A running archive of work in progress: seminar presentations, close readings of individual papers, and short methodological notes. Entries here are working documents rather than finished outputs — completed work moves to Research and Publications. Where a note records a seminar given by someone else, the presenter is credited in the entry.
Seminar Interpretable machine learning models: estimating preference coefficients with neural networks
How should the heterogeneity of individual preferences be recovered when preferences are never directly observed? The note compares MDN-Logit, which estimates a mixture distribution over attribute coefficients conditioned on individual characteristics, with MAPL, which estimates the distribution of aggregate preference per alternative. Both are placed within the 2023–2026 literature under the view that no model removes distributional assumptions — it only relocates them — and a line of follow-up work is proposed that would combine flexibility of functional form with interpretability at the coefficient level.
discrete choicetaste heterogeneitymixture density networkinterpretable MLvalue of time