A Novel Pruning Strategy for Mining Discriminative Patterns

Abstract

Discriminative patterns are sets of characteristics that differentiate multiple groups from each other, for example, successful and unsuccessful medical treatments. The objective of the discriminative pattern mining task is to discover a set of significant patterns that occur with disproportionate frequencies in different class-labeled datasets, generally dataset $D^{ + }$ against dataset $D^{ - }$. The discriminative pattern mining task faces two important problems: (1) the large search space problem where the search space exponentially increases with the number of items, and (2) the redundancy problem where the discriminative power of many patterns mainly derives from their sub-patterns. The common method to overcome the large search space problem is to discover frequent patterns in $D^{ + }$ and to use them as candidate discriminative patterns. In this paper, (1) we introduce a novel pruning strategy to reduce the search space. This strategy generates a new dataset $D^{new}$ = $D^{ + }$ − $D^{ - }$ and employs frequent patterns in it as candidate discriminative patterns. Following this idea, another problem appears: how to implement this idea efficiently? (2) Note that we do not explicitly calculate $D^{ + }$ − $D^{ - }$. To directly mine the frequent patterns in $D^{ + }$ − $D^{ - }$, we propose a prefix-tree, dubbed DDP-tree. This tree is directly built from $D^{ + }$ and $D^{ - }$, and contains the essential information about frequent patterns in $D^{ + }$ − $D^{ - }$. (3) To show the effectiveness of this strategy, we propose an algorithm, dubbed DiffNRDP-Miner (DiffNRDP-Miner: Difference based non-redundant discriminative pattern miner.), based on it. The advantages of DiffNRDP-Miner are that it removes redundant patterns and only needs to set one parameter, unlike other algorithms where several parameters must be set. Experimental results on benchmark datasets demonstrate that: this strategy (1) generates good patterns, where most of them are discriminative, (2) significantly reduces the search space, and (3) does not decrease the discriminative information of patterns.