Homogeneous Partitioning
A modeling approach that partitions feature space into regions with consistent conditional distributions, derived from sufficient dimension reduction and partition-based modeling ideas and typically realized via decision trees, clustering, or density-based segmentation methods.
Timeline (last 90 days)
The two-step method proposed in the paper is: first, learn a homogeneous partition (atom structure) from the data; second, greedily select atoms to maximize KL divergence under a given support constraint.
A homogeneous partition means dividing the feature space into several regions such that the conditional distribution of the target variable is the same within each region; if two data points differ in all feature values but have identical conditional distributions, they belong to the same atom.
Experiments in the paper show that even when the partition is imperfect, greedy selection still yields improvements, and the method is robust to model error.
All Facts (3)
The two-step method proposed in the paper is: first, learn a homogeneous partition (atom structure) from the data; second, greedily select atoms to maximize KL divergence under a given support constraint.
50%UnverifiedA homogeneous partition means dividing the feature space into several regions such that the conditional distribution of the target variable is the same within each region; if two data points differ in all feature values but have identical conditional distributions, they belong to the same atom.
50%UnverifiedExperiments in the paper show that even when the partition is imperfect, greedy selection still yields improvements, and the method is robust to model error.
50%