Abstract
Traditional frequent pattern mining algorithms become inefficient in dynamic e-learning environments because newly arriving learner interactions require repeated full recomputation. This study proposes the Incremental Intelligent Maximal Frequent Itemsets (IIMFIs) framework, which incrementally maintains maximal frequent itemsets while incorporating novelty-based post-generation filtering controlled by a threshold . Experiments were conducted on the Open University Learning Analytics Dataset (OULAD), comprising 180,596 learner-session transactions divided chronologically into four batches. At a minimum support of 0.005, the maximal-pattern representation reduced the pattern count from 2,597 frequent itemsets to 342 maximal itemsets, corresponding to an 86.8% reduction and a 7.6- fold compression. Across evaluated support thresholds, IIMFIs-Incremental achieved speedups of 2.31× to 5.86× relative to full recomputation. At minSup=0.01, IIMFIs-Incremental required 659 ms compared with 1,071 ms for FPGrowth and 60,964 ms for Apriori. At minSup=0.005, runtimes were comparable (1,465 ms vs. 1,447 ms), indicating that the incremental advantage is more pronounced at moderate support levels. Recommendation evaluation showed comparable 1@10 values between IIMFIs and FP-Growth, while @5 was lower for IIMFIs. Recommendation diversity exceeded 0.95 for the evaluated configurations. Novelty filtering served primarily as a computational optimization. These findings indicate that IIMFIs provides a compact and efficient approach for maintaining behavioral patterns and supporting adaptive e-learning recommendations under the evaluated conditions.
Keywords: Adaptive Recommendation Systems, Educational Data Mining, Incremental Data Mining, Maximal Frequent Itemset Mining, Novelty Filtering.