Document Type
Presentation
Publication Date
8-1-2014
Abstract
Causal inference or causal relationship discovery is an important task in hydrological study to explore the causes of abnormal hydrology phenomena such as drought and flood, which will help improving our prediction and response ability to natural disasters. Different from generic causality study where causalrelation discovery is sufficient, for extreme hydrological situation prediction and modeling, we need not only to construct a causal graph to reveal the contributing factors, but also to provide the lead time of each cause to its effect. Lead time is the time difference between the occurrence of lead and effect. Though causal inference or causal relationship discovery has been a major topic in many science problems, majority of the work has been focused on the validity of such relationship with no knowledge on cause-effect time lead information. Such insight is critical for hydrological modeling and prediction, in which time lead information is desired for knowing how long different factors will affect certain extreme situations such as flood or drought. The most commonly used computational algorithms for causality discovered can be categorized as using regression approaches or Bayesian approaches. Regression based approaches such as Granger's causality assume linear causality and first order causal relationship. Bayesian approaches, such as the PC algorithm from Pearl's causality definition, have exponential runtime complexity which makes it difficult to be applied to hydrological systems with a high number of variables. Furthermore, no existing approaches incorporate the lead time concept in the discovery of causal relationship. In this paper, we propose a new approach, mutual information causal (MI-Causal), for causal relationship discovery, which embodies the advantages of existing approaches and overcomes the limitations to satisfy the hydrologic need. The experimental results from both synthetic and real time hydrological data show that our proposed method outperforms regression approaches and Bayesian based approaches.
Comments
Session S5-02, Special Session: Computational Intelligence in Data Driven and Hybrid Models and Data Analysis II