Bilinear graph neural network prediction and interpretation cooperation method
By employing a bilinear graph neural network prediction and interpretation collaborative method, the problem of inconsistent interpretation in high-risk scenarios of graph neural network models is solved, achieving collaborative optimization of prediction and interpretation, and improving the transparency and credibility of the model.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-31
- Publication Date
- 2026-04-14
AI Technical Summary
Existing graph neural network models struggle to provide transparent and reliable decision-making support in high-risk, high-reliability scenarios, and traditional self-explanatory methods suffer from inconsistencies between explanation and decision-making logic.
A bilinear graph neural network prediction and interpretation co-operation method is adopted. The prediction representation and interpretation signal are generated through the bilinear transformation layer, and a joint objective function is constructed for end-to-end training to achieve co-optimization of prediction and interpretation.
It achieves traceability and interpretability of model prediction results, improves the generalization performance and robustness of the model in complex relational graph data scenarios, enhances the transparency and verifiability of the model, and avoids the performance trade-off bottleneck in traditional methods.
Smart Images
Figure CN121860076A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of artificial intelligence, and in particular to a bilinear graph neural network prediction and interpretation collaborative method. Background Technology
[0002] With the deepening development of artificial intelligence technology, Graph Neural Networks (GNNs) have been widely used in key areas such as drug discovery, social recommendation, and financial risk control due to their powerful modeling capabilities for non-Euclidean data structures (such as social networks, molecular graphs, and knowledge graphs). However, GNN models based on deep neural networks are often regarded as complex "black boxes," and their internal feature propagation, aggregation, and decision-making reasoning processes are difficult for humans to intuitively understand. This lack of interpretability severely undermines users' trust in the model's prediction results and hinders the secure deployment of GNNs in high-risk, high-reliability scenarios (such as medical diagnosis and judicial assistance). Therefore, developing interpretable graph neural networks that can provide transparent and credible decision-making basis has become one of the core issues driving the development of trustworthy artificial intelligence.
[0003] To overcome the aforementioned "black box" dilemma, existing research mainly follows two paths: ex post-explanation methods and self-explanation methods. Ex post-explanation methods, after model training, infer important features by analyzing the correlation between inputs and outputs. However, their explanation process is independent of the model's own reasoning mechanism, risking inconsistency between the explanation and the actual decision-making logic. Conversely, constructing traditional self-explanatory graph neural networks can lead to a fundamental discrepancy between the internal information flow upon which the architecture relies and the information flow upon which the explanation is generated. Optimization objectives may conflict, and the resulting explanation often fails to accurately and truthfully reflect the actual basis for the model's specific predictions. Summary of the Invention
[0004] The purpose of this invention is to provide a bilinear graph neural network prediction and interpretation collaborative method, which aims to identify key nodes, edge structures and feature contribution relationships that affect the model output, thereby providing a traceable and interpretable reasoning process.
[0005] To achieve the above objectives, the present invention adopts the following technical solution: This invention provides a method for coordinated prediction and interpretation in a bilinear graph neural network, comprising: S1: acquiring an input graph to be processed, and extracting the node feature matrix and adjacency matrix from the input graph, wherein the node feature matrix is a matrix containing the feature information of each node in the input graph, and the adjacency matrix is a matrix representing the connection relationship between nodes in the input graph; S2: inputting the preprocessed node feature matrix into the bilinear graph neural network, and generating prediction representation and interpretation signal simultaneously through a bilinear transformation layer; wherein the bilinear transformation layer is a linear transformation module used to simultaneously model the correlation relationship between feature dimensions and the interaction pattern between nodes, the prediction representation is a node feature representation used for downstream task judgment, and the interpretation signal is information reflecting the correlation strength of feature dimensions and the interaction strength between nodes; S3: performing local interpretation analysis and global interpretation analysis based on the interpretation signal output by the bilinear transformation layer, respectively, to obtain local interpretation results and global interpretation results; S4: constructing a joint objective function including prediction loss and interpretation consistency loss, training and optimizing the bilinear graph neural network to obtain a trained bilinear graph neural network, thereby achieving coordinated prediction and interpretation.
[0006] In step S2, a bilinear transformation is performed between the feature dimensions of the node feature matrix to encode the coupling relationship between each feature dimension, resulting in a first-stage hidden feature matrix. A bilinear transformation is then performed between nodes on the first-stage hidden feature matrix to model the feature interaction patterns between nodes, resulting in a second-stage hidden feature matrix. The predicted representation and interpretation signals are generated based on the second-stage hidden feature matrix.
