Research map / Graph learning / GNN architectures
Spectral-Based Graph Neural Networks for Complementary Item 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 SComGNN, a spectral GNN with low-pass and mid-pass filters plus a two-stage attention mechanism to model relevance and dissimilarity in complementary item graphs, evaluated on four Amazon datasets.
Problem
Complementary item recommendation requires modeling items that are both relevant and dissimilar, a trade-off that prior methods oversimplify or ignore.
Approach
Spectral analysis shows complementary item graphs are dominated by low- and mid-frequency components, which are mapped to relevance and dissimilarity; low-pass and mid-pass GCN filters extract them and a two-stage (pairwise then self) attention adaptively balances them under a contrastive link-prediction loss.
Key results
- SComGNN achieves HR@10 of 0.7127 on Appliances, 0.8589 on Toys, 0.8565 on Grocery, and 0.8789 on Home, best on all datasets and metrics.
- Reported HR@10 gains over the best baseline are 7.8%, 14.3%, 13.3%, and 4.4% on the four datasets.
- Removing the low-frequency branch drops HR@10 to 0.1855 on Appliances, showing the low-frequency component is most important.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use SComGNN instead of the strongest baseline | complementary item recommendation · Appliances | HR@10: 7.8% | Helped |
| Use SComGNN instead of the strongest baseline | complementary item recommendation · Toys | HR@10: 14.3% | Helped |
| Use SComGNN instead of the strongest baseline | complementary item recommendation · Grocery | HR@10: 13.3% | Helped |
| Use SComGNN instead of the strongest baseline | complementary item recommendation · Home | HR@10: 4.4% | Helped |
| Add the low-frequency GCN filter and two-stage attention (ablation) | SComGNN · Appliances, Toys, Grocery, Home | NDCG: 0.2457 -> 0.4377; 0.2236 -> 0.5501; 0.2560 -> 0.5959; 0.2375 -> 0.6610 | Helped |
| Add the mid-frequency GCN filter and two-stage attention (ablation) | SComGNN · Appliances, Toys, Grocery, Home | NDCG: 0.4235 -> 0.4377; 0.4907 -> 0.5501; 0.5389 -> 0.5959; 0.6239 -> 0.6610 | Helped |
| Use two-stage attention instead of simple concatenation (ablation) | SComGNN · Appliances, Toys, Grocery, Home | NDCG: 0.4195 -> 0.4377; 0.5149 -> 0.5501; 0.5330 -> 0.5959; 0.6223 -> 0.6610 | Helped |
Surprising findings
- Removing the low-frequency filter causes a far larger drop than removing the mid-frequency filter, indicating relevance dominates.
- Without the two-stage attention, the model with both components can underperform the model using only the mid-frequency branch in some cases.
Limitations
- Only one-layer GCN filters work best; deeper networks degrade performance.
- Evaluation limited to Amazon also-bought link prediction on four datasets.
- Spectral analysis performed on a single small dataset for spectrum plots.
Experimental setup
Datasets: Amazon Appliances Amazon Grocery and Gourmet Food Amazon Toys and Games Amazon Home and Kitchen
Benchmarks: Amazon Appliances Amazon Toys Amazon Grocery Amazon Home
Metrics: HR@5 HR@10 NDCG
Baselines: GIN GraphSage Popularity DCF P-Companion ALCIR
Builds on: GCN P-Companion ALCIR DCF
Closest papers by meaning
- Semantic Enhanced Heterogeneous Hypergraph Network for Collaborative Filtering (AAAI 2025 · similarity 0.86)
- LS-TGNN: Long and Short-Term Temporal Graph Neural Network for Session-Based Recommendation (AAAI 2025 · similarity 0.86)
- ContextGNN: Beyond Two-Tower Recommendation Systems (ICLR 2025 · similarity 0.86)
- Beyond Graph Convolution: Multimodal Recommendation with Topology-aware MLPs (AAAI 2025 · similarity 0.86)
- Mind Individual Information! Principal Graph Learning for Multimedia Recommendation (AAAI 2025 · similarity 0.85)
- SHARP-Distill: A 68× Faster Recommender System with Hypergraph Neural Networks and Language Models (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.