Inter-Transaction Association Rules Mining for Rare Events Prediction

Rare events prediction is a very interesting and critical issue that has been approached within various contexts by research areas, such as statistics and machine learning. Data mining has provided a set of tools to treat this problem when the size as well as the inherent features of the data, such as noise, randomness and special data types, become an issue for the traditional methods. Transaction databases that contain sets of events require special approaches in order to extract valuable temporal knowledge. Sequential analysis aims to discover patterns or rules describing the temporal structure of data. In this paper we propose an approach that extends sequential analysis to predict rare events in transaction databases. We utilize the framework of inter-transaction association rules, which associate events across a window of transactions. The proposed algorithm produces rules for the accurate and timely prediction of a userspecified rare event, such as a network intrusion or an engine failure.