Competitive Feature Extraction for Activity Recognition based on Wavelet Transforms and Adaptive Pooling
Mubarak G. Abdu-Aguye, Walid E. Gomaa · 2019
Any application of machine learning requires feature extraction, whereupon the source data is processed to yield representations that are germane to obtaining the desired output. Traditional approaches to feature extraction include the estimation of statistical and structural properties of the data, the choice of which is mainly influenced by domain knowledge. However, Deep Learning has made it possible to learn the best features directly from the data itself, when sufficient data is available. For domains such as activity recognition where such plentiful data may be lacking, the application of deep learning may be limited. Therefore, methods which can yield deep learning-like performance without the need for massive amounts of data remain of interest.In this work we present a novel approach to feature extraction for sensor-generated activity recognition data. We first process the data using wavelet transforms, and subsequently use an adaptive pooling operator on the generated decomposition to obtain a compact, fixed-length representation of the data. Our experiments on seven different activity recognition datasets yield results comparable to those obtained from a deep neural network for all the considered datasets without the need for large amounts of data or any training overhead.