Instant exceptional model mining using weighted controlled pattern sampling

When plugged into instant interactive data analytics processes, pattern mining algorithms are required to produce small collections of high quality patterns in short amounts of time. In the case of Exceptional Model Mining (EMM), even heuristic approaches like beam search can fail to deliver this requirement, because in EMM each search step requires a relatively expensive model induction. In this work, we extend previous work on high performance controlled pattern sampling by introducing extra weighting functionality, to give more importance to certain data records in a dataset. We use the extended framework to quickly obtain patterns that are likely to show highly deviating models. Additionally, we combine this randomized approach with a heuristic pruning procedure that optimizes the pattern quality further. Experiments show that in contrast to traditional beam search, this combined method is able to find higher quality patterns using short time budgets.