NeuroTree - A Differentiable Tree Operator for Tabular Data

Explore a conference talk on NeuroTree, a differentiable tree operator for tabular data, presented by Jeremie Desgagne-Bouchard at JuliaCon 2024. Dive into the innovative approach of NeuroTree, which addresses the greediness of traditional trees by simultaneously learning all nodes and leaves while incorporating the benefits of boosting and bagging through a built-in ensemble of trees. Discover how the computation of leaf weights is achieved through in-place element-wise operations and how custom reverse rules using ChainRules overcome auto-differentiation limitations for both CPU and GPU. Examine benchmarks comparing NeuroTree against state-of-the-art algorithms like XGBoost, LightGBM, CatBoost, and EvoTrees across various regression, classification, and ranking tasks. Learn about NeuroTree's performance on common regression datasets, including its top performance on the Higgs and YEAR datasets. Gain insights into the relevance of Julia's machine learning capabilities in the commercial context of portfolio management.