prior art, the present application has the following beneficial effects:

[0181] The present application dynamically optimizes the graph structure through the confidence-guided view generation mechanism, and realizes feature extraction and information enhancement by combining a dual-channel heterogeneous adaptive network, effectively solving the prediction accuracy problem in the presence of noise data interference and sparse annotation scenarios, improving the prediction accuracy in the presence of noise data and sparse annotation scenarios, enhancing the adaptive ability of the model to dynamic environments, and improving the local feature capture capability.

[0182] The present application calculates the confidence value of the edge through the edge predictor, and adaptively prunes the low-confidence edge by combining the confidence sensitivity factor, thereby realizing accurate identification and removal of noise connections. Unlike traditional random enhancement methods, the present application adopts a "conservative edge increase" strategy, which only increases edge connections in areas with high confidence, thereby maintaining the semantic consistency of the graph structure and enriching the diversity of local structure. Ablation experiments show that the model performance decreases by 9.51% after removing the confidence-guided graph enhancement module, verifying the key contribution of this module to the overall system performance.

[0183] The dual-channel structure can model the features of the worker and task nodes respectively, and realize efficient fusion of heterogeneous information through semantic alignment and message passing mechanism. The adaptive information enhancer in the network uses a learnable scale parameter to dynamically weight the task node features, improving the local semantic capture accuracy. Experiments show that the model performance on all datasets decreases significantly after removing the dual-channel structure, indicating that this module can effectively alleviate the information loss problem between heterogeneous data.

[0184] The label correction module adopts a dynamic weighting correction mechanism based on feature similarity, and adjusts the smoothness of the label weight distribution through a temperature parameter, effectively suppressing the influence of abnormal annotations. This module makes the generated pseudo-labels more consistent with the true distribution of the samples, improving the robustness of the model under the interference of malicious annotators.

[0185] The multi-view collaborative optimization part generates two complementary views, a confidence pruning view and a confidence enhancement view, and collaboratively optimizes based on the InfoNCE loss, thereby realizing cross-view feature alignment and feature invariance modeling, and improving the generalization ability and stability of the model on unseen data.

[0186] Embodiment two

[0187] The second embodiment of the present application also provides a confidence-guided adaptive graph representation enhancement learning device, which comprises a memory and a processor, the memory stores a computer program, and the computer program can be executed by the processor to realize the confidence-guided adaptive graph representation enhancement learning method as described above.

[0188] Embodiment three

[0189] The third embodiment of the present application also provides a computer readable storage medium, wherein the computer readable storage medium stores computer readable instructions, and the computer readable instructions are executed by a processor of a device where the computer readable storage medium is located to implement the confidence guided adaptive graph representation reinforcement learning method as described above.

[0190] In several embodiments provided by the embodiments of the present application, it should be understood that the disclosed apparatus and method can also be implemented by other manners. The apparatus and method embodiments described above are only illustrative, for example, the flowchart in the drawings shows the possible implementation architecture, function and operation of the apparatus, method and computer program product according to the embodiments of the present application. In this regard, each block in the flowchart or block diagram can represent a module, a program segment or a part of code, which contains one or more executable instructions for implementing the specified logical function. It should also be noted that in some alternative implementation manners, the functions noted in the blocks can also occur in different order from that noted in the drawings. For example, two consecutive blocks can actually be executed substantially in parallel, and sometimes they can be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagram and / or flowchart, and the combination of blocks in the block diagram and / or flowchart, can be implemented by a dedicated hardware-based system for implementing the specified function or action, or can be implemented by a combination of dedicated hardware and computer instructions.

[0191] In addition, each functional module in each embodiment of the present application can be integrated together to form an independent part, or each module can exist independently, or two or more modules can be integrated to form an independent part.

