Efficient Depth-First Search Approach for Mining Injective General Episodes

Abstract

Frequent episode mining is useful for finding temporal patterns in sequential data. Episodes represent partially ordered sets of event-types and frequently occurring episodes can capture temporal dependencies in the data. There are many algorithms in the literature that find a subset of frequent episodes that best summarize the data using serial episodes (which are episodes with total order). But there are no such algorithms for the case of general episodes. An efficient Depth-First search (DFS) approach for mining general episodes would be a crucial tool and a necessary first step for generating a subset of general episodes that best represent the given temporal data. In this paper, we present an efficient algorithm to mine for general injective episodes in a DFS manner. Our algorithm uses both the apriori principle and the idea of bidirectional evidence to prune the search space and it returns closed frequent episodes. Through simulation studies, we show that the algorithm is quite effective and efficient when compared with the existing algorithms.