Typical characterization and memory guided heteromorphism perception graph characterization learning method
Through the method of typical characterization and memory guidance, the node characterization learning of GNN on the heterogram is optimized, which solves the problem of insufficient performance of traditional GNN on the heterogram data, and improves the quality of node characterization and the effect of downstream tasks.
Patent Information
- Application Number
- CN202411137150.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-19
- Publication Date
- 2024-12-10
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
The traditional GNN model cannot adapt to the heterogeneous graph data, resulting in a reduced distinction in node characterization in the fields of fraud detection and other fields, affecting the downstream classification effect.
Using the heterogeneous perceptual graph characterization learning method of class typical representation and memory-guided, the message delivery process is optimized by introducing global class typical representation and learnable edge weight generators, and the message delivery process is transmitted through global class typical memory to improve nodes. Characterized mass.
Learn high-quality node representation on the heterogram data, improve the effect of GNN in downstream graph analysis tasks, and enhance the distinction of node representation and the ability to capture global information.
Smart Images
Figure CN119106703A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of graph representation learning, and in particular to a heterogamous perceptual graph representation learning method guided by typical representation and memory. Background Art
[0002] Graph representation learning represented by GNN (Graph Neural Network) learns node-level, edge-level or graph-level representations in an end-to-end manner, which can be applied to different downstream tasks. Node representation can be used for node classification, which is a very widely used downstream task. For example, it has application value in social network analysis, recommendation systems, fraud detection and other fields. GNN is a neural network structure specially designed for graph structure data. The node vector representation generated by GNN can utilize both the structural information of the graph and the feature information of the node itself. GNN aims to learn low-dimensional, continuous and meaningful vector representations of nodes, edges or the entire graph in the graph. These representations can effectively capture and encode information and features in the graph structure. As a deep learning method, GNN has low dependence on feature engineering and can be easily learned end-to-end. It has become the de facto standard way of learning graph representation.
[0003] The learning process of most GNN models today can be expressed in the message passing paradigm. The core idea of message passing is to exchange and update information between nodes in the graph, so that each node can aggregate information from its neighbors and update its own state. However, the traditional GNN method is based on the homogeneity assumption, assuming that the graph data input to the GNN is homogeneous graph data, that is, the connected nodes tend to belong to the same category, so the traditional GNN method cannot adapt to heterogeneous graph data. In heterogeneous graph data, some two nodes connected by edges rich in key information are more likely to be dissimilar or of different categories. For example, in a fraud graph, since fraud nodes are a minority of nodes, most of the neighbors of fraud nodes are normal nodes, showing heterogeneity. In this case, when GNN aggregates neighbor information, the representation of fraud nodes is easily overwhelmed by a large amount of normal neighbor node information, resulting in a decrease in the distinction between fraud nodes and normal nodes, affecting the downstream classification effect. Summary of the invention
[0004] The present invention aims to solve at least one of the technical problems existing in the prior art.
[0005] The technical solution of the present invention is: a typical representation and memory-guided heterogamous perceptual map representation learning method, comprising the following steps:
[0006] Step 1: Initialize all learnable neural network model parameters and start training iterations;
[0007] Step 2: In each round of iteration, the graph neural network is used for message passing to aggregate neighbor information for each target node to obtain the target node representation, and a learnable edge weight generator is used to generate edge weights for each edge. The edge weights are used for each target node to aggregate its neighbor information.
[0008] Step 3: Introduce a global class-typical representation, aggregate similar nodes in the graph data used for training into a class-typical representation, which represents the typical pattern of the node representation of this category, and guide the representation of the target node to be as similar as possible to its corresponding class-typical representation. Conversely, guide the representation of the target node to be as dissimilar as possible to the class-typical representation that it does not correspond to, so as to ensure that the optimization direction of the aggregation result of message passing is in a direction that is beneficial to the downstream classification task;
[0009] Step 4: The guidance of the global class typical representation can enable the node representation to capture global information. In order to make the target node pay attention to the rich global information in a more fine-grained manner, the global class typical memory is introduced to interact with the node representation to transmit useful information to the node representation over a long distance, and the final node representation is obtained;
[0010] Step 5: Use the final node representation to perform downstream graph analysis tasks, calculate the total loss, back-propagate and update all learnable neural network model parameters in this method, and complete a training iteration.
