Research map / Architectures and efficiency / CNN design
TV-Rec: Time-Variant Convolutional Filter for 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 TV-Rec, a self-attention-free sequential recommender that replaces fixed convolutional kernels with time-variant (node-variant) graph filters derived from graph signal processing, and reports gains over SASRec, BSARec, AdaMCT and other baselines on six benchmarks.
Problem
Sequential recommendation models that use convolutional filters are typically fixed and lack the flexibility to capture position-specific, evolving user preferences, so hybrid models add self-attention at extra cost.
Approach
Reinterpret the user sequence as a (padded directed cyclic) graph and apply node-variant graph filters whose taps differ per position, parameterized as a coefficient matrix times an L2-normalized basis with an orthogonal regularizer; this replaces both fixed convolution kernels and self-attention and removes positional embeddings.
Key results
- TV-Rec outperforms the strongest baselines by an average of 7.49% across six benchmarks, e.g. 22.17% HR@10 gain on Foursquare and 19.13% HR@10 on LastFM.
- With maximum length 200 on ML-1M and Foursquare, TV-Rec achieves an average 4.74% improvement over top baselines.
- On XLong, TV-Rec beats LRURec on HR@5, HR@10 and all NDCG metrics but is slightly lower on HR@20 (-1.70%).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use time-variant filters instead of the strongest reported baseline | TV-Rec · Beauty, Sports, Yelp, LastFM, ML-1M, and Foursquare | NDCG@20: 0.0705 vs 0.0691; 0.0425 vs 0.0422; 0.0321 vs 0.0304; 0.0572 vs 0.0514; 0.1955 vs 0.1901; 0.0176 vs 0.0147 | Helped |
| use time-variant filters instead of the strongest baseline | TV-Rec · ML-1M (N=200) and Foursquare (N=200) | NDCG@20: 0.2157 vs 0.2079; 0.0158 vs 0.0142 | Helped |
| use TV-Rec rather than LRURec | TV-Rec · XLong | HR@20: 0.5874 vs 0.5774 | Hurt |
| use TV-Rec rather than LRURec | TV-Rec · XLong | NDCG@20: 0.3697 vs 0.4178 | Helped |
| add learnable positional embeddings (ablation) | TV-Rec · Beauty, Sports, Yelp, LastFM, ML-1M, and Foursquare | NDCG@20: 0.0705 -> 0.0702; 0.0425 -> 0.0396; 0.0321 -> 0.0320; 0.0572 -> 0.0496; 0.1955 -> 0.1936; 0.0176 -> 0.0160 | Mixed |
| replace the time-variant filter with a basic graph filter (ablation) | TV-Rec · Beauty, Sports, Yelp, LastFM, ML-1M, and Foursquare | NDCG@20: 0.0692; 0.0408; 0.0307; 0.0543; 0.1933; 0.0113 vs full TV-Rec 0.0705; 0.0425; 0.0321; 0.0572; 0.1955; 0.0176 | Helped |
| replace the learned basis matrix with an identity matrix (ablation) | TV-Rec · Beauty, Sports, Yelp, LastFM, ML-1M, and Foursquare | NDCG@20: 0.0698; 0.0410; 0.0317; 0.0539; 0.1930; 0.0141 vs full TV-Rec 0.0705; 0.0425; 0.0321; 0.0572; 0.1955; 0.0176 | Helped |
| remove basis matrix normalization (ablation) | TV-Rec · Beauty, Sports, Yelp, LastFM, and ML-1M | NDCG@20: 0.0689; 0.0412; 0.0264; 0.0418; 0.1912 vs full TV-Rec 0.0705; 0.0425; 0.0321; 0.0572; 0.1955 | Helped |
| vary the number of basis vectors (ablation) | TV-Rec · Beauty | NDCG@20 and HR@20 | Mixed |
| vary the number of basis vectors (ablation) | TV-Rec · LastFM | NDCG@20 and HR@20 | Mixed |
| increase dropout rate (ablation) | TV-Rec · Sports | recommendation performance | Helped |
| increase dropout rate (ablation) | TV-Rec · Foursquare | recommendation performance | Hurt |
Surprising findings
- Adding learnable positional embeddings gives inconsistent results across datasets rather than a benefit.
- On XLong, TV-Rec underperforms LRURec on HR@20 (0.5774 vs 0.5874).
Limitations
- Training cost per epoch is higher than some baselines (e.g., 13.20s vs 11.21s SASRec on Beauty), and the paper notes slightly higher inference time than FMLPRec.
- Hyperparameter search space for m, alpha and p is small and tuned per dataset.
Experimental setup
Datasets: Amazon Beauty Amazon Sports Yelp LastFM ML-1M Foursquare XLong
Benchmarks: Amazon Beauty Amazon Sports Yelp LastFM ML-1M Foursquare XLong
Metrics: HR@5 HR@10 HR@20 NDCG@5 NDCG@10 NDCG@20
Baselines: Caser GRU4Rec SASRec BERT4Rec NextItNet FMLPRec DuoRec LRURec AdaMCT BSARec SR-GNN GC-SAN
Builds on: node-variant graph filters FMLPRec BSARec AdaMCT
Closest papers by meaning
- An Attentive Inductive Bias for Sequential Recommendation beyond the Self-Attention (AAAI 2024 · similarity 0.89)
- Temporal Graph Contrastive Learning for Sequential Recommendation (AAAI 2024 · similarity 0.88)
- Lightweight Yet Fine-Grained: A Graph Capsule Convolutional Network with Subspace Alignment for Shared-Account Sequential Recommendation (AAAI 2025 · similarity 0.88)
- Iterative Sparse Attention for Long-sequence Recommendation (AAAI 2025 · similarity 0.87)
- Actions Speak Louder than Words: Trillion-Parameter Sequential Transducers for Generative Recommendations (ICML 2024 · similarity 0.86)
- Beyond Graph Convolution: Multimodal Recommendation with Topology-aware MLPs (AAAI 2025 · similarity 0.86)
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.