Research map / Video / Video-language understanding
Exploiting Fine-Grained Skip Behaviors for Micro-Video 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.
Refines micro-video skip signals into highly positive, less positive (delayed skip after 5s), and negative (quick skip) categories, and builds a dual-level positive graph plus hierarchical BPR and BCE losses on top of FRAME.
Problem
Conventional micro-video recommenders treat skip behavior as simply positive or negative, ignoring that delayed skips after the first few seconds carry a different, less negative signal than quick skips.
Approach
Split interactions into highly positive, less positive (skipped after 5 seconds), and negative (skipped within 5 seconds) signals; build two user-video graphs (highly and less positive) with dual-path GCN propagation, and train with a hierarchical BPR loss (highly vs less positive, highly vs negative) combined with BCE.
Key results
- On MVA, the proposed method achieves NDCG@5 of 0.812 vs 0.787 for BM3 and 0.784 for FRAME, with all eight metrics best and significant at p=0.01.
- On KuaiRand-Pure, the proposed method achieves Precision@3 of 0.279 vs 0.263 for FRAME and NDCG@3 of 0.591 vs 0.551 for FRAME.
- Ablation shows the dual-level graph outperforms both using all interactions as positive and using only highly positive interactions.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use fine-grained skip signals and dual-level positive graphs instead of conventional interaction modeling | Proposed micro-video recommender · MVA | Precision@3: 0.546 -> 0.573 | Helped |
| Use fine-grained skip signals and dual-level positive graphs instead of conventional interaction modeling | Proposed micro-video recommender · KuaiRand-Pure | Precision@3: 0.263 -> 0.279 | Helped |
| Use dual-level positives instead of treating all interactions as positive (ablation) | Proposed micro-video recommender · MVA | recommendation performance | Helped |
| Use dual-level positives instead of treating only highly positive interactions as positive (ablation) | Proposed micro-video recommender · MVA | recommendation performance | Helped |
| Use proposed BPR loss with quick-skip negatives instead of negative sampling for unseen interactions (ablation) | Proposed micro-video recommender · MVA | recommendation performance | Helped |
| Include less-positive interactions in dual BPR ranking (ablation) | Proposed micro-video recommender · MVA | recommendation performance | Helped |
Limitations
- Relies on a fixed 5-second threshold for quick vs delayed skip, which may not transfer across platforms.
- Compared against only three baselines.
- Evaluation limited to two datasets.
Experimental setup
Datasets: Micro-video-A (MVA) KuaiRand-Pure
Benchmarks: MVA KuaiRand-Pure
Metrics: Precision@k Recall@k MAP@k NDCG@k
Baselines: FRAME LightGT BM3
Builds on: FRAME BPR LightGCN
Closest papers by meaning
- When Top-ranked Recommendations Fail: Modeling Multi-Granular Negative Feedback for Explainable and Robust Video Recommendation (AAAI 2026 · similarity 0.86)
- CREAD: A Classification-Restoration Framework with Error Adaptive Discretization for Watch Time Prediction in Video Recommender Systems (AAAI 2024 · similarity 0.81)
- Infer Induced Sentiment of Comment Response to Video: A New Task, Dataset and Baseline (NeurIPS 2024 · similarity 0.81)
- Calibrating Video Watch-time Predictions with Credible Prototype Alignment (ICML 2025 · similarity 0.81)
- MRMLREC: A Two-Stage Approach for Addressing Data Sparsity in MOOC Video Recommendation (Student Abstract) (AAAI 2024 · similarity 0.81)
- TV-Rec: Time-Variant Convolutional Filter for Sequential Recommendation (NeurIPS 2025 · similarity 0.81)
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.