Seminario del IC - Noviembre

Adaptive Decision Tree Methods

Matías D. Cattaneo

Professor of Operations Research and Financial Engineering

Princeton University

Martes 26 de noviembre a las 14hs

Cero+Infinito, 1er piso, Salón 2119 (Esquina)

Resumen:

Adaptive Decision Trees are a popular methodology in modern machine learning and data science. While they are widely used in practice, many of their theoretical and methodological properties remain unknown. This talk presents negative and positive results concerning the statistical properties of different variants of adaptive decision tree procedures. First, it is demonstrated that classical adaptive decision trees implemented using CART are pointwise (and hence uniformly) over the feature space inconsistent. For example, it is discussed how this finding implies important negative implications for heterogeneous causal inference analysis or personalized recommendation systems. Second, it is shown that adaptive oblique trees achieve near optimal mean square convergence under specific conditions, making them competitive relative to one-layer neural network procedures.