Contribution of low support association rules in understanding the mined knowledge

In the realm of numerical association rule mining using population-based nature-inspired algorithms, the evaluation of results usually depends on a crucial metric – a fitness value crafted from weighted support and confidence measures. Traditionally, this metric separates the mined rules into a spectrum, ranging from the low-quality, low-fitness rules to the peaks of their high-quality, high-fitness counterparts. Until recently, little attention has been directed towards low-fitness rules, as users engage predominantly with high-quality rules. However, in our pioneering research, we go into the enigmatic realm of low-support, more precisely, the often-overlooked association rules. Through meticulous analysis of our rule repository, we seek to uncover the profound insights concealed within these seemingly unremarkable rules. Our method was applied to the Abalone UCI ML dataset, where three association rules, mined using the universal association rule miner based on evolutionary algorithms, were taken into consideration, i.e., the AR-1 was mined in the first generation, the AR-2 mined in the last generation, and the AR-3 mined at the moment before the phenomenon of the disappearing features has arisen. Significantly, the analysis of the AR-3 revealed that the association rules of the lower support can contribute to understanding the mined knowledge.