Approximate Approach to Finding Generic Utility of Sequential Patterns

Majority of utility pattern mining (UPM) literature assumes that the utility of particular patterns is known beforehand. At the same time, in frequent pattern mining (FPM), all patterns assume the same value. A problem with many datasets is that information about the utility of patterns is not easily or at all available and the utility is not directly proportional to the frequency of the specific pattern. One approach to finding the utility could be formulated in terms of how a particular pattern contributes to effectiveness of a specific machine learning (ML) algorithm. In this paper, we focus in particular on sequential patterns, e.g. time series. We propose an approximate approach to finding the generic utility of sequential patterns based on intergroup separation calculated using subsequence Dynamic Time Warping (DTW) similarity. Such utility could then be used in supervised algorithms and with some limitations in unsupervised algorithms. It can be applied both for regular and streaming application. We provide a proof of effectiveness of our approach using PAMAP2 Physical Activity Monitoring Data Set, an open dataset from the UCI Machine Learning Repository. For dissimilar activities (e.g. lying and ascending the stairs) we identify subsequences with normalized utility exceeding 0.5, where 1 would be a difference between total lack of heart rate and maximal heart rate all the time. For very similar activities (e.g. Nordic walking and ascending stairs) we find subsequences with utility exceeding 0.12.