Research map / Learning theory and algorithms / Combinatorial and search optimization
Optimizing Recall in Deep Graph Hashing Framework for Item Retrieval (Student Abstract)
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 SRLoss, a smooth differentiable approximation of the Recall@N objective for hashing-based recommendation on LightGCN, with an approximation-adjustable gradient estimator based on sign-swish.
Problem
Hashing-based recommendation trains with BPR while evaluation uses Recall, and the discrete hash constraints make directly optimizing recall NP-hard; straight-through estimation gives coarse gradients.
Approach
Replace the non-differentiable indicator functions in Recall@N with smooth approximations (sigmoid, exponential, hinge, softplus) and estimate hash-function gradients with a sign-swish function whose temperature beta controls approximation accuracy.
Key results
- On Gowalla, the proposed method reaches Recall@100 of 0.31396 versus 0.26590 for BGCH (the strongest baseline).
- On Yelp2018, the proposed method reaches Recall@100 of 0.16348 versus 0.13450 for BGCH.
- Hash code length is 64 bits; data split 8:1:1 train/valid/test.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use the proposed SRLoss and approximation-adjustable gradient estimator instead of the strongest baseline, BGCH | Deep graph hashing recommendation · Gowalla | Recall@50: 0.19160 -> 0.23082 | Helped |
| Use the proposed SRLoss and approximation-adjustable gradient estimator instead of the strongest baseline, BGCH | Deep graph hashing recommendation · Gowalla | Recall@100: 0.26590 -> 0.31396 | Helped |
| Use the proposed SRLoss and approximation-adjustable gradient estimator instead of the strongest baseline, BGCH | Deep graph hashing recommendation · Gowalla | NDCG@50: 0.12740 -> 0.15109 | Helped |
| Use the proposed SRLoss and approximation-adjustable gradient estimator instead of the strongest baseline, BGCH | Deep graph hashing recommendation · Gowalla | NDCG@100: 0.14840 -> 0.17424 | Helped |
| Use the proposed SRLoss and approximation-adjustable gradient estimator instead of the strongest baseline, BGCH | Deep graph hashing recommendation · Yelp2018 | Recall@50: 0.08350 -> 0.10140 | Helped |
| Use the proposed SRLoss and approximation-adjustable gradient estimator instead of the strongest baseline, BGCH | Deep graph hashing recommendation · Yelp2018 | Recall@100: 0.13450 -> 0.16348 | Helped |
| Use the proposed SRLoss and approximation-adjustable gradient estimator instead of the strongest baseline, BGCH | Deep graph hashing recommendation · Yelp2018 | NDCG@50: 0.05000 -> 0.06016 | Helped |
| Use the proposed SRLoss and approximation-adjustable gradient estimator instead of the strongest baseline, BGCH | Deep graph hashing recommendation · Yelp2018 | NDCG@100: 0.06700 -> 0.08035 | Helped |
Limitations
- Evaluated only on two datasets with a short student-abstract length.
- Limited baselines; excludes DCD-based hashing recommendation baselines.
- No ablation isolating the smooth recall loss from the gradient estimator.
Experimental setup
Datasets: Gowalla Yelp2018
Benchmarks: Gowalla Yelp2018
Metrics: Recall@50 Recall@100 NDCG@50 NDCG@100
Baselines: BGCH HashGNN HashRec
Builds on: LightGCN Bayesian Personalized Ranking straight-through estimation sign-swish
Closest papers by meaning
- Stochastically Robust Personalized Ranking for LSH Recommendation Retrieval (AAAI 2020 · similarity 0.87)
- Discrete Personalized Ranking for Fast Collaborative Filtering from Implicit Feedback (AAAI 2017 · similarity 0.86)
- Multi-Feature Discrete Collaborative Filtering for Fast Cold-Start Recommendation (AAAI 2020 · similarity 0.85)
- Hashing as Tie-Aware Learning to Rank (CVPR 2018 · similarity 0.85)
- Deep Graph Online Hashing for Multi-Label Image Retrieval (AAAI 2025 · similarity 0.83)
- Ada-Retrieval: An Adaptive Multi-Round Retrieval Paradigm for Sequential Recommendations (AAAI 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.