Distribution Learning for Molecular Regression
Published in arXiv preprint arXiv:2407.20475, 2024
Citation: Nima Shoghi, Pooya Shoghi, Anuroop Sriram, Abhishek Das, arXiv preprint arXiv:2407.20475, 2024 https://arxiv.org/abs/2407.20475
This research proposes Distributional Mixture of Experts (DMoE), a novel method for molecular property regression that trains models to predict probability distributions of targets. By combining cross-entropy and L1 distance in the loss function, DMoE addresses key biases present in existing soft target regression methods. Evaluation on multiple datasets and architectures demonstrates the effectiveness of DMoE as a promising alternative to classical regression for molecular property prediction.