[0007] The specific process of the first-stage bilinear transformation of features includes: mapping each node feature vector in the node feature matrix to the hidden feature space, encoding the correlation between feature dimensions through the weight matrix of the bilinear transformation layer during the mapping process, and outputting the first-stage hidden feature matrix.
[0008] The specific process of the second-stage node bilinear transformation includes: transposing the first-stage hidden feature matrix, calculating the node pair interaction representation of the transpose matrix of the first-stage hidden feature matrix and the first-stage hidden feature matrix through the weight matrix of the bilinear transformation layer, and obtaining the second-stage hidden feature matrix by combining element-level operations.
[0009] In step S3, the specific process of local interpretation analysis includes: constructing a node pair interaction information matrix based on the node interaction strength information in the interpretation signal; performing logical operations on the node pair interaction information matrix and the adjacency matrix of the input graph to obtain the interpretation subgraph connection relationship matrix; and extracting node subsets and edge subsets from the input graph based on the interpretation subgraph connection relationship matrix to form the interpretation subgraph corresponding to the local interpretation result.
[0010] In step S3, the specific process of global interpretation analysis includes: organizing the interaction intensity information corresponding to all dimensions of the feature bilinear transformation output into a third-order interaction matrix; performing a dot product operation between the third-order interaction matrix and the given output direction vector to obtain the global node feature interaction pattern matrix; performing feature decomposition on the global node feature interaction pattern matrix to obtain feature vectors with importance weights, thus forming the global interpretation result.
[0011] The construction process of the joint objective function in step S4 includes: determining the prediction loss based on the downstream task type, which includes node classification and link prediction; calculating the interpretation consistency loss, which is determined based on the similarity between the output representation of the original input graph and the output representation of the interpretation subgraph; and constructing the joint objective function by balancing the prediction loss and the interpretation consistency loss through weight coefficients.
[0012] The calculation process of the interpretation consistency loss includes: calculating the cosine similarity between the predicted representation corresponding to the original input graph and the output representation corresponding to the interpretation subgraph, and constructing the interpretation consistency loss based on the cosine similarity to maximize the similarity between the two.
[0013] After obtaining the second-stage hidden feature matrix, the process also includes an adjacency matrix fusion step: based on the weight information of the node bilinear transformation layer, the potential interaction information matrix of the nodes is analyzed and obtained; the potential interaction information matrix of the nodes is fused with the adjacency matrix of the original input graph to construct a weight matrix for hidden node interaction perception; based on the weight matrix, neighborhood information is aggregated to update the second-stage hidden feature matrix.
[0014] In step S4, the collaborative optimization training adopts an end-to-end training method, which simultaneously optimizes the discriminative performance of the predicted representation and the consistency of the interpreted signal during the training process, so that the interpretation results are consistent with the decision basis of the bilinear graph neural network.
[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This application provides a bilinear graph neural network prediction and interpretation collaborative method that generates prediction representations and interpretation signals simultaneously during the same forward propagation process through a bilinear transformation layer, achieving feature space sharing and deep parameter coupling between the two. This fundamentally ensures the intrinsic consistency between the interpretation signal and the model's decision information flow, enabling the interpretation results to accurately reflect the model's true dependence on feature dimension associations and node interaction patterns. This invention employs a two-stage bilinear transformation: the first stage focuses on encoding associations between feature dimensions, and the second stage strengthens the interaction modeling between nodes, directly capturing high-order, complex feature interaction patterns through the weight matrix of the bilinear operator. Simultaneously, the adjacency matrix fusion step integrates the potential interaction information of nodes with the original graph structure, ensuring that the node representation simultaneously covers structural information and interaction features, effectively reducing spurious feature dependencies. This design not only improves the model's prediction and discrimination capabilities (especially suitable for complex graph data scenarios) but also significantly enhances the model's generalization performance across graph data of different distributions and scales, as well as its robustness against noise interference.
[0016] 2. The method provided in this application constructs a multi-granularity interpretation and analysis framework, based on the same bilinear parameter system, to simultaneously output local and global interpretations: at the local level, key node subsets and edge subsets are extracted through the interaction strength matrix to form a semantically readable interpretation subgraph; at the global level, the global importance distribution of feature dimensions is revealed through the third-order interaction matrix and eigenvalue decomposition. The two are logically consistent and mutually verifiable, forming a complete interpretation chain of "local structure - global features." This satisfies users' need to trace individual prediction results and supports macro-level control over the overall decision-making logic of the model, significantly improving the model's transparency and verifiability, and adapting to interpretation requirements in different scenarios.
