Multi-view adaptive spectrum dynamic graph convolutional network method suitable for link prediction
By using a multi-view adaptive spectral dynamic graph convolutional network, the problem of insufficient identification of causal relationships and periodic structural transitions in dynamic graph link prediction is solved, and efficient prediction of future interaction relationships is achieved.
Patent Information
- Application Number
- CN202510872895.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-06-23
- Publication Date
- 2026-01-06
AI Technical Summary
Existing technologies have failed to effectively reveal the causal relationships between node interactions in dynamic graph link prediction, and have difficulty identifying frequent interaction patterns and periodic structural changes, resulting in limited predictive ability for future interaction relationships.
A multi-view adaptive spectral dynamic graph convolutional network is adopted. Through a time-interaction-aware graph embedding module, a dual-channel adaptive spectrum enhancement module, and a time-domain prediction module, node, edge, time, and interaction frequency information are encoded and transformed into the frequency domain for information enhancement. Finally, prediction is performed in the time domain.
It improves the accuracy of dynamic graph link prediction and the ability to capture nonlinear dynamic changes, thereby enhancing the ability to predict future interaction relationships.
Smart Images
Figure CN121283879A_ABST
Abstract
Description
Technical Field
[0001] The technical field involved in this invention is frequency coding and frequency domain methods in dynamic graph neural networks. Background Technology
[0002] Link prediction aims to predict future connections by analyzing the current network structure. With the continuous evolution of the internet, dynamic graph link prediction has shown broad application prospects in various fields such as social networks, transportation, and biological networks.
[0003] Dynamic graph neural networks (DTMs) provide a powerful tool for extracting features and learning representations from dynamically evolving graph data. Based on different temporal resolutions, dynamic graphs are generally classified into discrete-time dynamic graphs (DTDGs) and continuous-time dynamic graphs (CTDGs). DTDGs divide a dynamic graph into multiple static graphs, each reflecting the structural features of the graph at a specific point in time; however, this approach often ignores important dynamic information over time intervals. In contrast, CTDGs can more accurately characterize the dynamic processes of nodes and edges in a graph under high-frequency changes. However, despite the emergence of numerous CTDG-based methods in recent years, they still have many limitations.
[0004] The shortcoming of existing technologies lies in the fact that most current methods process interaction sequences using independent encoding, thus failing to reveal the potential causal relationships between these interactions. Some methods introduce a time encoder, attempting to capture the temporal dependencies of interactions through timestamp mapping, but struggle to identify hidden "transitions" within the time series. For example, in social networks, sustained interaction between user A and user B might stem from a specific topical event, while a sudden interaction between user A and user C could signify a shift in topic.
[0005] Furthermore, existing research has shown that frequent interaction patterns, periodic structural transitions, and other subtle dynamic changes between nodes are difficult to identify effectively in the time domain, but are easier to capture in the frequency domain. Therefore, this paper proposes to transform information from the time domain to the frequency domain to improve the effectiveness of information representation in link prediction. Summary of the Invention
[0006] This invention presents a multi-view adaptive spectral dynamic graph convolutional network suitable for link prediction. Existing research primarily focuses on graph analysis in the time domain, with limited attention paid to the structural evolution of dynamic graphs. Furthermore, existing frequency encoding methods for interaction sequences are based on neighbor co-occurrence mechanisms to model interactions between nodes. However, these methods rely heavily on the number of shared neighbors between node pairs in historical data, failing to reveal the complexity of node interaction patterns evolving over time, thus limiting their predictive power for future interactions. This invention overcomes the limitations of existing technologies, such as limited attention to structural evolution and reliance on shared neighbors for interaction frequency encoding.
[0007] To achieve the above objectives, a frequency-domain dynamic graph convolution method suitable for link prediction is adopted. First, this invention encodes multi-view information, including nodes, edges, time, and interaction frequencies, accurately depicting the dynamic interactions between nodes. Then, the encoded representation is mapped to the frequency domain, and the information expression is enhanced through a dual-channel transmission mechanism, effectively overcoming the limitations of single-channel methods in information transmission. Finally, the enhanced information representation is transformed to the time domain using an inverse Fourier transform, and the time-domain prediction module outputs the final prediction result.
[0008] A multi-view adaptive spectral dynamic graph convolutional network suitable for link prediction includes a time-interactive graph embedding module, a dual-channel adaptive spectrum enhancement module, and a temporal prediction module.
[0009] The time-interaction-aware graph embedding module aims to enhance the representation capability of dynamic graph data through multi-view feature extraction. It consists of node and edge feature encoding, Fourier time encoding, and matrix encoding based on neighbor interaction frequency.
[0010] Node and edge feature encoding is based on the historical interaction records of nodes. For each source node u and target node v, we select its K nearest direct neighbors to construct interaction sequences Tu and Tv. If a node has fewer than K neighbors, the input format is aligned using zero-padding.
[0011] We encode the features of nodes and edges based on the interaction sequences Tu and Tv. Specifically, for each neighbor of u and v and its corresponding edge, we generate node embeddings. and edge embedding Where d N With d E denoted as , where is the node and edge feature dimension, and K is the number of selected neighbors.
