A hybrid framework for sequential data prediction with end-to-end optimization

We investigate nonlinear prediction in an online setting and introduce a hybrid model that effectively mitigates, via end-to-end architecture, the need for hand-designed features manual selection issues of conventional prediction/regression methods. In particular, we use recursive structures to extract from sequential signals, while preserving state information, i.e., history, boosted decision trees produce final output. The connection is fashion jointly optimize whole architecture using stochastic gradient descent, which also provide backward pass update equations. employ recurrent neural network (LSTM) adaptive feature extraction data boosting machinery (soft GBDT) effective supervised regression. Our framework generic so one can other deep learning architectures (such as RNNs GRUs) machine algorithms making long they are differentiable. demonstrate behavior our algorithm on synthetic significant performance improvements over methods various real life datasets. Furthermore, openly share source code proposed method facilitate further research.