[0011] The process of generating edge weight for each edge includes:
[0012] Construct an edge representation for each edge using the representations of the two nodes it connects.
[0013] Input the edge representation into the edge weight generator to get the edge weight.
[0014] The process of constructing an edge representation for each edge using the representations of the two nodes it connects is:
[0015] For an edge, the representations of the two connected nodes, the result of adding the two node representations, and the result of subtracting the two node representations are concatenated.
[0016] After constructing an edge representation for each edge using the representations of the two nodes connected by each edge, the edge representation is input into an edge weight generator composed of a trainable neural network model to obtain the corresponding output as the edge weight, and the optimization of the edge weight is guided by using heterogeneity to calculate the edge weight loss value. The optimization goal of the edge weight loss value is to strengthen the weight size of the edges connecting nodes of the same category and reduce the weight size of the edges connecting nodes of different categories.
[0017] The process of generating and using global class typical representations is as follows:
[0018] Aggregate nodes of the same category into a typical representation of the class, which represents the typical pattern of the node representation of this category;
[0019] Calculate the similarity between the node representation and the global class-typical representation of each category;
[0020] Calculating the class-typical representation loss guides the node representation to be closer to the global class-typical representation of its corresponding category and away from the global class-typical representation of other categories.
[0021] The process of interaction between global typical memory and node representation is:
[0022] The global class typical memory consists of several learnable vectors, which are used to calculate the similarity scores with all node representations one by one;
[0023] For a target node, the similarity score calculated between the target node and the global class typical memory is used to weight the global class typical memory and add it to the target node representation;
[0024] Calculating the class-typical memory similarity loss value enables the global class-typical memory to learn knowledge from the global class-typical representation;
[0025] Calculating diversity loss makes the global class representative memory contain richer knowledge;
[0026] Calculating the coefficient of variation loss allows global class representative memories to be selected uniformly;
[0027] The class-typical memory loss is obtained by adding the class-typical memory similarity loss value, diversity loss value, and coefficient of variation loss value.
[0028] The process of making the global class-typical memory learn knowledge from the global class-typical representation is:
[0029] Calculate the similarity between the global class-typical representation and the global class-typical memory for each node category;
[0030] The corresponding class-typical memory similarity loss value is calculated to improve the similarity between the class-typical memory that is most similar to each global class-typical representation and the corresponding global class-typical representation.
[0031] In order to prevent all memories from becoming similar to the global class-typical representation, the Euclidean distance between each memory in the global class-typical memory and all other memories is calculated, that is, the Euclidean distance between the global class-typical memories is calculated pairwise to obtain the class-typical memory diversity loss value, so that the global class-typical memories are different;
[0032] For each representation in the global class representative representation, the Euclidean distance is calculated with all other representations to obtain the class representative representation diversity loss value, so that the global class representative representations are different;
[0033] The above two loss values, namely the class-typical memory diversity loss value and the class-typical representation diversity loss value, are added together to obtain the diversity loss.
[0034] The coefficient of variation loss is calculated by calculating the similarity between the node representation and the global class typical memory. A lower coefficient of variation indicates that the selection of the global class typical memory is more balanced, that is, more memory is used.
[0035] Use the obtained node representation to perform downstream graph analysis tasks, and compare the task analysis results with the correct answer to obtain the downstream graph analysis task loss, and add this loss to all the above-mentioned related loss values - edge weight loss, class-typical representation loss, and class-typical memory loss to obtain the total loss;
[0036] Use the total loss to backpropagate and update all learnable neural network model parameters in this method.
[0037] In the work, the present invention optimizes the message passing process and results of the graph neural network. First, the ideal edge weights are generated according to the homogamy hypothesis through heterogamy-guided edge weight generation to improve the aggregation process of naive message passing; then, global information is introduced by constructing a global class-typical representation to guide the optimization direction of the generated edge weight message passing aggregation results; finally, the global class-typical memory transmits long-distance information, and the global class-typical memory learns from the global class-typical representation, and finally learns high-quality node representations on heterogamy graph data.
