Research map / Label-efficient and robust learning / Recognition and long-tail
Knowledge-Aware Explainable Reciprocal 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 KAERR, a knowledge-graph metapath model for reciprocal recommendation that encodes metapaths from each party's perspective with a BiLSTM, fuses them by attention, and trains with a bilateral quadruple loss.
Problem
Reciprocal recommendation (e.g., job matching, online dating) suffers from sparse bilateral interaction data, and existing content-based methods model side information from a unified view, failing to capture each party's distinct intentions and giving limited interpretability.
Approach
Extract knowledge-graph metapaths between the two parties, encode each metapath independently from candidate and job perspectives using a BiLSTM whose two directions start at each side, fuse metapaths with sigmoid attention whose weights serve as explanations, and optimize a bilateral quadruple loss that also uses unilateral rejections.
Key results
- On Zhaopin, KAERR reaches Recall@5 0.9477 and NDCG@5 0.8275, versus 0.7777 and 0.6144 for the best baseline DPGNN.
- On UEM, KAERR reaches Recall@5 0.9146 and NDCG@5 0.8014, versus 0.8259 and 0.6532 for DPGNN (Recall@5 is lower than DPGNN on UEM? No: KAERR 0.9146 exceeds DPGNN 0.8259).
- Ablations show removing the dual-perspective encoder, attentive fusion, or bilateral quadruple loss each lowers Recall@5 on both datasets.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use KAERR instead of DPGNN | reciprocal recommendation · Zhaopin, candidates | R@5: 0.7777 -> 0.9477 | Helped |
| Use KAERR instead of DPGNN | reciprocal recommendation · Zhaopin, jobs | R@5: 0.7460 -> 0.9499 | Helped |
| Use KAERR instead of DPGNN | reciprocal recommendation · UEM, candidates/researchers | R@5: 0.8259 -> 0.9146 | Helped |
| Use KAERR instead of DPGNN | reciprocal recommendation · UEM, jobs/demands | R@5: 0.8064 -> 0.9202 | Helped |
| Add dual-perspective metapath encoding instead of shared LSTM encoding (ablation) | KAERR · Zhaopin, candidates | R@5: 0.8936 -> 0.9477 | Helped |
| Add dual-perspective metapath encoding instead of shared LSTM encoding (ablation) | KAERR · UEM, candidates/researchers | R@5: 0.8653 -> 0.9146 | Helped |
| Add attentive metapath fusion instead of mean pooling (ablation) | KAERR · Zhaopin, candidates | R@5: 0.9105 -> 0.9477 | Helped |
| Add attentive metapath fusion instead of mean pooling (ablation) | KAERR · UEM, candidates/researchers | R@5: 0.8702 -> 0.9146 | Helped |
| Use bilateral quadruple loss instead of BPR loss (ablation) | KAERR · Zhaopin, candidates | R@5: 0.9079 -> 0.9477 | Helped |
| Use bilateral quadruple loss instead of BPR loss (ablation) | KAERR · UEM, candidates/researchers | R@5: 0.8669 -> 0.9146 | Helped |
Surprising findings
- The content-based PJFNN outperforms several collaborative-filtering baselines by a large margin on Zhaopin.
Limitations
- Evaluated on only two datasets, one of which is private.
- Metapath patterns are predefined and must be specified manually.
Experimental setup
Datasets: Zhaopin (Aliyun Person-Job Fitting competition) UEM (Sun Yat-sen University university-enterprise collaboration records)
Benchmarks: Zhaopin UEM
Metrics: Recall@5 Precision@5 NDCG@5 MRR@5
Baselines: BPRMF NCF LFRR LightGCN PJFNN BPJFNN APJFNN DPGNN
Builds on: TransR BPR DPGNN LightGCN
Closest papers by meaning
- Online Reciprocal Recommendation with Theoretical Performance Guarantees (NeurIPS 2018 · similarity 0.89)
- Semantic Convergence: Harmonizing Recommender Systems via Two-Stage Alignment and Behavioral Semantic Tokenization (AAAI 2025 · similarity 0.85)
- REASONER: An Explainable Recommendation Dataset with Comprehensive Labeling Ground Truths (NeurIPS 2023 · similarity 0.85)
- Bridging the User-side Knowledge Gap in Knowledge-aware Recommendations with Large Language Models (AAAI 2025 · similarity 0.85)
- Tree of Preferences for Diversified Recommendation (NeurIPS 2025 · similarity 0.85)
- Counterfactual Task-augmented Meta-learning for Cold-start Sequential Recommendation (AAAI 2025 · 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.