From redundant association rules to product networks: A scalable confidence-improvement pruning and Maximum Spanning Tree approach.

Abstract

Association rule mining is a powerful tool for market basket analysis, yet it frequently produces an overwhelming number of redundant and overlapping rules that hinder practical interpretation. This study presents a scalable hybrid framework that integrates efficient rule reduction with network-based structural analysis to transform large rule sets into concise, actionable product networks. The proposed approach proceeds in two stages. First, association rules are generated using the Apriori algorithm and redundant rules are substantially pruned using a confidence-improvement criterion. The procedure groups rules by their consequent and compares nested antecedent sets, which provides a scalable approximation of structural redundancy detection. Second, the reduced rule set is projected into a weighted product graph using the composite score Lift × Confidence, from which a Maximum Spanning Tree (MaxST) is extracted to identify the highest-scoring non-redundant associations among product categories. The methodology was applied to a large real-world retail dataset from Montenegro comprising approximately 2.64 million basket transactions and 14.9 million product records, aggregated into 46 product groups. The framework achieved a significant reduction in rule volume (e.g., from 403,817-137,186 rules at maxlen = 6) while preserving structurally important relationships. In the baseline maxlen = 3 configuration, the resulting MaxST identified Delicatessen as the dominant hub and uncovered coherent purchasing chains linking delicatessen, beauty and personal care, fresh meat, and healthy produce categories. Results demonstrate that the hybrid method helps bridge the gap between exhaustive rule mining and interpretable network insights, providing retailers with empirically grounded indications for cross-category promotions, bundle design, and shelf-layout testing. The approach combines computational efficiency with managerial relevance, making it suitable for large-scale retail analytics applications.