[0038] The present invention can solve the problem that the traditional GNN model cannot adapt to the node representation learning on heterogeneous graphs, so as to finally learn high-quality node representations and improve the effect of GNN in downstream graph analysis tasks. BRIEF DESCRIPTION OF THE DRAWINGS
[0039] Figure 1 It is a principle block diagram of the present invention;
[0040] Figure 2 Schematic diagram of the constraints of typical representation and memory of global classes;
[0041] Figure 3 The figure is a flow chart of the method of the present invention. DETAILED DESCRIPTION
[0042] The present invention will be further described in detail below in conjunction with the accompanying drawings and specific implementation methods.
[0043] The present invention is mainly carried out in different matching pictures or the same matching pictures environment, such as Figure 1 As shown, it includes the following parts: edge weight generator, global class typical representation, and global class typical memory.
[0044] like Figure 3As shown, the present invention provides a typical representation and memory-guided heterogamous perceptual map representation learning method, comprising the following steps:
[0045] Step 1: Initialize all learnable neural network model parameters required by this method and start training iterations;
[0046] Step 2: In each round of iteration, GNN is used for message passing to aggregate neighbor information for each target node to obtain the target node representation, and a learnable edge weight generator is used to generate edge weights for each edge. The edge weights are used to aggregate the neighbor information of each target node. Under the mechanism of step 2, the edge weights in the message passing aggregation process are generated, and the message passing of GNN is optimized from the aggregation process of message passing. In order to further guide the message passing to optimize the node representation results, further optimization will be considered from the perspective of the message passing aggregation results.
[0047] Step 3: Introduce a global class-typical representation, that is, aggregate similar nodes in the graph data used for training into a class-typical representation, which represents the typical pattern of the node representation of this category, and guide the representation of the target node to be as similar as possible to its corresponding class-typical representation, and vice versa - guide the representation of the target node to be as dissimilar as possible to its non-corresponding class-typical representation, so as to ensure that the optimization direction of the aggregation result of message passing is beneficial to the downstream classification task;
[0048] Step 4: The guidance of the global class typical representation can enable the node representation to capture global information. In order to make the target node pay attention to the rich global information in a more fine-grained manner, the global class typical memory is introduced to interact with the node representation to transmit useful information to the node representation over a long distance, and the final node representation is obtained;
[0049] Step 5: Use the final node representation to perform downstream graph analysis tasks, calculate the total loss, back-propagate and update all learnable neural network model parameters in this method, and complete a training iteration.
[0050] The process of generating edge weight for each edge includes:
[0051] Construct an edge representation for each edge using the representations of the two nodes it connects.
[0052] Input the edge representation into the edge weight generator to get the edge weight.
[0053] The process of constructing an edge representation for each edge using the representations of the two nodes it connects is:
[0054] For an edge, the representations of the two connected nodes, the result of adding the two node representations, and the result of subtracting the two node representations are concatenated.
[0055] After constructing an edge representation for each edge using the representations of the two nodes connected by each edge, the edge representation is input into an edge weight generator composed of a trainable neural network model to obtain the corresponding output as the edge weight. The edge weight is optimized using heterogeneity to calculate the edge weight loss value. The optimization goal of the edge weight loss value is to strengthen the weights of edges connecting nodes of the same category and reduce the weights of edges connecting nodes of different categories.
[0056] The process of generating and using global class typical representations is as follows:
[0057] Aggregate nodes of the same category into a typical representation of the class, which represents the typical pattern of the node representation of this category;
[0058] Calculate the similarity between the node representation and the global class-typical representation of each category;
[0059] The class representative representation loss is calculated to guide the node representation to be closer to the global class representative representation of its corresponding category and away from the global class representative representation of other categories.
[0060] The node representation should be most similar to the global class-typical representation of its corresponding category, and dissimilar to the global class-typical representations of other categories, so that the class-typical representation loss is calculated using the calculated similarity and the expected similarity.
