Conditional Random Field Feature Generation of Smart Home Sensor Data using Random Forests

Abstract

A typical approach to building a feature set for a conditional random field model is to build a large set of conjunctions of atomic tests, all of which adhere to a small number of relatively simple templates. Building more complex features in this way can be difficult, as the more complex templates needed to do this can result in a combinatoric explosion in the number of features. We use the inherent instability of decision trees to produce a small set of more complex conjunctions that are particularly suitable for the problem to be solved, using the same techniques used in generating random forest ensemble classifiers, and build a CRF on these features. We apply this method to an activity recognition problem on a dataset from the CASAS smart home project, in which we predict activities of daily living from sensor activations.