[0192] If the functions are implemented in the form of software function modules and sold or used as independent products, they can be stored in a computer readable storage medium. Based on such understanding, the technical solutions of the present application essentially or the parts that contribute to the prior art or parts of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for causing a computer device (which can be a personal computer, an electronic device, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present application. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), magnetic disk or optical disk, and various media that can store program codes. It should be noted that in this paper, the terms "include", "contain" or any other variants thereof are intended to cover non-exclusive inclusion, so that the processes, methods, articles or devices that include a series of elements not only include those elements, but also include other elements not explicitly listed, or include elements inherent to such processes, methods, articles or devices. Without more limitations, the element defined by the statement "includes a" does not exclude the presence of other identical elements in the process, method, article or device that includes the element.

[0193] The terms used in the embodiments of the present application are only for the purpose of describing specific embodiments and are not intended to limit the present application. The singular forms "a", "said" and "the" used in the embodiments of the present application and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise.

[0194] It should be understood that the term "and / or" used herein is only to describe the association relationship of the associated objects, which means that there can be three relationships, for example, A and / or B, which can represent the three cases of A alone, A and B together, and B alone. In addition, the character " / " in this paper generally represents that the front and rear associated objects are a "or" relationship.

[0195] Depending on the context, the word "if" as used herein can be interpreted as "when" or "upon" or "in response to determining" or "in response to detecting". Similarly, depending on the context, the phrase "if it is determined" or "if (a stated condition or event) is detected" can be interpreted as "when it is determined" or "in response to determining" or "when (a stated condition or event) is detected" or "in response to detecting (a stated condition or event)".

[0196] The "first / second" mentioned in the embodiments are only to distinguish similar objects, and do not represent a specific order for the objects. Understandably, the "first / second" can be interchanged in a specific order or sequence as appropriate. It should be understood that the objects distinguished by "first / second" can be interchanged as appropriate, so that the embodiments described herein can be implemented in an order other than those illustrated or described herein.

[0197] The above only describes the preferred embodiments of the present application and is not intended to limit the present application. Those skilled in the art can make various modifications and changes to the present application. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A confidence-guided self-adaptive graph representation reinforcement learning method, characterized in that, The method comprises the following steps: acquiring a crowd-sourced data set composed of multi-source heterogeneous electronic data of different characteristics; inputting the crowd-sourced data set into an adaptive graph representation learning model, learning and training using a multi-objective collaborative optimization loss function, and obtaining a trained adaptive graph representation learning model for subsequent classification and prediction tasks; wherein the adaptive graph representation learning model comprises a node mapping module, a view generation module, a dual-channel adaptive learning module, and a label correction module; the crowd-sourced data set is mapped and constructed into an original bipartite graph containing worker nodes and task nodes through the node mapping module; based on the original bipartite graph, the view generation module is used to generate a pruned view and an enhanced view through dynamic confidence pruning and conservativeness edge addition; according to the pruned view and the enhanced view, the dual-channel heterogeneous adaptive network of the dual-channel adaptive learning module is used to extract features, adaptively enhance information, and learn the worker nodes and task nodes, and an initial prediction label is obtained through a predictor; combined with the original bipartite graph, the label correction module is used to correct the initial prediction label to obtain a final prediction label; wherein the generation process of the pruned view is as follows: first, based on the original bipartite graph, the maximum class confidence of each edge is calculated using an edge predictor, and the expression of the class confidence is: = ； wherein, denotes an edge between the ith worker node and the jth task node a class confidence; is an edge predictor; is a weight matrix for worker node features; denotes the ith worker node feature vector; is a weight matrix for task node features; denotes the jth task node feature vector; is a weight matrix for edge features; is a linear transformation function to convert the output of the edge predictor to a probability distribution; is a feature vector for edge ; is a bias term for the edge predictor; is a feature concatenation operation; then, the maximum class confidence of each edge is taken as the confidence, and the edges of the higher probability class are selected to retain high-quality labeled edges, and the expression is: ； wherein, is the edge confidence; C denotes a set of all possible label classes; c denotes a class; denotes a class label; denotes taking the maximum over all possible label classes c; denotes a class confidence function; then, dynamic edge pruning is performed, and the confidence sensitivity factor is adaptively adjusted to control the sensitivity of the confidence edge, and the expression of the dynamic edge pruning probability is: ； wherein, denotes the edge of the dynamic pruning probability; is a very small constant; is a confidence sensitivity factor; is a pruning weight coefficient; denotes the L2 norm, which is used to measure the difference procedure of features; based on the dynamic edge pruning probability, the edges of the original bipartite graph are pruned to obtain the pruned view; the generation process of the enhanced view is as follows: the maximum class confidence of each edge calculated by the edge predictor is taken as the probability of edge addition, and the formula is: ； wherein, represents the probability of edge increase; is the edge increase weight coefficient; is a nonlinear activation function; is the weight matrix of edge increase; T is the transpose symbol; is the confidence of edge ; based on the calculated edge addition probability, the edges of the original bipartite graph are added to obtain the enhanced view.

