A Neural-embedded Choice Model: TasteNet-MNL Modeling Taste Heterogeneity with Flexibility and Interpretability

Discrete choice models (DCMs) require a priori knowledge of the utility\nfunctions, especially how tastes vary across individuals. Utility\nmisspecification may lead to biased estimates, inaccurate interpretations and\nlimited predictability. In this paper, we utilize a neural network to learn\ntaste representation. Our formulation consists of two modules: a neural network\n(TasteNet) that learns taste parameters (e.g., time coefficient) as flexible\nfunctions of individual characteristics; and a multinomial logit (MNL) model\nwith utility functions defined with expert knowledge. Taste parameters learned\nby the neural network are fed into the choice model and link the two modules.\n Our approach extends the L-MNL model (Sifringer et al., 2020) by allowing the\nneural network to learn the interactions between individual characteristics and\nalternative attributes. Moreover, we formalize and strengthen the\ninterpretability condition - requiring realistic estimates of behavior\nindicators (e.g., value-of-time, elasticity) at the disaggregated level, which\nis crucial for a model to be suitable for scenario analysis and policy\ndecisions. Through a unique network architecture and parameter transformation,\nwe incorporate prior knowledge and guide the neural network to output realistic\nbehavior indicators at the disaggregated level. We show that TasteNet-MNL\nreaches the ground-truth model's predictability and recovers the nonlinear\ntaste functions on synthetic data. Its estimated value-of-time and choice\nelasticities at the individual level are close to the ground truth. On a\npublicly available Swissmetro dataset, TasteNet-MNL outperforms benchmarking\nMNLs and Mixed Logit model's predictability. It learns a broader spectrum of\ntaste variations within the population and suggests a higher average\nvalue-of-time.\n