Improving Ranking Quality Via Consequent Set Quartile Partitioning In Rule-Based Recommendation

Accurate ranking in session-based recommender systems remains a challenge, especially in dynamic domains such as e-commerce and digital news where ranking quality directly shapes user satisfaction and platform effectiveness. Neural models achieve high accuracy but incur heavy computational costs, while association rule (AR)-based methods are efficient yet struggle to prioritize relevant items when multiple candidates share similar weights. This study aims to improve the ranking quality of AR-based recommendation while retaining computational efficiency. A lightweight framework, ART-Q (Association Rule with Top-k Quartile Filtering), is proposed to extend traditional single top-k ranking into multiple top-k selection across quartiles. The approach follows three phases: (i) modeling—constructing an association rule dictionary from training data, maintaining antecedents as keys with consequent sets and rule weights as values; (ii) inference—partitioning each consequent set into four quartiles and selecting top-k items from each, thereby producing a multiple top-k recommendation that surfaces candidates across weight strata; and (iii) evaluation—benchmarking ART-Q against non-neural (AR, k-Nearest Neighbor variants) and neural baselines (recurrent, convolutional, and graph-based models) using standard ranking metrics. Experiments on two real-world datasets from e-commerce and e-news domains show that ART-Q consistently improves hit rate, reciprocal rank, and ranking stability while preserving computational efficiency. Nevertheless, reliance on static quartile partitioning and the limited dataset scope constrain generalizability. Future work may explore adaptive weighting strategies and broader validation to strengthen applicability across domains.