Discovering Temporal Features and Relations of Activity Patterns

TLDR

Mining data with temporal dependencies is a key challenge in emerging domains, and activity discovery and recognition—used in assisted living and security systems—has been extensively studied, yet temporal features and relations remain underexploited by current algorithms. This study proposes a temporal model to uncover activity patterns’ temporal features and relations from sensor data. The algorithm identifies activity order, typical start times, and durations via rule mining and clustering, and was validated on four months of smart‑home sensor data.

Abstract

An important problem that arises during the data mining process in many new emerging application domains is mining data with temporal dependencies. One such application domain is activity discovery and recognition. Activity discovery and recognition is used in many real world systems, such as assisted living and security systems, and it has been vastly studied in recent years. However, the temporal features and relations which provide useful insights for activity models have not been exploited to their full potential by mining algorithms. In this paper, we propose a temporal model for discovering temporal features and relations of activity patterns from sensor data. Our algorithm is able to discover features and relations, such as the order of the activities, their usual start times and durations by using rule mining and clustering techniques. The algorithm has been validated using 4 months of data collected in a smart home.