[0061] The process of interaction between global typical memory and node representation is:
[0062] The global class typical memory consists of several learnable vectors, which are used to calculate the similarity scores with all node representations one by one;
[0063] For a target node, the similarity score calculated between the target node and the global class typical memory is used to weight the global class typical memory and add it to the target node representation;
[0064] Calculating the class-typical memory similarity loss value enables the global class-typical memory to learn knowledge from the global class-typical representation;
[0065] Calculating diversity loss makes the global class representative memory contain richer knowledge;
[0066] Calculating the coefficient of variation loss allows global class representative memories to be selected uniformly;
[0067] The class-typical memory loss is obtained by adding the class-typical memory similarity loss value, diversity loss value, and coefficient of variation loss value.
[0068] The process of making the global class-typical memory learn knowledge from the global class-typical representation is:
[0069] Calculate the similarity between the global class-typical representation and the global class-typical memory for each node category;
[0070] The corresponding class-typical memory similarity loss value is calculated to improve the similarity between the class-typical memory that is most similar to each global class-typical representation and the corresponding global class-typical representation.
[0071] The process of calculating diversity loss is:
[0072] For each memory in the global class typical memory, the Euclidean distance is calculated with all other memories, that is, the Euclidean distance is calculated for the global class typical memory pair by pair, and the class typical memory diversity loss value is obtained, so that the global class typical memories are different;
[0073] Similarly, this process is applied to the global class representative representation to obtain the class representative representation diversity loss value, so that the global class representative representations are different;
[0074] The above two loss values, namely the class-typical memory diversity loss value and the class-typical representation diversity loss value, are added together to obtain the diversity loss.
[0075] The process of calculating the coefficient of variation loss is:
[0076] The coefficient of variation loss is calculated based on the similarity scores obtained by calculating the similarity between the node representation and the global class typical memory. A lower coefficient of variation indicates that the selection of the global class typical memory is more balanced, that is, more memory is used;
[0077] Use the obtained node representation to perform downstream graph analysis tasks, and compare the task analysis results with the correct answer to obtain the downstream graph analysis task loss, and add this loss to all the above-mentioned related loss values - edge weight loss, class-typical representation loss, and class-typical memory loss to obtain the total loss;
[0078] Use the total loss to backpropagate and update all learnable neural network model parameters in this method.
[0079] Specifically, step 2 includes:
[0080] Given a graph For the target node , using weighted message passing for neighbor aggregation and node update:
[0081]
[0082]
[0083] in is node v i is the set of neighbors, and ReLU(·) is the activation function. (l) is the bias vector, and is the learnable parameter of the lth layer of GNN, which is used for the aggregation of neighbor node information and the update of its own information. The total number of layers is L; in particular, due to Initialized by the node initial feature vector xi, c ij is the product of the square roots of the node degrees, Used to normalize the information passing of neighbor nodes. is from node v j To node v i The weight on the edge is a learnable parameter used to adjust the importance of neighbor node information and is generated by the edge weight generator.
[0084] The advantage of learnable edge weights is that they do not need to be predefined through a priori analysis. Instead, the most appropriate edge weights can be learned adaptively and end-to-end through training data, thereby improving the node representation learning performance of heterogeneous graph data. i In the case of unknown labels, the most appropriate weight is adaptively found. The present invention adopts a neural network, and here a two-layer MLP network is used as an example to i All neighbor nodes of Generate edge weights before aggregating neighbors at each layer of GNN. i and its neighbors v j The edge weight e between ij is the output of the two-layer MLP; the input of the two-layer MLP is the edge representation Edge Representation By v i With v j Node representation at layer l-1 and After transformation, the specific edge weight generation process is as follows:
[0085]
[0086] in, and is node v i and v j Node representation at layer l-1 and By W self The mapped ones are used to construct edge representations. Is and The edge representation is composed of splicing; here in the intuitive On the basis of and Respectively represent the node v i and v jThe differences between the two nodes and the information that can be obtained after the aggregation of the two nodes increase the information implied in the edge representation and fully perceive the homogeneity / heterogeneity between the nodes. and is a learnable parameter of the two-layer MLP, which is used to train the generated edge weights. Next, the generated edge weights are mapped to the [0, 1] interval to normalize the importance of neighbor node information:
[0087]
[0088] Define the edge training set as The edges that contain the two connected nodes have known category labels, that is, For each edge (v i , v j )∈ε L , define its edge associativity for:
[0089]
[0090] in is the goal to be optimized here, the specific edge weight loss function The definition is as follows:
[0091]
[0092] Furthermore, step 3 includes:
[0093] For the training set of layer l Node, each node category All node representations under are aggregated together (the present invention adopts an average method) to obtain the typical representation of the global class as follows:
[0094]
[0095] For nodes , the node representation of the lth layer can be calculated Similarity to the global class representative representation of its corresponding category Here, Euclidean distance is used as the similarity metric:
[0096]
[0097] Where c is the node v i The category label of .
