Abstract:
Extracting high occupancy itemsets (HOIs) from transaction databases is a significant and emerging area in data mining. This process aims to identify itemsets whose occupancy exceeds a user-defined minimum threshold. Over the years, numerous studies have sought to address this challenge, yet determining an appropriate minimum occupancy threshold remains difficult. Setting the threshold too low can lead to an overwhelming number of HOIs, rendering many mining algorithms inefficient and potentially exhausting memory resources. Conversely, if the threshold is set too high, the mining algorithm may fail to identify any HOIs whatsoever. Establishing an effective minimum threshold frequently entails a time-consuming generate-and-test process. This work introduces a novel method for mining top-k HOIM. The primary innovation of the algorithm lies in its candidate itemset pruning strategy. Specifically, if transactions containing items from the newly expanded itemset have the same length, the algorithm omits the Upper Bound (UBO) calculation, thereby significantly enhancing computational efficiency. In cases where UBO calculation is required, the process concludes as soon as the threshold is reached, without completing the entire loop. Additionally, the method leverages the comparison of UBO values stored in the Estimated Occupancy Co-Occurrence Structure (EOCS) matrix to facilitate early pruning of extended itemsets comprising two items. The proposed algorithm utilizes a depth-first search technique in conjunction with an equivalence class to process potential itemsets. Experiments conducted on both synthetic and real-world datasets show that this innovative approach is 20-25 times faster and requires 2 times less memory than baseline algorithms.
Read more