Dynamic Frechet Regression with Feature Selection for Distributional Data
Kiran Adhikari, Amrutha Dinesh, Mathew Kuttolamadom, Ying Lin
Read on arXiv →
Key claim
DFR improves predictions of dynamic distributional responses.
In plain English
Many applications generate responses that are complex statistical objects rather than simple numbers. Current regression methods struggle to relate these complex responses to scalar predictors, especially when the responses change over time or other indices. Dynamic Fréchet Regression (DFR) addresses this by modeling these responses with an index-aware approach, allowing for more accurate and interpretable predictions. Builders might find this useful for analyzing data that evolves over time, such as in manufacturing processes.
Introduces a novel framework for modeling distribution-valued responses over ordered indices.
Demonstrates improved predictive accuracy and feature recovery through simulations and a real-world application.
Deep reliability assessment
The methodology supports modeling index-dependent trajectories of distribution-valued responses with improved predictive accuracy and feature recovery. However, the claims about interpretability and general applicability may be overextended without broader empirical validation across diverse datasets.
Reproducibility
No open source code or dataset is mentioned in the paper.
Key figure
Figure 1 visually confirms distinct forms of distributional evolution in both location and spread across three scenarios: Linear, Quadratic, and Periodic trends.