Causal structure learning for travel mode choice using structural restrictions and model averaging algorithm

This work contributes to develop a new methodology to identify empirical-data-driven causal structure of a domain knowledge. We propose an algorithm as a two-stage procedure by first drawing relevant prior partial relationships between variables and using them as structure constraints in a structure learning task of Bayesian networks (BNs). The latter is then based on a model averaging approach to obtain a statistically sound BN. The empirical study focuses on modeling commuters' travel mode choice. We present experimental results on testing the design of prior restrictions, the effect of resampling size and learning algorithms, and the effect of random draw on fitted BN structure. The results show that the proposed method can capture more sophisticated relationships between the variables that are missing in both decision tree models and random utility models.