Iterative collaborative learning method for node characterization and topological structure

Abstract

The invention discloses an iterative collaborative learning method for node characterization and a topological structure, which comprises the following steps: acquiring initial graph data containing a node set and an initial edge set, and initializing node characterization; constructing a dynamic edge predictor, calculating the structure confidence degree between the node pairs based on the current node representation, generating a candidate edge weight matrix, and obtaining an optimized adjacent structure through a sparsification strategy; inputting the optimized adjacent structure into a graph neural network encoder, performing information propagation and aggregation in combination with the node feature matrix, and outputting enhanced node representation; the enhanced node representation is fed back to the dynamic edge predictor, and the edge weight and the node representation are iteratively updated until the convergence condition is met; and outputting a final optimized graph structure and node representation for downstream graph analysis. According to the method, through iterative collaborative optimization of the graph structure and the node representation, the graph topology can be dynamically corrected, the discrimination capability of the node representation is enhanced, the method is suitable for various complex scenes such as a noise graph and a sparse graph, and the performance of a downstream graph analysis task is improved.

Description

Technical Field

[0001] This invention relates to an iterative learning method, and more particularly to an iterative collaborative learning method for node representation and topology structure. Background Technology

[0002] Graph-structured data is widely found in fields such as social networks, knowledge graphs, and bioinformatics. Graph Neural Networks (GNNs) have become the mainstream method for processing graph data, learning node representations through neighborhood information aggregation mechanisms. However, traditional GNN methods typically assume that the graph structure is fixed and accurate, while real-world graphs often contain noise, missing edges, or redundant edges, leading to a decline in model performance.

[0003] Existing methods have attempted to modify graph structures through predefined rules or heuristics, but lack collaborative optimization with node representation learning. Other methods attempt to jointly learn structures and representations, but these often rely on static graph structures or fixed thresholds, making it difficult to adapt to dynamic graph evolution and prone to getting stuck in local optima or oversmoothing problems.

[0004] Therefore, there is an urgent need for a method that can dynamically and collaboratively optimize graph structure and node representation to improve the robustness and expressive power of graph representation learning. Summary of the Invention

[0005] Purpose of the invention: This invention proposes an iterative collaborative learning method for node representation and topology structure, which can dynamically and collaboratively optimize graph structure and node representation.

[0006] Technical solution: This invention includes the following steps:

[0007] S1: Obtain the initial graph data containing the set of nodes and the initial set of edges, and initialize the node representations;

[0008] S2: Construct a dynamic edge predictor, calculate the structural confidence between node pairs based on the current node representation, generate a candidate edge weight matrix, and obtain the optimized adjacency structure through the corresponding strategy;

[0009] S3: The optimized adjacency structure is input into the corresponding encoder, and information propagation and aggregation are performed in combination with the node feature matrix to output the enhanced node representation;

[0010] S4: Feed the enhanced node representation back to the dynamic edge predictor in S2 to update the edge weights and node representations until the conditions are met;

[0011] S5: Outputs the final optimized graph structure and node representations for downstream graph analysis tasks.

[0012] The initial map data ,in, It is a set of nodes. It is the initial set of edges. It is the node feature matrix.

[0013] The dynamic edge predictor adopts a dual-channel structure, including a third channel and a fourth channel.

[0014] The dynamic edge predictor outputs the third channel weight and the fourth channel weight, and fuses the third channel weight and the fourth channel weight through the first mechanism to output the edge weight.

[0015] The formula for calculating the weight of the third channel is as follows:

[0016]

[0017] in, For nodes The current representation, For learnable projection matrices, For element-wise multiplication, For learnable weight vectors, This is a bias term.

[0018] The formula for calculating the weight of the fourth channel is as follows:

[0019]

[0020] in, For nodes The set of first-order neighbors, For learnable matrices, These are learnable weights; This is a vector concatenation operation; This is a smoothing term.

[0021] The formula for calculating the edge weight is:

[0022]

[0023] in, The sigmoid function outputs the gate value. ; These are learnable parameters.

[0024] The calculation formula for the corresponding strategy is as follows:

[0025]

[0026] in For dynamic thresholds, The preset sparsity is set.

[0027] The conditions in S4 include: structural stability and characterization consistency.

[0028] An iterative collaborative learning system for node representation and topology includes:

[0029] Initial graph data acquisition module: Acquires initial graph data containing a set of nodes and an initial set of edges, and initializes node representations;

[0030] Dynamic edge predictor building module: Constructs a dynamic edge predictor, calculates the structural confidence between node pairs based on the current node representation, generates a candidate edge weight matrix, and obtains the optimized adjacency structure;

[0031] Enhanced node representation output module: The optimized adjacency structure is input into the encoder, and information propagation and aggregation are performed in combination with the node feature matrix to output the enhanced node representation;

[0032] Edge weight and node representation update module: Feeds the enhanced node representation back to the dynamic edge predictor to update the edge weight and node representation until the conditions are met;

[0033] Output module: Outputs the final optimized graph structure and node representation.