[0012] Fourier time encoding is used to model the temporal structure features in interaction sequences. This module introduces Fourier features to encode time differences. Let the interaction time of node i be t. i Then the time difference between any two nodes is Δt. ij =|ti -t j The time difference sequence is embedded into a high-dimensional periodic space via Fourier transform:
[0013]
[0014] in Embedding matrix for time features It is the Fourier frequency matrix d T For the time feature dimension, ΔT∈R K It is the time difference sequence with the current node.
[0015] This method effectively captures periodic patterns in node interactions by using sine and cosine mappings to represent multi-scale periodic changes, thereby enhancing the model's ability to model temporal behavior.
[0016] Matrix-based neighbor interaction frequency coding can simultaneously capture interaction frequency and temporal characteristics. Let the neighbor sets of source node u and target node v be:
[0017] Both sequences are arranged in chronological order. Taking node u as an example, if T... u ={b,a,v,b,c},T v ={g,a,u,f,a}, then T u The elements in T v The frequency of occurrence in the sequence is: 0, 2, 0, 0, 0. After considering the direct interaction between u and v, the final frequency sequence becomes {0, 2, 1, 0, 0}, reflecting the interaction strength from u to v.
[0018] In addition, the matrix-based neighbor interaction frequency coding also counts the interaction frequency between a node and its neighbors. For example, the frequency sequence of u {2, 1, 1, 2, 1} shows its local interaction activity, which helps to highlight key neighbors.
[0019] To improve efficiency, a neighbor interaction matrix M is constructed:
[0020]
[0021] Matrix operations can be used to efficiently solve for interaction frequencies. For example, the interaction strength of u→v is C. u→v =∑ i M ij v→uC u→v =∑ j M ji Furthermore, the interaction frequency feature vector of the nodes is obtained:
[0022]
[0023] By leveraging batch matrix operations, MIF is more computationally efficient than pointwise methods, fully utilizing the advantages of parallel computing and improving the scalability of large-scale dynamic graph analysis.
[0024] To unify feature dimensions and integrate multi-perspective information, we perform linear mapping on various feature types, transforming them into a unified embedding space:
[0025]
[0026] Where *∈{N, E, T, I} represents the node, edge, time, and interaction frequency features. For the mapping matrix, This is a bias term.
[0027] During fusion, we concatenate the mapped features and reduce their dimensionality, resulting in the final representation of the source node as follows:
[0028]
[0029] in For a dimension reduction matrix, As a bias term, the target node H v The calculation method is the same.
[0030] The dual-channel adaptive spectrum enhancement module aims to effectively capture nonlinear, fine-grained dynamic changes in the frequency domain. We transform historical interaction information into a frequency domain representation and obtain a complex spectrum representation through Fourier transform:
[0031]
[0032] Where j is the j-th node, N is the total number of nodes, K is the number of neighbors, and d f It is the frequency dimension.
[0033] To extract effective frequency domain features, we use a multilayer perceptron (MLP) to generate a mask coefficient matrix M. * The input is the concatenation result of the real and imaginary parts:
[0034]
[0035] Furthermore, multi-view spectral information is obtained by constructing two complementary channel masks:
[0036]
[0037] Where τ is a temperature parameter that controls the smoothness of the mask. The enhanced signal is obtained by calculating the Hadamard product of the obtained mask with the real and imaginary parts of the signal.
[0038]
[0039] Then, learnable parameters α and β are introduced for adaptive weighted fusion:
[0040]
[0041] Among them W * ∈C K×d It is a learnable frequency domain mapping tensor.
[0042] Finally, through inverse Fourier transform Φ * By restoring the spectral representation to the time domain, the enhanced dynamic graph feature representation X is obtained. * :
[0043]
[0044] The temporal prediction module introduces an MLP-Mixer structure to capture nonlinear features, enhancing the temporal features at each time step. Perform Token-Mixing and Channel-Mixing operations separately:
[0045]
[0046] Where σ is the nonlinear activation function and LN is the layer normalization. (* represents token or channel) are trainable parameters.
[0047] The time-aware representations of nodes u and v are obtained through weighted convergence:
[0048]
[0049] Finally, the node representations are fed into a two-layer MLP, and after ReLU and Softmax activation, the predicted probabilities are output:
[0050]
[0051] The training objective is to minimize the binary cross-entropy loss:
[0052]
[0053] This loss function exhibits good numerical stability and convergence properties, effectively measuring the difference between the model's predicted probabilities and the true labels, thus guiding the model's continuous optimization during training. In dynamic graph link prediction tasks, the ratio of positive to negative samples is often severely imbalanced. This loss function maintains stable optimization performance even in such scenarios, helping to improve the model's ability to discriminate sparse connections and its overall generalization performance. Attached Figure Description
[0054] Figure 1This is the training graph for a multi-view adaptive spectral dynamic graph convolutional network.
[0055] Figure 2 This describes the calculation process for matrix-based neighbor interaction frequency encoding.
