Binary Classification of Sequences Possessing Unilateral Common Factor with AMS and APR

Most real-world sequence data for binary classification tasks appear to possess unilateral common factor. That is, samples from one of the classes occur because of common underlying causes while those from the other class may not. We are interested in resolving these tasks using convolutional neural networks (CNN). However, due to both the technical specification and the nature of the data, learning a classifier is generally associated with two problems: (1) defining a segmentation window size to sub-sequence for sufficient data augmentation and avoiding serious multiple-instance learning issue is non-trivial; (2) samples from one of the classes have common underlying causes and thus present similar features, while those from the other class can have various latent characteristics which can distract CNN in the learning process. We mitigate the first problem by introducing a random variable on sample scaling parameters, whose distribution’s parameters are jointly learnt with CNN and leads to what we call adaptive multi-scale sampling (AMS). To address the second problem, we propose activation pattern regularization (APR) on only samples with the common causes such that CNN focuses on learning representations pertaining to the common factor. We demonstrate the effectiveness of both proposals in extensive experiments on real-world datasets.