[0098] Guide the aggregation results to be close to the global typical representation to improve the node representation learning performance of heterogeneous graph data:
[0099]
[0100] Among them, Y L Is the training set , CE(·,·) is the cross entropy loss function. is the Euclidean distance between each node and all global class representative representations, and Y L Calculate the cross entropy loss to get the typical representation loss of the class Optimizing this loss can make the node representation closer to the global class-typical representation of its corresponding category and away from the global class-typical representation of other categories.
[0101] Furthermore, step 4 includes:
[0102] Global Class Typical Memory Network are learnable parameters as follows:
[0103]
[0104] The hyperparameter K is the number of global class typical memories. Different from the fixed number of global class typical representations, more types of global class typical memories and the consideration that nodes of the same category may also have multiple characteristic patterns, so the value of K is greater than
[0105] Next, we will introduce how node representation interacts with the global typical memory network:
[0106]
[0107] in
[0108] like Figure 2 As shown, in order to make the learnable parameter H of the global class typical memory mem It has a more practical meaning, namely, "typicality". The present invention calculates the global class typical representation by node category. and H mem , and reduce the Euclidean distance between the class-typical memory that is most similar to each global class-typical representation and the corresponding global class-typical representation. mem right learning, so that at least one global class-typical memory is similar to a global class-typical representation, ensuring that H mem Memory with practical significance. The specific process is as follows:
[0109]
[0110]
[0111] in, is the typical memory similarity loss.
[0112] Accordingly, in order to prevent all memories from becoming similar to the global class typical representation and failing to mine the potential distribution pattern, the present invention introduces diversity loss , calculate the Euclidean distance for each pair of global class typical memories, so that the global class typical memories are as dissimilar as possible:
[0113]
[0114]
[0115]
[0116] This loss is also used for global class representative representation To ensure the diversity among the typical representations of the global class, the node representations of different categories are encouraged to have higher discrimination.
[0117] In addition, in order to ensure that more memories are useful and avoid the tendency of only a few class-typical memories to be selected during training, the present invention calculates the coefficient of variation (CV). A lower coefficient of variation indicates that the selection of global class-typical memories is more balanced, that is, more memories are used, so the coefficient of variation loss is introduced:
[0118]
[0119] So far, we have obtained a typical memory loss as follows
[0120]
[0121] Finally, By adding it to the update formula of GNN, the interaction between node representation and global class typical memory can be realized. The aggregation update of GNN layer is modified to the following form:
[0122]
[0123] Where α is a hyperparameter used to adjust the importance of global class representative memory.
[0124] Furthermore, step 5 includes:
[0125] After L layers, the final node representation H can be obtained (L) , which is fed into the classifier MLP CLS In the training, the cross entropy loss function is used:
[0126]
[0127] The final overall training objective function is:
[0128]
[0129] Among them, β is a hyperparameter.
[0130] This paper mainly aims at the problem that the traditional GNN model cannot adapt to the node representation learning on heterogeneous graphs, and proposes a typical representation and memory-guided heterogeneous perception graph representation learning method. Since the traditional GNN is based on the homogeneity assumption, that is, it is assumed that the connected nodes in the graph data are of the same category, these graphs are homogeneous graphs; however, many graph data in the real world violate this assumption, and these graph data are heterogeneous graphs, then the traditional GNN method based on this homogeneity assumption cannot learn high-quality node representations on heterogeneous graphs.