[0034] Beneficial effects: This invention can dynamically correct graph topology and enhance the discrimination ability of node representation by iteratively and collaboratively optimizing graph structure and node representation. It is applicable to various complex scenarios such as noisy graphs and sparse graphs, and improves the performance of downstream graph analysis tasks. Attached Figure Description

[0035] Figure 1 This is a flowchart of the present invention;

[0036] Figure 2 This is a network framework diagram of the dynamic edge predictor of the present invention. Detailed Implementation

[0037] The invention will now be further described with reference to the accompanying drawings.

[0038] Example 1

[0039] like Figure 1 As shown, the iterative collaborative learning method for node representation and topology structure in this embodiment includes the following steps:

[0040] S1: Obtain the initial graph data containing the set of nodes and the initial set of edges, and initialize the node representations.

[0041] Initial graph data ,in It is a set of nodes. It is the initial set of edges. It is the node feature matrix; initialize the node representation. The node feature matrix can be used directly. Or through a simple linear transformation We obtained, among which, This is the learnable parameter matrix.

[0042] S2: Construct a dynamic edge predictor, calculate the structural confidence between node pairs based on the current node representation, generate a candidate edge weight matrix, and obtain the optimized adjacency structure through the corresponding strategy.

[0043] The dynamic edge predictor employs a dual-channel structure: a third channel and a fourth channel. In this embodiment, the third channel is the semantic channel, and the fourth channel is the topological channel. The dynamic edge predictor outputs the semantic channel weights. and topology channel weights And by fusing semantic channel weights through the first mechanism and topology channel weights Output candidate edge weights The first mechanism employs a learnable gating mechanism.

[0044] Semantic Channel: Calculating Semantic Association Weights Based on Node Representation Similarity The calculation formula is:

[0045]

[0046] in, For nodes The current representation, To create a learnable projection matrix, an asymmetric design is used to break the symmetry of cosine similarity, adapting to directed graph scenarios. This is an element-wise product used to capture interactive features; For learnable weight vectors, This is a bias term.

[0047] Semantic channel node representation captures the semantic associations of nodes in the feature space, avoiding the limitations of relying solely on cosine similarity.

[0048] Topology Channel: Calculating Topology Rationality Weights Based on Structural Context Scoring The calculation formula is:

[0049]

[0050] in, For nodes The set of first-order neighbors, This is a learnable matrix used to extract the importance of neighboring nodes; These are learnable weights; This is a vector concatenation operation; To smooth out terms and prevent division by zero, the topological channel structure context score quantifies local structural consistency by weighting the representations of shared neighbors. This ensures that it incorporates both neighbor consensus terms and node-specific features, avoiding reliance solely on neighbor feature similarity.

[0051] Learnable gating mechanism: By fusing semantic and topological weights, the final edge weights are output. The calculation formula is:

[0052]

[0053] in, The sigmoid function outputs the gate value. ; These are learnable parameters. The learnable gating mechanism uses an adaptive fusion algorithm, where the gating value changes with the node pair. Furthermore, it introduces a non-linear tradeoff through the sigmoid function, which is superior to linear weighting, achieving a dynamic balance between the contributions of semantics and structure, thus avoiding the need for manually setting weights.

[0054] The corresponding strategy in this embodiment adopts a sparsity strategy, namely an adaptive threshold pruning algorithm. In each iteration, a threshold is set based on the dynamic quantile of the edge weight distribution, retaining only edges above the threshold to form a sparse and semantically enhanced graph structure. The adaptive threshold pruning algorithm is used to adjust the weights of candidate edges. Sparsification operations generate semantically enhanced graph structures after optimization;

[0055] The calculation formula for the adaptive threshold pruning algorithm is as follows:

[0056]

[0057] in The threshold is dynamic, and its value changes with iteration to avoid over-pruning caused by a fixed threshold. The preset sparsity is set to 0.9, which means retaining the first 10% of edges. Adaptive threshold pruning can avoid dense graphs and maintain a sparse and semantically strong topology.

[0058] S3: The optimized adjacency structure is input into the corresponding encoder, and information is propagated and aggregated in combination with the node feature matrix to output an enhanced node representation.

[0059] The encoder is a graph neural network encoder that uses a residual information aggregation mechanism. It introduces residual connections represented by nodes in the previous layer into each layer to prevent over-smoothing and preserve multi-level neighborhood information.

[0060] This embodiment uses a two-layer graph convolutional network with residual connections:

[0061]

[0062] in, for The normalized version, and This is the weight matrix. This represents the node representation from the previous iteration.

