Feature-Aware Attentive Convolutional Neural Network for Sequence Processing

Sequence processing has attracted increasing attention recently due to its broad range of applications. The development of deep neural network has made great progress in research. However, the domain features used and their interactions that play crucial roles in feature learning are still under-explored. In this paper, we propose a feature-aware Sequence Attentive Convolutional Neural Network (SeqANN) that interactively learns the sequence representations by harnessing the sequence global and local information as well as the expressive domain knowledge. In specific, we develop a one-channel Convolutional Neural NetWork (CNN) and a multi-channel CNN to learn the information of global sequence and local sequence, respectively. In addition, a featureaware attention mechanism is designed to learn the interactions between features and adaptively determine the significant parts of local representations based on the sequence representations of domain features. SeqANN is evaluated on a widely-used biomedical dataset: RBP-24. Experiment results demonstrate that SeqANN has robust superiority over competitors and achieves competitive performance.