[0131] The present invention optimizes from two perspectives: the aggregation process and aggregation results of the core message transmission mechanism of GNN. The aggregation process of the transmission mechanism is optimized through the edge weight generation guided by heterogeneity, and the aggregation results are optimized through global class typical representation and memory guidance. Finally, high-quality node representation is learned to improve the effect of GNN in downstream graph analysis tasks.
[0132] The core mechanism of the present invention mainly includes three parts, namely, a heterogeneity-aware edge weight generator, a global class-typical representation-guided message transmission aggregation optimization, and a global long-distance information transmission of a global class-typical memory interaction.
[0133] The main designs of each part are as follows:
[0134] Part I: Heterogeneity-guided Edge Weight Generation. The main design of the edge weight generator is to use a neural network to generate edge weights for all neighbor nodes of the target node before aggregating neighbors at each layer of the GNN. The core idea is to adaptively and end-to-end learn the most appropriate edge weights through training data, thereby improving the node representation learning performance of heterogeneous graph data. The input of the edge weight generator is the edge representation, which is composed of the transformed node representation, including the concatenation of the representations of two nodes, the difference of the two node representations, and the sum of the two node representations, which respectively represent the difference and the information that may be obtained after aggregation. The generated edge weights are mapped to the interval of 0 to 1 to normalize the importance of the neighbor node information. This design does not require pre-defining edge weights through a priori analysis, but instead adaptively learns the most appropriate edge weights through training data.
[0135] Part II: Guidance of global class-typical representations. The main design of message passing aggregation optimization guided by global class-typical representations is to further optimize the aggregation results of message passing by introducing global class-typical representations. The core idea is to aggregate the representations of nodes of the same type into class-typical representations as the typical pattern of node representations of this category, based on the assumption that nodes of the same type should have similar features and learn similar representations. Then, by guiding the representation of the target node to be as similar as possible to its corresponding class-typical representation, and at the same time dissimilar to other class-typical representations, the aggregation results of message passing are guaranteed to be helpful for downstream classification tasks. This method calculates the similarity between the node representation and the class-typical representation and optimizes the corresponding loss function, so that the node representation is closer to the class-typical representation of its corresponding category, thereby improving the node representation learning performance of heterogeneous graph data.
[0136] Part III: Global Class-Typical Memory Interaction. The main design of global long-distance information transmission of global class-typic memory interaction is to enable node representation to capture longer-distance dependencies and rich global information by introducing a global class-typic memory network. The core idea is to use a learnable global class-typic memory network to interact with node representation and expand the perception range of node representation without increasing the number of layers of GNN to avoid the degradation of representation quality caused by too many layers. Specifically, the representation of each node is dot-producted with the global class-typic memory to obtain a similarity score and the interaction result between the node and the global information. In addition, to ensure that the global class-typic memory is practical and diverse, similarity loss and diversity loss are introduced, and the balance of memory selection is guaranteed by the coefficient of variation loss. Finally, these interaction results are integrated into the update formula of GNN to achieve effective interaction between node representation and global class-typic memory.
[0137] The present invention may also have many other embodiments. Without departing from the spirit and essence of the present invention, those skilled in the art may make various corresponding changes and modifications based on the present invention. These corresponding changes and modifications should all fall within the scope of protection of the claims attached to the present invention.
Claims
1. A typical representation and memory-guided heterogamous perceptual graph representation learning method, characterized in that: The following steps are involved: Step 1: Initialize all learnable neural network model parameters and start training iterations; Step 2: In each round of iteration, use the graph neural network for message passing to aggregate neighbor information for each target node to obtain the target node representation, and use the edge weight generator to generate edge weights for each edge, and use the edge weights for each target node to aggregate its neighbor information; Step 3: Introduce a global class-typical representation, aggregate similar nodes in the graph data used for training into a class-typical representation, represent the typical pattern of the node representation of this category, and guide the representation of the target node to be similar to its corresponding class-typical representation, and guide the representation of the target node to be dissimilar to the class-typical representation that it does not correspond to; Step 4: Introduce global typical memory to interact with node representation to transmit useful information to node representation over a long distance and obtain the final node representation; Step 5: Use the final node representation to perform downstream graph analysis tasks, calculate the total loss, backpropagate and update all learnable neural network model parameters, and complete a training iteration.
