Research map / Graph learning / Knowledge and temporal graphs
Lightweight Yet Fine-Grained: A Graph Capsule Convolutional Network with Subspace Alignment for Shared-Account Sequential Recommendation
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 LightGC$^2$N, a lightweight graph capsule convolutional network with subspace alignment that captures fine-grained user-interaction ownership in shared-account sequential recommendation at linear complexity.
Problem
Shared-account sequential recommendation must distinguish interactions from multiple latent users in hybrid account sequences, while existing graph/RNN SSR methods are coarse-grained and have quadratic computational complexity unsuitable for resource-constrained devices.
Approach
Builds capsule graphs where account-level latent users are capsules that attentively propagate messages to interaction items (fine-grained ownership matching), merges users via account-level dynamic routing, and refines sequence embeddings with low-rank subspace affinities aligned to the originals through an InfoNCE contrastive loss.
Key results
- LightGC$^2$N reaches Recall@5 of 61.35% on HV-E versus 54.11% for the next best method TiDA-GCN.
- On HA-M, LightGC$^2$N reports Recall@5 of 46.02% and MRR@5 of 41.30%, versus 23.55% and 20.91% for TiDA-GCN.
- Ablation removing subspace alignment (w/o S) drops HV-E Recall@5 from 61.35% to 57.96%; removing all components (w/o A) drops it to 22.90%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use LightGC²N instead of TiDA-GCN | Shared-account sequential recommendation · HV-E | Recall@5: 54.11 -> 61.35 | Helped |
| Use LightGC²N instead of TiDA-GCN | Shared-account sequential recommendation · HV-V | Recall@5: 76.37 -> 79.29 | Helped |
| Use LightGC²N instead of TiDA-GCN | Shared-account sequential recommendation · HA-M | Recall@5: 23.55 -> 46.02 | Helped |
| Use LightGC²N instead of TiDA-GCN | Shared-account sequential recommendation · HA-B | Recall@5: 24.69 -> 47.36 | Helped |
| Use linear attention instead of Conv1D for item capsule projection (ablation) | LightGC²N · HV-E | Recall@5: 58.33 -> 61.35 | Helped |
| Add account-level dynamic routing (ablation) | LightGC²N · HV-E | Recall@5: 59.86 -> 61.35 | Helped |
| Use GC²N instead of traditional graph convolution (ablation) | LightGC²N · HV-E | Recall@5: 53.46 -> 61.35 | Helped |
| Add contrastive learning to subspace alignment (ablation) | LightGC²N · HV-E | Recall@5: 59.96 -> 61.35 | Helped |
| Add subspace alignment (ablation) | LightGC²N · HV-E | Recall@5: 57.96 -> 61.35 | Helped |
| Add LightGC²N components rather than remove all components (ablation) | LightGC²N · HV-E | Recall@5: 22.90 -> 61.35 | Helped |
Surprising findings
- Removing the contrastive learning drops HV-E Recall@5 only slightly, and the w/o A variant (all components removed) falls far below all baselines (22.90% Recall@5 on HV-E).
Limitations
- Assumes a fixed number of latent users (alpha) per shared account, which must be tuned per dataset.
- Baselines are reported under cross-domain settings; results depend on hyper-parameter tuning of baselines on each dataset.
- Hyper-parameters such as beta and gamma are sensitive, with best performance only at certain values (0.8 or 0.9).
Experimental setup
Datasets: HV-E HV-V HA-M HA-B
Benchmarks: HV-E HV-V HA-M HA-B
Metrics: Recall@5 Recall@20 MRR@5 MRR@20
Baselines: NCF LightGCN HRNN NAIS TGSRec π-Net PSJNet DA-GCN TiDA-GCN
Builds on: LightGCN DA-GCN TiDA-GCN dynamic routing InfoNCE
Closest papers by meaning
- TV-Rec: Time-Variant Convolutional Filter for Sequential Recommendation (NeurIPS 2025 · similarity 0.88)
- Sparse Enhanced Network: An Adversarial Generation Method for Robust Augmentation in Sequential Recommendation (AAAI 2024 · similarity 0.85)
- ContextGNN: Beyond Two-Tower Recommendation Systems (ICLR 2025 · similarity 0.85)
- Temporal Graph Contrastive Learning for Sequential Recommendation (AAAI 2024 · similarity 0.84)
- Beyond Graph Convolution: Multimodal Recommendation with Topology-aware MLPs (AAAI 2025 · similarity 0.84)
- Actions Speak Louder than Words: Trillion-Parameter Sequential Transducers for Generative Recommendations (ICML 2024 · similarity 0.83)
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.