Scalable Nonlinear Learning with Adaptive Polynomial Expansions

TL;DR

This paper introduces a highly efficient algorithm that adaptively expands features to learn nonlinear representations quickly, matching linear learning times, and demonstrates superior computational tradeoffs in experiments.

Contribution

The paper presents a novel adaptive polynomial expansion algorithm that significantly improves nonlinear learning efficiency while maintaining computational speed comparable to linear models.

Findings

Algorithm achieves fast nonlinear learning with linear time complexity.

Experimental results show favorable tradeoffs against strong baselines.

Method effectively expands features adaptively for better representation.

Abstract

Can we effectively learn a nonlinear representation in time comparable to linear learning? We describe a new algorithm that explicitly and adaptively expands higher-order interaction features over base linear representations. The algorithm is designed for extreme computational efficiency, and an extensive experimental study shows that its computation/prediction tradeoff ability compares very favorably against strong baselines.