TFWaveFormer: Temporal-Frequency Collaborative Multi-level Wavelet Transformer for Dynamic Link Prediction

Hantong Feng, Yonggang Wu, Duxin Chen, Wenwu Yu

TL;DR

TFWaveFormer is a novel Transformer architecture that masters dynamic link prediction by integrating learnable, multi-resolution wavelet decomposition directly into the model. By synergizing local temporal details with global frequency patterns, it achieves state-of-the-art performance across ten diverse benchmarks while maintaining high computational efficiency.

What Problem Does This Paper Address?

Predicting future connections in a dynamic graph—like a researcher joining a new collaboration or a user editing a specific Wikipedia page—is notoriously difficult. Real-world networks don't just evolve linearly; they exhibit "multi-scale" temporal dynamics. This includes high-frequency periodic fluctuations (e.g., daily social media bursts), long-range dependencies (e.g., career-long research trends), and abrupt topological shifts caused by sudden events.

Existing methods usually fall into one of three traps:

  1. Recurrent models (RNNs) struggle with long-range dependencies due to vanishing gradients.
  2. Standard Temporal Transformers (like TGAT) are great at capturing instantaneous interactions but often fail to distinguish between different frequency patterns in non-stationary data.
  3. Pure Frequency-domain methods capture global trends but miss the localized "micro-details" essential for precise prediction.

The authors of TFWaveFormer aim to bridge this gap by creating a model that "sees" the network in both the time and frequency domains simultaneously.

Methods

The TFWaveFormer framework consists of three sophisticated stages designed to handle the "heavy lifting" of temporal analysis:

  • Feature Extraction & Alignment: The model first extracts multi-modal features, including edge semantics, node attributes, and a "Node Interaction Frequency" (NIF) encoder to capture topological relationships.
  • Learnable Multi-Level Wavelet Module: This is the core innovation. Instead of using traditional, fixed mathematical wavelets (like Haar or Daubechies), the authors use parallel, multi-scale convolutional kernels. This allows the model to learn the optimal wavelet-like filters directly from the data. It decomposes the signal into multiple scales—capturing everything from fine-grained short-term bursts to coarse-grained long-term trends—without the computational overhead of iterative transforms.
  • Temporal-Frequency Hybrid Transformer: The model uses a gating mechanism to reconstruct the most important features across different scales. These "frequency-aware" features are then fused with standard temporal features using a multi-head self-attention mechanism. This ensures the final node embeddings reflect both the immediate context and the broader evolutionary rhythm of the network.

Datasets

The researchers evaluated TFWaveFormer on ten real-world datasets spanning various domains and scales:

  • Social/Collaboration: Wikipedia, Reddit.
  • Interaction-based: MOOC (online learning), LastFM (music listening).
  • Communication: Enron (corporate email), UCI (student messaging).
  • Proximity/Mobility: Social Evolution, Contact (physical proximity), Flights (aviation network).
  • Economics: UN Trade (international commodity trade).

These datasets range from small-scale (Enron, 184 nodes) to large-scale (Flights, 13,000+ nodes), providing a robust test of the model's generalizability.

Key Insights and Ideas

  • Learnable Wavelets are Superior: By replacing fixed-basis transforms with learnable convolutions, TFWaveFormer can adapt its "vision" to the specific temporal signature of a dataset.
  • The "Dual-Domain" Advantage: The paper demonstrates that time-domain features (microscopic evolution) and frequency-domain features (macroscopic trends) are complementary. Removing either component leads to a significant drop in performance, especially on sparse datasets.
  • Scale Sensitivity: The optimal number of wavelet scales ($m$) varies by dataset. Dense social networks with complex, bursty activity benefit from more scales (e.g., $m=5$), while simpler communication networks reach peak performance with fewer scales ($m=3$).
  • Efficiency Matters: Unlike some recent high-accuracy models that are computationally expensive, TFWaveFormer sits on the "Pareto front," offering the best accuracy-to-training-time ratio among its peers.

Key Results

TFWaveFormer established new state-of-the-art (SOTA) results across nearly all tested metrics:

  • Transductive Setting: It achieved 99.33% AP on Wikipedia and 99.32% AP on Reddit, outperforming previous leaders like DyGMamba and DyGFormer.
  • Inductive Setting: On unseen nodes, it showed remarkable generalization, notably gaining 1.17% AP on the LastFM dataset over the second-best method.
  • Consistency: Across ten datasets, it maintained an average rank of 1.20 to 1.70, indicating it is more stable and reliable than specialized models that only perform well on specific types of graphs.

Limitations and Future Work

While TFWaveFormer is highly effective, the authors noted a few areas for growth:

  • Scale Exploration: Due to time constraints, the authors only tested up to five wavelet scales ($m=5$). They suggest that even more complex, larger-scale datasets might benefit from higher values of $m$.
  • Adaptive Configuration: Currently, the number of scales is a hyperparameter. Future work could involve making the model automatically determine the optimal number of scales based on the input graph's structural and temporal characteristics.
  • Extreme Sparsity: While it performs well on sparse graphs like UN Trade, the performance gap between it and other models narrows in these scenarios, suggesting room for further optimization in extremely low-data environments.

Summary updated

Abstract

Dynamic link prediction plays a crucial role in diverse applications including social network analysis, communication forecasting, and financial modeling. While recent Transformer-based approaches have demonstrated promising results in temporal graph learning, their performance remains limited when capturing complex multi-scale temporal dynamics. In this paper, we propose TFWaveFormer, a novel Transformer architecture that integrates temporal-frequency analysis with multi-resolution wavelet decomposition to enhance dynamic link prediction. Our framework comprises three key components: (i) a temporal-frequency coordination mechanism that jointly models temporal and spectral representations, (ii) a learnable multi-resolution wavelet decomposition module that adaptively extracts multi-scale temporal patterns through parallel convolutions, replacing traditional iterative wavelet transforms, and (iii) a hybrid Transformer module that effectively fuses local wavelet features with global temporal dependencies. Extensive experiments on benchmark datasets demonstrate that TFWaveFormer achieves state-of-the-art performance, outperforming existing Transformer-based and hybrid models by significant margins across multiple metrics. The superior performance of TFWaveFormer validates the effectiveness of combining temporal-frequency analysis with wavelet decomposition in capturing complex temporal dynamics for dynamic link prediction tasks.
BibTeX
@inproceedings{feng2026tfwaveformer,
  author = {Hantong Feng and Yonggang Wu and Duxin Chen and Wenwu Yu},
  title = {{TFWaveFormer: Temporal-Frequency Collaborative Multi-level Wavelet Transformer for Dynamic Link Prediction}},
  booktitle = {The Web Conference},
  year = {2026},
  doi = {10.1145/3774904.3792422},
}

Loading related content...

Figure