IMPROVING SENSOR NETWORK PREDICTIONS THROUGH THE IDENTIFICATION OF GRAPHICAL FEATURES
Akter, Syeda Selina
MetadataShow full item record
We propose a framework that represents sensor network data as a graph, extracts graphical features, and applies feature selection methods to identify the most useful features to be used by a classifier for prediction tasks in a particular sensor network. The purpose of this graph-based framework is to provide a generic tool for sensor network application builders and practitioners to improve prediction task performance in general through the use of inherent graph structure that exists in sensor networks and through the use of generic graph-based features. We apply our graphical feature based approach to three different kinds of sensor network applications with different prediction tasks: activity recognition from motion sensors in a smart home, demographic prediction from GPS sensor data in a smart phone, and activity recognition from GPS sensor data in a smart phone. For smart home activity recognition, our graphical feature-based approach using Support Vector Machine outperformed three widely used methods, Naïve Bayes, Hidden Markov Model and Conditional Random Fields and other previous graph-based approaches on three different datasets from three smart apartments. For demographic prediction from smart phone sensors, we evaluated our approach on the Nokia Mobile Phone dataset for the three classification tasks: gender, age-group and job-type. Our approach produced comparable results with most of the state of the art methods while having the additional advantage of general applicability to sensor networks without using sophisticated and application-specific feature generation techniques or background knowledge. In activity recognition using smart phone sensors, we find that adding graph-based features using GPS to basic smart phone sensor data improves activity recognition accuracy compared to using only basic non-graphical features with existence of nodes performing the best. Adding selected edges as features reduced error for some activities. We can conclude that the graphical feature-based framework based on sensor categorization, node and edges as features, and feature selection techniques provides promising results compared to non-graph-based features.