Research map / Graph learning / Knowledge and temporal graphs
Deep Temporal Graph Clustering
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes TGC, a batch-processing framework that adds node-level Student's t assignment and batch-level cosine reconstruction clustering losses on top of HTNE-style temporal loss to perform deep clustering on temporal graphs, with new arXivAI/arXivCS datasets.
Problem
Deep clustering on temporal graphs is largely unexplored because adjacency-matrix-based clustering modules cannot be applied to the interaction-sequence batch-processing pattern of temporal graph learning, and existing temporal methods focus on link prediction rather than node clustering.
Approach
Adapt deep clustering losses to batch training: node-level soft assignment with KL-based target distribution computed per batch, plus a batch-level pseudo-reconstruction that applies cosine-similarity constraints to positive, historical, and negative node pairs, added to the HTNE temporal loss.
Key results
- TGC achieves best ACC on all four datasets in the main table, e.g., 48.75 on DBLP vs 46.82 for TREND (second best), and 99.69 on School.
- On large-scale arXivCS, TGC reaches 39.95 ACC and 43.89 NMI, while static methods such as GAE and MVGRL run out of memory.
- On arXivCS, TGC uses 212.77 MB of GPU memory versus 6946.73 MB for SDCN, the only static method that did not run out of memory.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply TGC to temporal graph clustering | TGC · DBLP | ACC: 46.82 -> 48.75 | Helped |
| Apply TGC to temporal graph clustering | TGC · Brain | ACC: 47.46 -> 44.30 | Hurt |
| Apply TGC to temporal graph clustering | TGC · Patent | ACC: 46.64 -> 50.36 | Helped |
| Apply TGC to temporal graph clustering | TGC · School | ACC: 99.38 -> 99.69 | Helped |
| Apply TGC to temporal graph clustering | TGC · arXivAI | ACC: 65.66 -> 73.59 | Helped |
| Apply TGC to temporal graph clustering | TGC · arXivCS | ACC: 34.42 -> 39.95 | Helped |
| Add the node-level distribution module (ablation) | TGC · three ablation-study datasets | clustering performance | Helped |
| Add the batch-level reconstruction module (ablation) | TGC · three ablation-study datasets | clustering performance | Helped |
| Add both node-level distribution and batch-level reconstruction (ablation) | TGC · three ablation-study datasets | clustering performance | Helped |
Surprising findings
- Temporal graph learning baselines such as JODIE and TGN perform poorly on clustering (e.g., JODIE ACC 20.79 on DBLP), and the authors note that most temporal methods focus on link prediction rather than clustering.
- On School, AE and several static deep clustering methods score far below temporal methods because the adjacency matrix compresses many repeated interactions (N^2 < E).
Limitations
- Performance is not uniformly best; the authors state TGC may not be optimal on all datasets and that temporal methods lag deep static clustering methods in general.
- Few public temporal graph datasets with suitable clustering labels; some global information is lost without an adjacency matrix.
- Requires K (number of clusters) to be known, which the authors flag as a direction for future work.
Experimental setup
Datasets: DBLP Brain Patent School arXivAI arXivCS
Benchmarks: DBLP Brain Patent School arXivAI arXivCS
Metrics: ACC NMI ARI F1
Baselines: DeepWalk AE node2vec GAE MVGRL AGE DAEGC SDCN SDCNQ DFCN HTNE TGAT
Builds on: HTNE SDCN node2vec TGN TREND
Closest papers by meaning
- State Space Models on Temporal Graphs: A First-Principles Study (NeurIPS 2024 · similarity 0.89)
- Learning Persistent Community Structures in Dynamic Networks via Topological Data Analysis (AAAI 2024 · similarity 0.87)
- DECRL: A Deep Evolutionary Clustering Jointed Temporal Knowledge Graph Representation Learning Approach (NeurIPS 2024 · similarity 0.86)
- Using Time-Aware Graph Neural Networks to Predict Temporal Centralities in Dynamic Graphs (NeurIPS 2024 · similarity 0.85)
- TGB 2.0: A Benchmark for Learning on Temporal Knowledge Graphs and Heterogeneous Graphs (NeurIPS 2024 · similarity 0.85)
- Every Node Is Different: Dynamically Fusing Self-Supervised Tasks for Attributed Graph Clustering (AAAI 2024 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.