Fast tree inference with weighted fusion penalties

TL;DR

This paper introduces a fast, tree-structured inference method for multidimensional fusion penalties that improves interpretability and computational efficiency in large condition datasets, with proven theoretical properties and practical performance.

Contribution

It develops a homotopy algorithm for weighted fusion penalties, especially for $ extit{l}_1$ and $ extit{l}_} norms, and demonstrates its efficiency and statistical guarantees in data structure recovery.

Findings

The path of solutions forms a tree structure for uniform weights.

The homotopy algorithm exactly recovers the tree structure for certain norms.

The method outperforms competitors in speed and accuracy in simulations.

Abstract

Given a data set with many features observed in a large number of conditions, it is desirable to fuse and aggregate conditions which are similar to ease the interpretation and extract the main characteristics of the data. This paper presents a multidimensional fusion penalty framework to address this question when the number of conditions is large. If the fusion penalty is encoded by an $\ell_q$-norm, we prove for uniform weights that the path of solutions is a tree which is suitable for interpretability. For the $\ell_1$ and $\ell_\infty$-norms, the path is piecewise linear and we derive a homotopy algorithm to recover exactly the whole tree structure. For weighted $\ell_1$-fusion penalties, we demonstrate that distance-decreasing weights lead to balanced tree structures. For a subclass of these weights that we call "exponentially adaptive", we derive an $\mathcal{O}(n\log(n))$ homotopy algorithm and we prove an asymptotic oracle property. This guarantees that we recover the underlying structure of the data efficiently both from a statistical and a computational point of view. We provide a fast implementation of the homotopy algorithm for the single feature case, as well as an efficient embedded cross-validation procedure that takes advantage of the tree structure of the path of solutions. Our proposal outperforms its competing procedures on simulations both in terms of timings and prediction accuracy. As an example we consider phenotypic data: given one or several traits, we reconstruct a balanced tree structure and assess its agreement with the known taxonomy.