Weighting-Based Strategies for Classification: Analysis and Local Peak Dynamics with CART Trees

Key points

• In this work, we compare the naive approach (unweighted) with the weighting-based rebalancing strategies.1
• Somewhat surprisingly, we demonstrate that, asymptotically, as the predictor complexity increases with sample size, weighting strategies yield predictive performance identical to unweighted approaches across all evaluation metrics, up to a shift in the final threshold.1
• To go beyond, we theoretically study the estimation of a linear regression function through weighted and unweighted CART trees with limited depths.1
• Thus, to guide practitioners, we introduce an indicator that aims at signaling whether a data set may benefit from weighting.1

Sources (1)

• [1]Weighting-Based Strategies for Classification: Analysis and Local Peak Dynamics with CART Trees
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 13, 09:27 AM

  Tabular learning tasks often suffer from class imbalance, which is often addressed via weighting-based strategies.

  In this work, we compare the naive approach (unweighted) with the weighting-based rebalancing strategies.

Extractive summary: sentences quoted from the sources.