Using Graphs to Improve Activity Prediction in Smart Environments Based on Motion Sensor Data

Activity Recognition in Smart Environments presents a difficult learning problem. The focus of this paper is a 10-class activity recognition problem using motion sensor events over time involving multiple residents and non-scripted activities. This paper presents the results of using three different graph-based approaches to this problem, and compares them to a non-graph SVM approach. The graph-based approaches are generating feature vectors using frequent subgraphs for classification by an SVM, an SVM using a graph kernel and nearest neighbor approach using a graph comparison measure. None demonstrate significantly superior accuracy compared to the non-graph SVM, but all demonstrate strongly uncorrelated error both against the base SVM and each other. An ensemble is created using the non-graph SVM, Frequent Subgraph SVM, Graph Kernel SVM, and Nearest Neighbor. Error is shown to be highly uncorrelated between these four. This ensemble substantially outperforms all of the approaches alone. Results are shown for a 10-class problem arising from smart environments, and a 2-class one-vs-all version of the same problem.