Discretized sequential pattern mining for behaviour classification

Abstract

This paper proposes an approach for behavioural analysis on continuous data based on sequential pattern mining (SPM) and classification. Discretization and granularity issues associated with SPM and classification with time series data are examined, and a thorough examination of the many factors and parameters that are used to control the granularity of the various dimensions of the problem is conducted. The key contribution of the paper is the seminal examination of machine learning-based classification driven by sequential pattern mining to classify behaviour patterns. Results are reported on the effectiveness of various levels and settings on acceleration data collected from the walking activity of 17 subjects obtained from a publicly available dataset. Using the optimal set of parameter settings identified, the classifier was able to correctly identify 70% percent of walking data segments from a held-out test set as belonging to a particular subject, and 78% of those not belonging to the subject. Moreover, a number of interesting findings were identified when varying the granularity and other settings for the various discretization parameters, solidifying the work as a significant advancement in the area of behavioural analysis with sequential pattern mining and classification on time series data.