Adaptive Self-Sufficient Itemset Miner for Transactional Data Streams

Feiyang Tang, David Tse Jung Huang, Yun Sing Koh, Philippe Fournier-Viger

Abstract: Most studies on pattern mining consider itemsets that have a high frequency of occurrence as useful, often determined by the support of the itemsets. However, current research has shown that we need to move beyond a pure “support-confidence” framework for pattern mining. Recently, there is an interest on finding statistically significant patterns and one of the most popular type of patterns is self-sufficient itemsets. One limitation is that these works do not consider concept drifts and cannot be used in a data stream. Learning in the online environment requires us to develop efficient and effective mechanisms to address the online characteristics of non-static data and non-stationary data distributions. In our research we will concentrate on detecting self-sufficient itemsets from data streams. These patterns have a frequency that is significantly different from the frequency of their subsets and supersets. We present a comprehensive framework for mining self-sufficient itemsets from data streams along with a drift detector. This supports mining self-sufficient itemsets in an online environment and provides the ability to adapt to changes in the stream. Our experimental evaluations show that our framework can mine self-sufficient itemsets faster in an online environment and with better precision and recall.