FSE: a Powerful Feature Augmentation Technique for Classification Task

Neural networks are powerful at discovering the hidden relation, such as classifying facial expressions to emotions. The performance of the neural network is typically limited by the number of informative features. In this paper, a novel feature augmentation is proposed for generating new informative features in an unsupervised manner. Current data augmentation focuses on synthesizing new samples according to data distribution. Instead, our approach, Feature Space Expansion (FSE), enriches data feature by providing their distribution information, which brings benefit based on model performance and convergence speed. To the best of our knowledge, FSE is the first feature augmentation method, which is developed based on feature distribution. We evaluate FSE performance on face emotion dataset and music effect dataset. We provide diverse comparisons with different alternative baselines. The experimental results indicate FSE provides significant improvement in model’s prediction accuracy when the number of features in original dataset is relatively small, and less remarkable improvement when the number of features in original dataset is large. In addition, training on FSE augmented training set can have at least ten times faster convergence speed than training on original training set.