[0056] Figure 3 This is a flowchart of the dual-channel adaptive spectrum enhancement module.
[0057] Figure 4 This is a flowchart of the time-domain prediction module. Detailed Implementation
[0058] To more clearly illustrate the technical solutions in the implementation of the present invention, the present invention will be further described and explained below with reference to the accompanying drawings and embodiments, but the content of the present invention is not limited to the scope described.
[0059] Example: Figure 1 As shown, a multi-view adaptive spectral dynamic graph convolutional network suitable for link prediction includes a time-interactive graph embedding module, a dual-channel adaptive spectrum enhancement module, and a temporal prediction module.
[0060] Step 1: Construct a time-interaction-aware graph embedding module
[0061] This module enhances dynamic graph modeling capabilities through multi-view information, specifically comprising the following three sub-processes:
[0062] 1. Node and edge feature encoding:
[0063] Assuming the source node is u and the target node is v, select the K nearest direct neighbors based on their historical interaction records to construct interaction sequences Tu and Tv. If the number of neighbors is less than K, zero-padding is used to align the sequence length. Based on the interaction sequences, the following steps are taken:
[0064] Node embedding:
[0065]
[0066] Edge embedding:
[0067]
[0068] 2. Fourier time coding:
[0069] For each pair of interacting nodes, based on their interaction time difference:
[0070] Δt ij =|t i -t j |
[0071] Embed it into a periodic space using Fourier transform:
[0072]
[0073] 3. Matrix coding based on neighbor interaction frequency (e.g.) Figure 2 As shown):
[0074] Construct a neighbor interaction matrix M.
[0075]
[0076] Constructing frequency eigenvectors:
[0077]
[0078] Finally, the above multi-perspective information is unified into a single feature dimension and then fused.
[0079] Step 2: Construct a dual-channel adaptive spectrum enhancement module (e.g., Figure 3 (As shown)
[0080] By applying a Fourier transform to the node history embedding representation, we obtain the complex spectrum:
[0081]
[0082] The real and imaginary parts are concatenated and then input into the MLP to generate a mask.
[0083] M * =MLP(Concat(f real ,f imag ))
[0084] Generate complementary spectral channels based on masks:
[0085]
[0086] Spectrum Enhancement:
[0087]
[0088] Adaptive fusion:
[0089]
[0090] Finally, the inverse Fourier transform is used to restore the time-domain features:
[0091]
[0092] Step 3: Construct the time-domain prediction module (e.g.) Figure 4 (As shown)
[0093] A time-domain modeling approach is employed using an MLP-Mixer architecture, where a two-stage mixing process is performed for the representation at each time step.
[0094] Token-Mixing:
[0095]
[0096] Channel-Mixing:
[0097]
[0098] The final representation of each node is weighted and aggregated:
[0099]
[0100] Step 4, link prediction and training
[0101] The representations of the source node and the target node are concatenated and then input into a two-layer MLP:
[0102]
[0103] The training process uses the binary cross-entropy loss function:
[0104]
[0105] The above description is merely one specific embodiment of the present invention, used to illustrate the technical solution of the present invention, and is not intended to limit the scope of protection of the present invention. In practical applications, the network structure and input method can be flexibly adjusted according to actual needs, and still fall within the scope of protection of the present invention.
Claims
1.A multi-view adaptive spectral dynamic graph convolutional network method suitable for link prediction, characterized in that The method comprises a matrix coding based on neighbor interaction frequency and a dual-channel adaptive spectrum enhancement module. The matrix coding based on neighbor interaction frequency is used for constructing a neighbor set based on historical interaction data of nodes and generating an interaction frequency feature representation between nodes, and specifically comprises the following steps: a) For any to-be-predicted node pair (u, v), the K latest interaction neighbors of the node u and the node v in time are extracted respectively to form time-ordered neighbor sequences Tu and Tv; b) constructing an interaction frequency matrix M, the elements M ij satisfy the following conditions: c) based on the interaction frequency matrix M, generating an interaction frequency feature vector H for node u and node v respectively by matrix operation u = [C u→u , C u→v ] and H v = [C v→u , C v→v ]; C v→u =∑ j M ji C u→v =∑ i M ij The dual-channel adaptive spectrum enhancement module enhances information expression through a dual-channel transmission mechanism, effectively overcoming the limitations of single-channel methods in information transmission, and specifically comprises the following steps: a) Fourier transforming the historical interaction sequence to obtain a complex spectral representation F comprising a real part and an imaginary part * ; b) generating a frequency mask M from the spectral representation F * inputting the multi-layer perception model, generating a frequency mask M * and further generating a causal mask and a pseudo correlation mask according to a set temperature parameter τ; c) Based on a causal mask and a pseudo-correlation mask, the spectrum representation is weighted to obtain a causal spectrum and a pseudo-correlation spectrum, respectively, And through the learnable parameters α * and β * Fusion, get the frequency domain features d) on the frequency domain features performing an inverse Fourier transform to obtain the enhanced time domain feature representation X * .