[0063] S4: Feed the enhanced node representations back to the dynamic edge predictor in S2, and iteratively update the edge weights and node representations until the convergence condition is met.

[0064] Convergence conditions include structural stability indices Consistency index with representation .

[0065] The structural stability index is the rate of change of the edges in the optimized adjacency structure over two consecutive iterations. Less than the preset edge change rate threshold , No. The iteration relative to the first Formula for calculating the rate of change of the edge in the next iteration for:

[0066]

[0067] in, For the first The optimized edge set obtained in the next iteration; This is the set of edges from the previous iteration; This is an indicator function that takes the value 1 when the condition is true and 0 otherwise.

[0068] The representation consistency index is the mean cosine similarity change of node representations. Less than the preset feature change threshold The node represents the mean change in cosine similarity. The calculation formula is:

[0069]

[0070] in, For nodes The representation vector; For nodes The representation vector from the previous iteration.

[0071] If and only if and When convergence is achieved, the iteration stops, where is... Preset edge change rate threshold, This is a preset threshold for feature changes.

[0072] S5: Outputs the final optimized graph structure and node representations for downstream graph analysis tasks.

[0073] Output final node representation and optimize graph structure It can be used for tasks such as node classification and link prediction.

[0074] Example 2

[0075] The iterative collaborative learning system for node representation and topology in this embodiment includes:

[0076] Initial graph data acquisition module: Acquires initial graph data containing a set of nodes and an initial set of edges, and initializes node representations;

[0077] Dynamic edge predictor building module: Constructs a dynamic edge predictor, calculates the structural confidence between node pairs based on the current node representation, generates a candidate edge weight matrix, and obtains an optimized adjacency structure through a sparsification strategy;

[0078] Enhanced node representation output module: The optimized adjacency structure input graph neural network encoder is combined with the node feature matrix to perform information propagation and aggregation, and outputs enhanced node representation;

[0079] Edge weight and node representation update module: Feeds the enhanced node representation back to the dynamic edge predictor and iteratively updates the edge weight and node representation until the convergence condition is met;

[0080] Output module: Outputs the final optimized graph structure and node representations for downstream graph analysis tasks.

Claims

1. An iterative collaborative learning method for node representation and topology structure, characterized in that, Includes the following steps: S1: Obtain initial graph data and initialize node representations; S2: Construct a dynamic edge predictor, calculate the structural confidence between node pairs based on the current node representation, generate a candidate edge weight matrix, and obtain the optimized adjacency structure through the corresponding strategy; S3: The optimized adjacency structure is input into the corresponding encoder, and information propagation and aggregation are performed in combination with the node feature matrix to output the enhanced node representation; S4: Feed the enhanced node representation back to the dynamic edge predictor in S2 to update the edge weights and node representations until the conditions are met; S5: Outputs the final optimized graph structure and node representations for downstream graph analysis tasks.

2. The iterative collaborative learning method for node representation and topology structure according to claim 1, characterized in that, The initial map data ,in, It is a set of nodes. It is the initial set of edges. It is the node feature matrix.

3. The iterative collaborative learning method for node representation and topology structure according to claim 1, characterized in that, The dynamic edge predictor adopts a dual-channel structure, including a third channel and a fourth channel.

4. The iterative collaborative learning method for node representation and topology structure according to claim 3, characterized in that, The dynamic edge predictor outputs the third channel weight and the fourth channel weight, and fuses the third channel weight and the fourth channel weight through the first mechanism to output the edge weight.

5. The iterative collaborative learning method for node representation and topology structure according to claim 4, characterized in that, The formula for calculating the weight of the third channel is as follows: in, For nodes The current representation, For learnable projection matrices, For element-wise multiplication, For learnable weight vectors, This is a bias term.

6. The iterative collaborative learning method for node representation and topology structure according to claim 4, characterized in that, The formula for calculating the weight of the fourth channel is as follows: in, For nodes The set of first-order neighbors, For learnable matrices, These are learnable weights; This is a vector concatenation operation; This is a smoothing term.

7. The iterative collaborative learning method for node representation and topology structure according to claim 4, characterized in that, The formula for calculating the edge weight is: in, The sigmoid function outputs the gate value. ; These are learnable parameters.

8. The iterative collaborative learning method for node representation and topology structure according to claim 1, characterized in that, The calculation formula for the corresponding strategy in S2 is as follows: in For dynamic thresholds, The preset sparsity is set.

9. The iterative collaborative learning method for node representation and topology structure according to claim 1, characterized in that, The conditions in S4 include: structural stability and characterization consistency.

10. A system for an iterative collaborative learning method of node representation and topology structure, characterized in that, The system is applicable to the iterative collaborative learning method for node representation and topology as described in any one of claims 1 to 9.

Images