<p>Association rule extraction is a critical research domain in data mining. However, discovering generalizable and positively correlated association rules from numerical datasets via metaheuristic algorithms remains a significant challenge. To address these limitations, this study proposes an improved Variable-Length Multi-Objective Harmony Search algorithm for Numerical Association Rule Mining (VLMOHSNAR), which integrates three key innovations that synergistically enhance performance. Firstly, the proposed variable-length coding cooperates with the multi-objective framework to eliminate redundant pitches while maintaining rule quality. Secondly, the simultaneous optimization of support and netconf is designed to complement the variable-length coding, ensuring Pareto-optimal rules that balance generality and positive correlation. Thirdly, the pitch selection mechanism (PSM) and harmony memory update mechanism work together to improve the balance between exploration and exploitation. Extensive experiments were conducted on ten real-world datasets to evaluate the proposed method against three classical algorithms and five state-of-the-art methods. Results indicate that, for datasets such as Basketball, Bolts, Pollution, Retail-service, Quake, Stulong, Fetal-health, and House16H, average support and average netconf values exceed 0.8, confirming excellent performance. Although VLMOHSNAR performs optimally on low-dimensional data, its scalability to high-dimensional datasets remains an open challenge. The proposed method significantly advances numerical association rule mining by combining variable-length representation with multi-objective harmony search, offering a more efficient and adaptable framework for practical rule discovery.</p>