Fast frequent episode mining based on finite-state machines

Abstract

Frequent Episode Mining (FEM) techniques play an important role in data mining, and have multiple applications, spanning from identifying user marketing habits to performing anomaly detection in computer networks. Most of the FEM approaches exhaustively search for frequent patterns, while using a threshold to efficiently reduce the search space. While this approach provides efficient results in small datasets, it fails in large datasets due to heavy processing, which leads to low performance. This paper, proposes a fast frequent episode mining method which utilizes Finite-State Machines (FSM). Initially, a FSM is created based on a subset of the data, in order to approximate the type and frequency of the most dominant episodes. Instead of applying traditional exhaustive search procedures, the parsing of the dataset is herein guided by the proposed FSM approach. Experimental results show that the proposed approach has better time performance than the traditional FEM algorithms, while still maintaining high accuracy.