Mining Predictive Patterns and Extension to Multivariate Temporal Data

[摘要] An important goal of knowledge discovery is the search for patterns in the data that can help explaining its underlying structure. To be practically useful, the discovered patterns should be novel (unexpected) and easy to understand by humans. In this thesis, we study the problem of mining patterns (defining subpopulations of data instances) that are important for predicting and explaining a specific outcome variable. An example is the task of identifying groups of patients that respond better to a certain treatment than the rest of the patients.We propose and present efficient methods for mining predictive patterns for both atemporal and temporal (time series) data. Our first method relies on frequent pattern mining to explore the search space. It applies a novel evaluation technique for extracting a small set of frequent patterns that are highly predictive and have low redundancy. We show the benefits of this method on several synthetic and public datasets.Our temporal pattern mining method works on complex multivariate temporal data, such as electronic health records, for the event detection task. It first converts time series into time-interval sequences of temporal abstractions and then mines temporal patterns backwards in time, starting from patterns related to the most recent observations. We show the benefits of our temporal pattern mining method on two real-world clinical tasks.

[发布机构] the University of Pittsburgh