[0017] 3. The method provided in this application designs a joint objective function that organically integrates the prediction loss (for downstream tasks such as node classification and link prediction) and the interpretation consistency loss (based on the representation similarity between the original graph and the interpretation subgraph) through weight coefficients. Combined with an end-to-end training approach, this achieves simultaneous optimization of the discriminative performance of the predicted representation and the consistency of the interpretation signal. This design avoids the objective conflict caused by traditional separate optimization, eliminates the need for additional training of the interpretation module in terms of training efficiency, and simultaneously ensures "high-precision prediction" and "high-fidelity interpretation" in terms of performance. It breaks through the performance trade-off bottleneck of existing methods and achieves a synergistic improvement in overall performance. Attached Figure Description
[0018] Figure 1 This is a flowchart of a bilinear graph neural network prediction and interpretation collaborative method provided in an embodiment of this application. Detailed Implementation
[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.
[0020] For example, refer to Figure 1 This application provides a bilinear graph neural network prediction and interpretation collaborative method, including: S1: Obtain the input graph to be processed, and extract the node feature matrix and adjacency matrix from the input graph. The node feature matrix is a matrix containing the feature information of each node in the input graph, and the adjacency matrix is a matrix representing the connection relationship between nodes in the input graph.
[0021] For example, let the input image be... , for The node feature matrix is given by N, where N is the number of nodes and F represents the node feature dimension.
[0022] S2: Input the preprocessed node feature matrix into the bilinear graph neural network, and generate prediction representation and interpretation signal simultaneously through the bilinear transformation layer; wherein the bilinear transformation layer is a linear transformation module used to simultaneously model the correlation between feature dimensions and the interaction pattern between nodes, the prediction representation is the node feature representation used for downstream task judgment, and the interpretation signal is information reflecting the correlation strength between feature dimensions and the interaction strength between nodes.
[0023] As one possible implementation, in step S2, a bilinear transformation is performed between the feature dimensions of the node feature matrix to encode the coupling relationship between each feature dimension, resulting in a first-stage hidden feature matrix; a bilinear transformation is performed between nodes on the first-stage hidden feature matrix to model the feature interaction pattern between nodes, resulting in a second-stage hidden feature matrix; and the predicted representation and interpretation signal is generated based on the second-stage hidden feature matrix.
[0024] For example, the specific process of the first-stage bilinear transformation of features includes: mapping each node feature vector in the node feature matrix to the hidden feature space, encoding the correlation between feature dimensions through the weight matrix of the bilinear transformation layer during the mapping process, and outputting the first-stage hidden feature matrix.
[0025] Specifically, for any node eigenvectors The mapping from the original feature space to the hidden feature space is achieved through a bilinear transformation layer: (1) in, express The hidden feature representation obtained after feature bilinear transformation The weight matrix is the weight matrix of the bilinear layer. This process not only completes feature mapping, but also encodes the coupling relationship between the feature dimensions at the parameter level.
[0026] For example, the specific process of the second-stage node bilinear transformation includes: transposing the first-stage hidden feature matrix, calculating the node pair interaction representation of the transpose matrix of the first-stage hidden feature matrix and the first-stage hidden feature matrix through the weight matrix of the bilinear transformation layer, and obtaining the second-stage hidden feature matrix by combining element-level operations.
[0027] Specifically, for the first-stage hidden feature matrix Transpose to obtain Then, the bilinear transformation layer is used to compute the interaction representation between node pairs: (2) in, , express The hidden node feature representation obtained after bilinear transformation of the nodes. This is the weight matrix of the bilinear layer. This represents element-wise multiplication. After completing the two-stage bilinear transformation, the node features not only contain higher-order correlations between dimensions but also incorporate information about interaction patterns between nodes.
[0028] For example, after obtaining the second-stage hidden feature matrix, the method further includes an adjacency matrix fusion step: based on the weight information of the node bilinear transformation layer, the node latent interaction information matrix B is obtained; the node latent interaction information matrix B is fused with the adjacency matrix A of the original input graph G to construct the weight matrix for hidden node interaction perception. The neighborhood information is aggregated based on the weight matrix, and the hidden feature matrix of the second stage is updated.
