Learning Dynamic Causal Mechanisms From Non-stationary Data

With the continuous development of Internet technology,daily life activities have gen-erated a huge amount of data with great value.Mining valuable information from data is an important topic in many fields.Causality is more valuable information than correlation.If we get causal information,then we can achieve intervention,counterfactual and other actions,so finding causal relationships from data is a key research topic in many fields.The traditional approach to causal discovery is to use interventions or randomized experiments,which is in many cases expensive,time-consuming,or even impossible.So revealing causal relationships through the analysis of observational data has attracted widespread attention in machine learn-ing,philosophy,statistics,and computer science.While there are exciting opportunities for causal discovery,there are also significant challenges,one of which is the non-stationary nature of the data over time.For example,f MRI recordings are often unstable:the flow of information in the brain may change depending on the stimulus,the task and the subject's at-tention,etc.Another example is when an adult and a child watch the same television.Because adults and children watch television in different patterns,the information received by the tele-vision may change with the person watching it,and there are many similar scenarios in real life.In these scenarios,many existing causal discovery methods may fail because they assume that the causal model behind the data is fixed,which means there is a fixed joint distribution of the observed data.To better address the problems associated with non-smooth causal mechanisms,which contain changes in causal coefficients and causal direction,we define this type of mechanism as a non-smooth causal mechanism(NSCM)and introduce a Gaussian-based variational time abstraction model(GVTA).The goal of the algorithm is to detect and learn non-smooth causal mechanisms from multiple time series,specifically the model has the following two steps.First,we detect steady states from non-stationary time series using a hierarchical cyclic state-space model via RNN.Second,we estimate the causal mechanism of each steady-state using a Gaussian process algorithm based on causal discovery.The model is continuously optimized by iterating over the previous two steps so that the model finally learns the correct causal structure.Our main contributions to this thesis are the follows:-We define the type of mechanism in which the causal intensity and causal direction of the causal structure of time-series data vary with time as the Non-Stationary Causal Mechanism(NSCM),and we developed an effective approach(GVTA)to address the problem related to the non-stationary causal mechanism.-This thesis introduces a state-space model based on Gaussian processes,using different states of the state variables to represent differences in the causal mechanisms behind the data,and using recurrent neural networks as the state transfer equations in the state-space model.It allows the model to be applied to time-series data and to obtain state variables from time-series data that are non-linearly related to the data.-GVTA can detect and estimate dynamic causal mechanisms,which include changes in causal strength and causal direction.The thesis applies the model into simulated data and open datasets,and the results show that the method proposed in this thesis is more effective than the baseline approach.