School of Information Technology and Engineering

Permanent URI for this collection

http://etd.aau.edu.et/handle/123456789/460

Browse

Browsing School of Information Technology and Engineering by Subject "Association rules"

Now showing 1 - 1 of 1
  • No Thumbnail Available
    Item
    A Hybrid Two-Stage Neuro-Symbolic Framework for Explainable Recommendation in Sparse Data
    (Addis Ababa University, 2026-02) Ermias Alemayehu; Beakal Gizachew
    Large-scale recommender systems commonly employ two-stage pipelines in which a fast retriever generates a candidate set and a more expressive reranker produces the final top-K recommendations. In sparse implicit-feedback settings, however, end-to-end performance is constrained by candidate availability, and explanation mechanisms are often weakly connected to ranking decisions. This thesis proposes a hybrid two-stage neuro-symbolic framework for explainable recommendation in sparse data that integrates a LightGCN retriever, train-only symbolic candidate expansion based on association-style co-occurrence rules, and an evidenceaware MLP reranker under a fixed candidate-budget constraint. The framework is evaluated using Candidate Recall@C, HR@K, NDCG@K, explanation coverage (EC@K), and a fidelity-style evidence-removal diagnostic to assess whether symbolic features influence ranking behavior under fixed candidates. Experiments\ on Amazon Books, MovieLens-1M, and Yelp show that the impact of symbolic expansion and explainable reranking is dataset-regime dependent; in retrievalbottlenecked settings, improvements in Candidate Recall@C are accompanied by corresponding gains in HR@K and NDCG@K, consistent with a retrieval-ceiling interpretation of two-stage recommendation. On Amazon Books, the full configuration achieves Candidate Recall@C = 0.3135/0.2527, HR@10 = 0.0698/0.0388, and NDCG@10 = 0.0391/0.0201 on validation/test, with EC@10 approximately 0.86. Overall, the thesis contributes a neuro-symbolic two-stage recommendation framework and an evaluation strategy that jointly analyzes retrieval constraints, ranking effectiveness, explanation coverage, and evidence dependence under fixed serving budgets.