TreNet: Hybrid Neural Networks for Learning the Local Trend in Time Series

Tao Lin, Tian Guo, Karl Aberer

Abstract: Local trends of time series characterize the intermediate upward and downward patterns of time series. Learning and forecasting the local trend in time series data play an important role in many real applications, ranging from investing in the stock market, resource allocation in data centers and load schedule in smart grid. Inspired by the recent successes of neural networks, in this paper we propose TreNet, a novel end-to-end hybrid neural network that predicts the local trend of time series based on local and global contextual features. TreNet leverages convolutional neural networks (CNNs) to extract salient features from local raw data of time series. Meanwhile, considering long-range dependencies existing in the sequence of historical local trends, TreNet uses a long-short term memory recurrent neural network (LSTM) to capture such dependency. Furthermore, for predicting the local trend, a feature fusion layer is designed in TreNet to learn joint representation from the features captured by CNN and LSTM. Our proposed TreNet demonstrates its effectiveness by outperforming conventional CNN, LSTM, HMM method and various kernel based baselines on real datasets.

Conflicts: epfl.ch