[0029] (3) After completing the two-stage bilinear transformation and adjacency matrix fusion steps, the bilinear graph neural network deeply fuses the structural information of hidden node interactions with the feature association information extracted by the bilinear transformation through a neighborhood information aggregation operation. The final output is a node representation that combines predictive and discriminative capabilities with interpretable readability. The specific form of this aggregation operation is as follows: (4) Here, H represents the final node representation matrix output by the aggregation operation. Each row vector in this matrix is the final feature representation of the corresponding node, which carries both prediction and discrimination information for downstream tasks (such as node classification and link prediction) and explanatory information that can be directly interpreted. This represents a non-linear activation function used to enhance the model's ability to fit complex feature relationships. Represents the weight matrix The corresponding degree matrix, It is a diagonal matrix, and the i-th element on the diagonal is equal to The sum of all elements in the i-th row is used to calculate the sum of all elements in the i-th row. Normalization is performed to avoid feature scale imbalance caused by differences in node degree during the aggregation process. Let represent the transpose of the hidden feature matrix in the second stage. This aggregation process utilizes the original graph structure while directly integrating the node association features extracted by the bilinear mechanism into the neighborhood information fusion, thereby outputting a node representation H that combines predictive and discriminative capabilities with interpretable readability.
[0030] S3: Based on the interpreted signal output by the bilinear transform layer, perform local interpretation analysis and global interpretation analysis respectively to obtain local interpretation results and global interpretation results.
[0031] As one possible implementation, in step S3, the specific process of local interpretation analysis includes: constructing a node pair interaction information matrix based on the node interaction strength information in the interpretation signal; performing logical operations on the node pair interaction information matrix and the adjacency matrix of the input graph to obtain the interpretation subgraph connection relationship matrix; and extracting node subsets and edge subsets from the input graph based on the interpretation subgraph connection relationship matrix to form the interpretation subgraph corresponding to the local interpretation result.
[0032] At the local interpretation level, the pairwise interaction information matrix of the nodes can be calculated using the parameters of the bilinear transformation layer to calculate the node interaction strength information that the different dimensions of the output vector depend on. Specifically, the j-th dimension output of the node bilinear layer can be expressed as the following equation (5): in, represent The j-th dimension feature, , This represents the interaction strength between different node pairs when the model outputs the j-th dimension feature. The obtained interaction strength matrix can then be used... and adjacency matrix Performing a logical AND operation yields the final interpretable subgraph connectivity matrix: (6) The resulting explanatory subgraph connection matrix Acting on the input graph This yields the final explanatory subgraph. ,in, They represent respectively by The induced subsets of nodes and edges.
[0033] As one possible implementation, in step S3, the specific process of global interpretation analysis includes: organizing the interaction strength information corresponding to all dimensions of the feature bilinear transformation output into a third-order interaction matrix. ; Combine the third-order interaction matrix with the given output direction vector Perform a dot product operation to obtain the global node feature interaction pattern matrix Q; perform eigenvalue decomposition on the global node feature interaction pattern matrix to obtain the eigenvalues with importance weights. ) eigenvectors This forms a global interpretation result.
[0034] (7) (8) Then the eigenvalues can be selected. Largest eigenvector This approach provides a direct, global perspective on the activation distribution of different feature dimensions within a node during the prediction process, revealing the global structural contribution pattern of node features to the overall prediction. Through this multi-granularity interpretation analysis method, the model can directly generate interpretations that encompass both global feature association patterns and local interpretations specific to individual samples during training and inference, thus achieving a comprehensive evaluation of the model's predictive behavior.
[0035] S4: Construct a joint objective function that includes prediction loss and interpretation consistency loss, train and optimize the bilinear graph neural network to obtain a trained bilinear graph neural network, so as to achieve synergy between prediction and interpretation.
[0036] Existing self-explanatory graph neural networks often set up separate optimization processes for prediction and interpretation during the training phase. This separate optimization approach makes it difficult to simultaneously guarantee the model's prediction accuracy and interpretation consistency, and may even lead to conflicting optimization objectives. For example, to improve prediction performance, the model may rely on high-dimensional feature association patterns that are difficult for humans to understand, while the interpretation module tends to select low-dimensional structural association patterns that are intuitively understandable to humans. The two lack inherent synergy at the feature space and parameter levels, thereby weakening the model's practicality and credibility in security-critical scenarios such as medical decision support, judicial intelligent assistance, and financial risk control.
