B0434
Title: Learning causal representations with Granger rotated PCA
Authors: Gherardo Varando - Universitat de Valencia (Spain) [presenting]
Homer Durand - Universitat de Valencia (Spain)
Miguel-Angel Fernandez-Torres - Universitat de Valencia (Spain)
Jordi Munoz-Mari - Universitat de Valencia (Spain)
Maria Piles - Universitat de Valencia (Spain)
Gustau Camps-Valls - Universitat de Valencia (Spain)
Abstract: Causal analysis over spatiotemporal, and generally high dimensional temporal data, is usually performed over a reduced variable space obtained by dimensionality reduction techniques such as principal component analysis (PCA) or the varimax rotated PCA. The feature extraction is thus generally disconnected and learned independently from the causal task, which leads to a quite arbitrary selection of the modes of variability explaining the phenomena. Granger-rotated PCA is proposed as a new supervised dimensionality reduction method able to extract the component most causally related with respect to an exogenous forcing signal. This is achieved by directly optimizing the Granger test statistic and a closed-form and efficient solution is obtained. The proposed Granger-rotated PCA is compared to PCA, varimax, partial least squares, and canonical correlation analysis over synthetic data. The results show that Granger PCA is able to extract the causal-related component in a variety of complex settings while other methodologies fail. Finally, the proposed feature extraction is applied to study teleconnection patterns between the ENSO index, soil moisture, and vegetation indices in Africa, retrieving known connection patterns.