2. The confidence-guided adaptive graph representation reinforcement learning method according to claim 1, characterized in that the crowd-sourced data set is obtained by preprocessing multi-source heterogeneous electronic data; the preprocessing includes data cleaning, data conversion, data desensitization, and data standardization.

3. The confidence-guided adaptive graph representation reinforcement learning method according to claim 1, characterized in that, According to the pruned view and the enhanced view, the dual-channel heterogeneous adaptive network is used to extract features, adaptively enhance information, and learn the worker nodes and task nodes, specifically: first, the worker feature embedding, task feature embedding, and edge feature embedding are extracted from the pruned view and the enhanced view, respectively; then, the initial embedding corresponding to the transformed weight matrix is generated through embedding initialization operation, and the expression is: ； ； ； wherein, , , respectively represent an initial worker feature embedding, an initial task feature embedding and an initial edge feature embedding; , , respectively represent an extracted worker feature embedding, a task feature embedding and an edge feature embedding; , , respectively represent , , a weight matrix of , , respectively represent , , a learnable parameter of denotes a feature index; denotes a total number of features; denotes a layer normalization operation; Then, based on the initial embedding, worker nodes for each iteration are obtained through hierarchical information passing and aggregation. and task nodes Hidden embedding representation and Then, in the i-th iteration, the first... The formula for the embedding of a layer is: ； ； in, , The i-th iteration and the i-th iteration are respectively Layer worker node and task nodes Embedded representation; It is a non-linear activation function; , They represent , The neighboring nodes; , They represent , The set of neighboring nodes; , They represent , Another neighboring node; This represents the original attention score calculated for each pair of nodes (i,m); This represents the attention score calculated for each pair of nodes (i,j) and the normalized result; This represents the temperature coefficient, used to control the smoothness of attention distribution; , They represent , The weight matrix; Indicates residual connection; , The i-th iteration and the i-th iteration are respectively Layer worker node and task nodes Embedded representation; For the i-th iteration, the th Embedded representation of layer edges; It is an exponential function; capturing semantic features of hidden embedding representations using linear auxiliary layers in conjunction with a multi-layer perceptron to generate final worker embeddings representations and final task embeddings , the expression being: ； ； wherein, , respectively denote the final worker embedding, the final task embedding at the i-th iteration; denotes an auxiliary linear transformation function; is a multi-layer perceptron function; , respectively denote the weight matrix at the i-th iteration; denotes a batch normalization operation; denotes the hidden worker embedding at the i-th iteration; denotes the hidden task embedding at the i-th iteration; denotes a learning parameter in the multi-layer perceptron function; denotes the embedding bias at the i-th iteration; based on the final task embedding, the task projection node is obtained by mapping using the basic projection head; the adaptive information enhancer and the learnable scale parameter are used to locally learn the task projection node, and the contribution of the task node features is adaptively adjusted through back propagation during the training process, and high-order task features are obtained to reduce the noise information generated by the original task features; the expression of the high-order task features is: ； ； wherein, is a high-order task feature; is a concatenation operation; is an element-wise multiplication; is a final task embedding; , , denotes a multi-headed attention weight matrix; is a standardization scaling factor; is a matrix variation parameter; denotes a regularization operation; denotes a weight matrix; denotes a bias; denotes a total number of high-order task features; denotes an adaptive information enhancer; is a linear transformation function; An updated final task embedding is generated based on the high-order task feature, and an initial prediction label is obtained by using a predictor.

