Spatial-temporal causal modeling: a data centric approach to climate change attribution (Invited)

Attribution of climate change has been predominantly based on simulations using physical climate models. These approaches rely heavily on the employed models and are thus subject to their shortcomings. Given the physical models’ limitations in describing the complex system of climate, we propose an alternative approach to climate change attribution that is data centric in the sense that it relies on actual measurements of climate variables and human and natural forcing factors. We present a novel class of methods to infer causality from spatial-temporal data, as well as a procedure to incorporate extreme value modeling into our methodology in order to address the attribution of extreme climate events. We develop a collection of causal modeling methods using spatio-temporal data that combine graphical modeling techniques with the notion of Granger causality. “Granger causality” is an operational definition of causality from econometrics, which is based on the premise that if a variable causally affects another, then the past values of the former should be helpful in predicting the future values of the latter. In its basic version, our methodology makes use of the spatial relationship between the various data points, but treats each location as being identically distributed and builds a unique causal graph that is common to all locations. A more flexible framework is then proposed that is less restrictive than having a single causal graph common to all locations, while avoiding the brittleness due to data scarcity that might arise if one were to independently learn a different graph for each location. The solution we propose can be viewed as finding a middle ground by partitioning the locations into subsets that share the same causal structures and pooling the observations from all the time series belonging to the same subset in order to learn more robust causal graphs. More precisely, we make use of relationships between locations (e.g. neighboring relationship) by defining a relational graph in which related locations are connected (note that this relational graph, which represents relationships among the different locations, is distinct from the causal graph, which represents causal relationships among the individual variables - e.g. temperature, pressure- within a multivariate time series). We then define a hidden Markov Random Field (hMRF), assigning a hidden state to each node (location), with the state assignment guided by the prior information encoded in the relational graph. Nodes that share the same state in the hMRF model will have the same causal graph. State assignment can thus shed light on unknown relations among locations (e.g. teleconnection). While the model has been described in terms of hard location partitioning to facilitate its exposition, in fact a soft partitioning is maintained throughout learning. This leads to a form of transfer learning, which makes our model applicable even in situations where partitioning the locations might not seem appropriate. We first validate the effectiveness of our methodology on synthetic datasets, and then apply it to actual climate measurement data. The experimental results show that our approach offers a useful alternative to the simulation-based approach for climate modeling and attribution, and has the capability to provide valuable scientific insights from a new perspective.