Automatic Lifestate Identification and Clustering

Introduction & BackgroundSummarising high-dimensional time series data across multiple entities is an increasingly prevalent problem because mass collection has become routine in most domains. We propose a method of automatically summarising data.
 Objectives ApproachSummarization such context both with regard to reduction the observations and large number temporal points. While numerous methods segment and/or summarise exist, properties often do not align needs consumers summaries or require unrealistic setting parameters. Addressing this, we define set broad that lead high utility class domains, which are determined by information theoretic notion optimality. Intuitively these reflect summarization into lifestates where (1) possible limited shared allow interpretation comparison (2) lifestate-transitions jointly controlled provide parameterless, optimal sample dimensionality.
 Relevance Digital FootprintsExample include: regular survey collection, consumer purchasing history from transactional (where items choose high), other repeatedly sampled digital data. Within Footprints domain, concise descriptions (summarizations) extremely important. For example, within health records could be identified used find critical patterns decline recovery patients.
 Conclusions ImplicationsThis work aims segmentations optimally trade off states segments humans must then interpret, while still capturing salient state changes. Building on prior work, model complexity normalised maximum likelihood (NML). In short, proposed generates automated summarizations informationally rich, according theory, branch mathematics.