[0037] The self-interpretive bilinear graph neural network provided in this application embodiment achieves deep coupling between the prediction process and the interpretation process at the feature representation and semantic mapping level, and introduces a "prediction-interpretation" collaborative optimization strategy, which simultaneously incorporates the optimization objectives related to prediction accuracy and the optimization objectives related to interpretation consistency into the joint objective function, achieving synergistic improvement of the two in the same training process, effectively avoiding the objective conflict and performance trade-off between prediction accuracy and interpretation consistency that are common in traditional separate optimization methods.
[0038] For example, the construction process of the joint objective function in step S4 includes: determining the prediction loss according to the downstream task type, which includes node classification and link prediction; calculating the interpretation consistency loss, which is determined based on the similarity between the output representation of the original input graph and the output representation of the interpretation subgraph; and constructing the joint objective function by balancing the prediction loss and the interpretation consistency loss through weight coefficients.
[0039] Specifically, suppose the objective function of the prediction task is: The objective function for interpreting consistency constraints is: Then the overall optimization objective can be defined as: (9) in, Weighting coefficients are used to balance prediction accuracy and interpretability consistency. In the prediction loss... In terms of specific tasks (such as node classification, link prediction, etc.), common indicators such as cross-entropy loss or mean squared error can be used.
[0040] As one possible implementation, the calculation process of the interpretation consistency loss includes: calculating the cosine similarity between the predicted representation corresponding to the original input graph and the output representation corresponding to the interpretation subgraph; constructing the interpretation consistency loss based on the cosine similarity to maximize the similarity between the two. For example, as shown in the following formula (10): in, and Let represent the representations output from the original input graph and the interpretation subgraph, respectively. Based on this similarity, the interpretation consistency loss can be expressed as: This loss aims to maximize the similarity between the original graph and the interpretation subgraph in the output representation, guiding the model to generate structural patterns that retain key information of the original prediction while being human-interpretable, thereby achieving deep synergy between the reasoning and interpretation processes.
[0041] In step S4, the collaborative optimization training adopts an end-to-end training method, which simultaneously optimizes the discriminative performance of the predicted representation and the consistency of the interpreted signal during the training process, so that the interpretation results are consistent with the decision basis of the bilinear graph neural network.
[0042] Step S4, the collaborative optimization training, adopts an end-to-end training approach. Its core logic is to integrate the prediction generation, interpretation signal extraction, and model parameter update of the bilinear graph neural network into a unified training process. This eliminates the need to separate the prediction and interpretation modules for independent training, ensuring the entire training process revolves around the core objective of "coordinated prediction and interpretation." During training, the model uses a joint objective function as its optimization guide, simultaneously promoting the coordinated improvement of the two core objectives: on the one hand, by optimizing the prediction loss, it enhances the discriminative performance of the prediction representation, ensuring accurate prediction results in downstream tasks such as node classification and link prediction; on the other hand, by optimizing the interpretation consistency loss, it ensures the consistency of the interpretation signal, that is, ensuring that the interpretation results generated based on the interpretation signal (including local interpretation subgraphs and global feature interaction patterns) match the prediction logic corresponding to the original input graph.
[0043] This synchronous optimization mode avoids the trade-off between prediction accuracy and interpretation quality in traditional separate training. It enables the model to simultaneously achieve high-precision prediction and high-fidelity interpretation during parameter iteration, ultimately ensuring that the interpretation results are completely consistent with the real decision-making basis of the bilinear graph neural network, significantly improving the model's credibility and practicality in safety-critical scenarios.
[0044] In the description of this specification, specific features, structures, materials, or characteristics may be combined in any suitable manner in one or more embodiments or examples.
[0045] The above are merely specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.
Claims
1. A bilinear graph neural network prediction and interpretation collaborative method, characterized in that, include: S1: Obtain the input graph to be processed, and extract the node feature matrix and adjacency matrix from the input graph. The node feature matrix is a matrix containing the feature information of each node in the input graph, and the adjacency matrix is a matrix representing the connection relationship between nodes in the input graph. S2: The preprocessed node feature matrix is input into a bilinear graph neural network, and a bilinear transformation layer is used to simultaneously generate predictive representations and explanatory signals. The bilinear transformation layer is a linear transformation module used to simultaneously model the correlation between feature dimensions and the interaction patterns between nodes. The predictive representation is the node feature representation used for downstream task judgment, and the explanatory signal is information reflecting the strength of the correlation between feature dimensions and the strength of the interaction between nodes. S3: Based on the explanatory signal output by the bilinear transformation layer, local explanatory analysis and global explanatory analysis are performed respectively to obtain local explanatory results and global explanatory results. S4: Construct a joint objective function that includes prediction loss and interpretation consistency loss, train and optimize the bilinear graph neural network to obtain a trained bilinear graph neural network, so as to achieve synergy between prediction and interpretation.
