The Potential of Household Specific Feature Selection for Analysing Smart Home Time-Series Data

Abstract:
The smart home domain is seen as a growth area with manifold applications, such as services for energy management, comfort, or safety and security. Many of such applications rely in part on machine learning to build models for the given homes. However, each home has an individual setup and installations may vary greatly amongst each other. In this paper we analyse the potential of automatically adapting the feature selection to drive household specific model learning. We illustrate the benefits of household specific feature selection along a sample case of device recognition using real world data. Furthermore we present and test an approach for automatic feature selection form time-series data, based on wrapper methods and genetic optimization.