Temporal modeling refers to analyzing the causal relationship between given time series data and modeling the predictive function based on this, and it is widely used in various applications such as weather forecasting, market price prediction or anomaly detection.
Generally, in the case of causality analysis, Granger causality has been used for a long time, but it becomes computationally intractable when it uses a lot of past information called delay terms.
In this thesis, to solve the problem of previous Granger causality, we propose Granger neural encoding which consists of non-linear encoding and lasso Granger model.
In non-linear encoding, we use deep learning models such as CNN (convolutional neural networks) and RNN-LSTM (recurrent neural network - long short-term memory) which have state-of-the-art performance in various applications. Then we analyze the causality on those encoded time series data by using lasso Granger model. In addition to Granger neural encoding, we propose a two-staged dynamic capacity network, each consisting of a simple Granger neural encoding model for causal analysis and a complex one for precise prediction based on them.
In the end of this thesis, we verify dynamic capacity Granger neural encoding has improved modeling performance than other comparative temporal models.