A0967
Title: Supervised learning of binary features by decision trees for weightless neural networks
Authors: Douglas Cardoso - University of Coimbra (Portugal) [presenting]
Abstract: An original method based on decision trees for the supervised learning of binary features is discussed. As a case study, it was used to improve the performance of weightless neural networks (WNNs), whose most usual mathematical modeling presumes binary inputs, unlike other well-known neural networks theorized to directly operate on real-valued data. A systematic methodology is introduced for the extraction of binary features through decision tree learning, optimizing for both classification accuracy and feature sparsity, leveraging the power of decision trees to efficiently identify discriminative binary features from input data. Experimental evaluation on various classification datasets from the OpenML CC18 benchmark suite confirms the effectiveness of the approach in generating compact yet discriminative representations, which provided a statistically significant accuracy gain compared to the arguably most used binarization method for WiSARD, a reference model of WNN.