4. The confidence-guided adaptive graph representation reinforcement learning method according to claim 3, characterized in that, The predictor comprises a linear transformation layer and a normalization layer, the final task embedding is mapped to a low dimension by the linear transformation layer, and the initial prediction label is obtained by normalization, and the formula is: ； wherein, is an initial prediction label; is a linear transformation function; is a normalization function; , respectively denote a normalized weight matrix; , is a normalized bias; is an activation function; is an updated final task embedding; , b respectively denote a weight matrix and a bias of the predictor.

5. The confidence-guided adaptive graph representation reinforcement learning method according to claim 4, characterized in that, In combination with the original bipartite graph, label correction is performed on the initial prediction label, specifically: First, a similarity weight matrix is calculated according to the original bipartite graph, then the initial prediction label is weighted and summed by using the similarity weight matrix to obtain a weighted label, and then the weighted label is normalized to obtain a final prediction label, and the expression is: ； ； wherein, is the initial predicted label is the corresponding final predicted label. is the similarity weight matrix for label i to sample j; is the initial label vector for label j; is the normalization of the label vector; is the total number of samples; is the multi-layer perception; is the layer normalization operation; is the task projection vector for label i, j , is the cosine similarity of , is the similarity weight coefficient; is the transpose symbol; is the task projection vector for label ; , are the query matrix weight and the key matrix weight of the attention network, respectively; is the standardization scaling factor; is the temperature parameter for controlling the degree of weight distribution between labels.

6. The confidence-guided adaptive graph representation reinforcement learning method according to claim 5, characterized in that, A multi-objective collaborative optimization loss function is used for learning and training to maximize the similarity between positive sample pairs and minimize the similarity between negative sample pairs; Wherein, the multi-target cooperative optimization loss function Including optimization loss functions between worker nodes and between task nodes , adaptive information enhancement loss function And label loss function , the expression is: ； Optimization loss between worker nodes and between task nodes The expression for the optimization loss is: ； ； wherein, , are a set of worker nodes and a set of task nodes, respectively; , represent the i-th worker node embedding under the q-th and k-th view, respectively; represent the j-th worker node embedding under the q-th and k-th view, respectively; represent the i-th task node embedding under the q-th and k-th view, respectively; is a temperature parameter; represents a similarity matrix of two vectors; represent two projection vectors, respectively; is a weight matrix; is the j-th query vector under the k-th node; , is a similarity weight coefficient; Adaptive information enhancement loss function is an Infonce loss used to adaptively adjust the weight of the task feature, and the expression is: ； wherein, respectively represent the high-order task features of the i-th task node under the q-th, k-th view; respectively represent the high-order task features of the j-th task node under the q-th, k-th view; The label loss function is calculated by cross-entropy to calculate the difference between the original label and the prediction label of different views, and the expression is: ； wherein, is a label loss function weight; is a set of task nodes; , denotes the i-th label prediction based on the pruned view and the augmented view, respectively; denotes the original label.

7. A confidence-guided adaptive graph representation reinforcement learning device, characterized in that, The computer readable storage medium stores computer readable instructions, and the computer readable instructions are executed by the processor of the device to which the computer readable storage medium belongs to realize the confidence guided adaptive graph representation reinforcement learning method according to any one of claims 1-6.

8. A computer-readable storage medium, characterized in that, The computer readable storage medium stores computer readable instructions, and the computer readable instructions are executed by the processor of the device to which the computer readable storage medium belongs to realize the confidence guided adaptive graph representation reinforcement learning method according to any one of claims 1-6.

Images