The development of sequential analysis provides a deeper understanding in the exploration of many different fields. In the application of sequential analysis, there are two main challenges: How to extract informative features from a high-dimensional noisy domain? How to model the interaction for the information flow from multiple domains? We explored the two core challenges in bio-informatics, sales forecasting and multimedia services. In biology field, a typical problem is the to evaluate... Show moreThe development of sequential analysis provides a deeper understanding in the exploration of many different fields. In the application of sequential analysis, there are two main challenges: How to extract informative features from a high-dimensional noisy domain? How to model the interaction for the information flow from multiple domains? We explored the two core challenges in bio-informatics, sales forecasting and multimedia services. In biology field, a typical problem is the to evaluate the interaction mechanism between non-coding DNA sequences and transcription. We propose CANEE, a convolutional self-attention architecture to analyze the function of non-coding DNA sequences. Compared to other existing models, CANEE achieves a better performance in overall prediction of 919 regulatory functions with respect to receiver operating characteristics and has a significant improvement on some responses in precision recall curve with shorter training time. In sales forecasting field, we extract a unique customers’ microbehavior dependency structure from clickstream data based on a Word-to-Vector model. Then, we build a clickstream informed LSTM model to forecast the car sales over 30 days. Our model significantly outperforms the classic seasonal autoregressive integrated moving average model. Besides, we demonstrate that transfer knowledge among different car models can further improve the performance. Other applications for multi-domain sequences happens in multimedia service field, where we focus on the understanding of multiple domain modalities, we propose new principles for audio visual learning and introduce a new framework as well as its training algorithm to set sight of videos’ themes to facilitate AVC learning. Show less