2. The method for learning heterogamous perceptual image representations using typical representation and memory guidance according to claim 1, characterized in that: The process of generating edge weight for each edge includes: Construct an edge representation for each edge using the representations of the two nodes it connects. Input the edge representation into the edge weight generator to get the edge weight.
3. The method for learning heterogamous perceptual image representation by typical representation and memory guidance according to claim 2, characterized in that: The process of constructing an edge representation for each edge using the representations of the two nodes it connects is: For an edge, the representations of the two connected nodes, the result of adding the two node representations, and the result of subtracting the two node representations are concatenated.
4. The method for learning heterogamous perceptual image representation by using typical representation and memory guidance according to claim 2, characterized in that: After constructing an edge representation for each edge using the representations of the two nodes connected by each edge, the edge representation is input into an edge weight generator composed of a trainable neural network model to obtain the corresponding output as the edge weight, and the edge weight loss value is calculated by using heterogeneity to guide the optimization of the edge weight.
5. The method for learning heterogametic perceptual image representation by using typical representation and memory guidance according to claim 1, characterized in that: The process of generating and using global class typical representations is as follows: Aggregate nodes of the same category into a typical representation of the class, which represents the typical pattern of the node representation of this category; Calculate the similarity between the node representation and the global class-typical representation of each category; The class representative representation loss is calculated to guide the node representation to be closer to the global class representative representation of its corresponding category and away from the global class representative representation of other categories.
6. The method for learning heterogamous perceptual image representations using quasi-typical representation and memory guidance according to claim 1, characterized in that: The process of interaction between global typical memory and node representation is: The global class typical memory consists of several learnable vectors, which are used to calculate the similarity scores with all node representations one by one; For a target node, the similarity score calculated between the target node and the global class typical memory is used to weight the global class typical memory and add it to the target node representation; Calculating the class-typical memory similarity loss value enables the global class-typical memory to learn knowledge from the global class-typical representation; Calculating diversity loss makes the global class representative memory contain richer knowledge; Calculating the coefficient of variation loss allows global class representative memories to be selected uniformly; The class-typical memory loss is obtained by adding the class-typical memory similarity loss value, diversity loss value, and coefficient of variation loss value.
7. The method for learning heterogamous perceptual image representations using quasi-typical representation and memory guidance according to claim 6, characterized in that: The process of making the global class-typical memory learn knowledge from the global class-typical representation is: Calculate the similarity between the global class-typical representation and the global class-typical memory for each node category; The corresponding class-typical memory similarity loss value is calculated to improve the similarity between the class-typical memory that is most similar to each global class-typical representation and the corresponding global class-typical representation.
8. The method for learning heterogamous perceptual image representations using quasi-typical representations and memory guidance according to claim 6, characterized in that: The process of calculating diversity loss is: For each memory in the global class typical memory, the Euclidean distance is calculated with all other memories to obtain the class typical memory diversity loss value, so that the global class typical memories are different; For each representation in the global class representative representation, the Euclidean distance is calculated with all other representations to obtain the class representative representation diversity loss value, so that the global class representative representations are different; The diversity loss is obtained by adding the class-typical memory diversity loss value and the class-typical representation diversity loss value.
9. The method for learning heterogamous perceptual image representations using quasi-typical representation and memory guidance according to claim 6, characterized in that: The process of calculating the coefficient of variation loss is: The coefficient of variation loss is calculated by calculating the similarity between the node representation and the global class typical memory.
10. The method for learning heterogamous perceptual image representations guided by typical representation and memory according to claim 1, characterized in that: Use the obtained node representation to perform downstream graph analysis tasks, and compare the task analysis results with the correct answer to obtain the downstream graph analysis task loss, and add this loss to the edge weight loss, class-typical representation loss, and class-typical memory loss to obtain the total loss; Use the total loss to backpropagate and update all learnable neural network model parameters.