2. The bilinear graph neural network prediction and interpretation collaborative method according to claim 1, characterized in that, In step S2, a bilinear transformation is performed on the node feature matrix between feature dimensions to encode the coupling relationship between each feature dimension, thus obtaining the first-stage hidden feature matrix; A bilinear transformation is performed between nodes on the first-stage hidden feature matrix to model the feature interaction patterns between nodes, resulting in a second-stage hidden feature matrix. The predicted representation and interpretation signals are generated based on the second-stage hidden feature matrix.
3. The bilinear graph neural network prediction and interpretation collaborative method according to claim 2, characterized in that, The specific process of the first-stage bilinear transformation of features includes: mapping each node feature vector in the node feature matrix to the hidden feature space, encoding the correlation between feature dimensions through the weight matrix of the bilinear transformation layer during the mapping process, and outputting the first-stage hidden feature matrix.
4. The bilinear graph neural network prediction and interpretation collaborative method according to claim 2, characterized in that, The specific process of the second-stage node bilinear transformation includes: transposing the first-stage hidden feature matrix, calculating the node pair interaction representation of the transpose matrix of the first-stage hidden feature matrix and the first-stage hidden feature matrix through the weight matrix of the bilinear transformation layer, and obtaining the second-stage hidden feature matrix by combining element-level operations.
5. The bilinear graph neural network prediction and interpretation collaborative method according to claim 1, characterized in that, In step S3, the specific process of the local interpretation analysis includes: constructing a node pair interaction information matrix based on the node interaction strength information in the interpretation signal; performing logical operations on the node pair interaction information matrix and the adjacency matrix of the input graph to obtain the interpretation subgraph connection relationship matrix; and extracting node subsets and edge subsets from the input graph based on the interpretation subgraph connection relationship matrix to form the interpretation subgraph corresponding to the local interpretation result.
6. The bilinear graph neural network prediction and interpretation collaborative method according to claim 1, characterized in that, In step S3, the specific process of the global interpretation analysis includes: organizing the interaction intensity information corresponding to all dimensions of the feature bilinear transformation output into a third-order interaction matrix; performing a dot product operation between the third-order interaction matrix and the given output direction vector to obtain a global node feature interaction pattern matrix; performing feature decomposition on the global node feature interaction pattern matrix to obtain feature vectors with importance weights, thus forming a global interpretation result.
7. The bilinear graph neural network prediction and interpretation collaborative method according to claim 1, characterized in that, The process of constructing the joint objective function in step S4 includes: determining the prediction loss based on the downstream task type, wherein the downstream tasks include node classification and link prediction; calculating the interpretation consistency loss, wherein the interpretation consistency loss is determined based on the similarity between the output representation of the original input graph and the output representation of the interpretation subgraph; and constructing the joint objective function by balancing the prediction loss and the interpretation consistency loss through weight coefficients.
8. The bilinear graph neural network prediction and interpretation collaborative method according to claim 7, characterized in that, The calculation process of the interpretation consistency loss includes: calculating the cosine similarity between the predicted representation corresponding to the original input graph and the output representation corresponding to the interpretation subgraph, and constructing the interpretation consistency loss based on the cosine similarity to maximize the similarity between the two.
9. The bilinear graph neural network prediction and interpretation collaborative method according to claim 2, characterized in that, After obtaining the second-stage hidden feature matrix, the adjacency matrix fusion step is also included: based on the weight information of the node bilinear transformation layer, the node potential interaction information matrix is analyzed and obtained; the node potential interaction information matrix is fused with the adjacency matrix of the original input graph to construct the weight matrix for hidden node interaction perception. Based on the weight matrix, neighborhood information is aggregated, and the second-stage hidden feature matrix is updated.
10. The bilinear graph neural network prediction and interpretation collaborative method according to claim 1, characterized in that, The collaborative optimization training described in step S4 adopts an end-to-end training method, which simultaneously optimizes the discriminative performance of the predicted representation and the consistency of the interpreted signal during the training process, so that the interpretation results are consistent with the decision basis of the bilinear graph neural network.