Research map / Graph learning / Graph self-supervised learning
Beyond Flat Walks: Compositional Abstraction for Autoregressive Graph Generation
Kaiwen Bian, Andrew H Yang, Ali Parviz, Gal Mishne, Yusu Wang
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 a hierarchical, compositional abstraction for autoregressive graph generation that replaces flat node-edge walks, aimed at large structurally complex biological graphs.
Problem
Autoregressive graph generation linearizes graphs into flat walks over nodes and edges, which scales poorly to large, structurally complex biological graphs where higher-order organization is dense and important.
Approach
Introduces a hierarchical graph abstraction that composes higher-order structures so the autoregressive generation operates over abstracted components rather than a flat node-edge sequence.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace flat SENT walk with hierarchical HDT-HAC-R | SENT · Five molecular benchmarks | FCD: 51.6% to 64.6% lower FCD | Improved |
| replace flat SENT walk with hierarchical HDT-HAC-R | SENT · Five molecular benchmarks | MoleculePGD: 34.3% to 86.6% lower | Improved |
| replace SENT with HDT-HAC-R (within-community randomization) | SENT · Five molecular benchmarks | V.U.N.: 3.6% to 37.2% deficit | Hurt |
| use hierarchy-wide randomization (FR) instead of within-community (R) (ablation) | HDT-HAC · Five molecular benchmarks | V.U.N.: Macrocycle 0.537 -> 0.762 | Improved |
| replace SENT with HDT-HAC-R for zero-shot motif prompting | SENT · Motif-conditioned generation | motif discrepancy / V.U.N. | Hurt |
| replace SENT with HDT-HAC-R | SENT · Proteins (D&D) | V.U.N.: +62.8% | Improved |
| replace SENT with HDT-HAC-R | SENT · SSBM | V.U.N.: +59.3% | Improved |
| replace SENT with default HDT-HAC-R | SENT · SBM | V.U.N.: 0.669 -> 0.276 | Hurt |
| supply planted hierarchy instead of HAC at tau=10 (FR) (ablation) | HDT-HAC-FR · Connected RSBM (R=4) | V.U.N.: 0.083 -> 0.744 | Improved |
| increase HAC stopping threshold (ablation) | HDT-HAC · SBM | V.U.N. | Improved |
| use hierarchical serialization | SENT · Planar | graph MMD | Hurt |
| replace HDT with declarative HSENT or typed HDTC (ablation) | HDT-R · GuacaMol and MOSES | validity, V.U.N., MoleculePGD | Hurt |
Closest papers by meaning
- Flatten Graphs as Sequences: Transformers are Scalable Graph Generators (NeurIPS 2025 · similarity 0.87)
- SNACK: A Sequential Notation Framework for Probabilistic Graph Generation (NeurIPS 2026 · similarity 0.86)
- DMol: A Highly Efficient and Chemical Motif-Preserving Molecule Generation Platform (NeurIPS 2025 · similarity 0.85)
- Diffusion on Language Model Encodings for Protein Sequence Generation (ICML 2025 · similarity 0.85)
- TEMPO: Temporal Multi-scale Autoregressive Generation of Protein Conformational Ensembles (NeurIPS 2025 · similarity 0.85)
- DeFoG: Discrete Flow Matching for Graph Generation (ICML 2025 · 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.
