A Projected Upper Bound for Mining High Utility Patterns from Interval-Based Event Sequences
S. Mohammad Mirbagheri
TL;DR
This paper introduces a projected upper bound $\mathcal{P}_k$ for mining high utility patterns from interval-based event sequences (e-sequences), aiming to accelerate HUIPMiner through tighter pruning than the existing $\mathrm{LWU}_k$ bound. By defining $\mathcal{P}_k(L)=u_{max}(L)+\mathrm{LWU}_{k-|L|}(L)$ and proving $\mathcal{P}_k(L) \le \mathrm{LWU}_k(L)$, the authors establish a projected downward-closure property that dynamically tightens as pattern length grows. Empirical evaluations on the BlocksmorchenSensor dataset show that using $\mathcal{P}_k$ yields notable improvements in execution time (up to about 2x) and memory usage (a few percent on average) without losing completeness, demonstrating practical gains for high utility interval-based pattern mining. The approach advances efficient mining in domains with interval-based events by enhancing pruning while maintaining exact pattern discovery.
Abstract
High utility pattern mining is an interesting yet challenging problem. The intrinsic computational cost of the problem will impose further challenges if efficiency in addition to the efficacy of a solution is sought. Recently, this problem was studied on interval-based event sequences with a constraint on the length and size of the patterns. However, the proposed solution lacks adequate efficiency. To address this issue, we propose a projected upper bound on the utility of the patterns discovered from sequences of interval-based events. To show its effectiveness, the upper bound is utilized by a pruning strategy employed by the HUIPMiner algorithm. Experimental results show that the new upper bound improves HUIPMiner performance in terms of both execution time and memory usage.