multilinearRegression

Matlab code for bilinear and trilinear least-squares regression

---

This repository contains code for least squares and ridge regression problems where the regression weights are parametrized bilinearly or trilinearly. Formally, bilinear regression solves the problem:

$\hat w = \arg \min_{\vec w} || \vec Y - X \vec w||^2_2 + \lambda ||\vec w||^2_2$,

subject to the constraint that $\vec w = \mathrm{vec}(UV^\top)$, for some matrices $U$ and $V$, where $\vec Y$ are the outputs, $X$ is the design matrix, and $\lambda$ is the ridge parameter.

Problem settings

• "mixed" bilinear regression, where we allow some coefficients of the regression weights $\vec w$ to be parametrized bilinearly, while others are parametrized linearly.

• "multiple" bilinear regression, where we allow different bilinear parametrizations (e.g., with different rank) for different segments of the regression weights

• trilinear regression, where the regression weights are parametrized by a low-rank 3rd order tensor.

Algorithms

There are implementations of two different methods for solving the optimization problem for $\hat w$:

• Alternating coordinate ascent - this involves alternating between closed-form updates for $U$ and $V$ until convergence.

• Joint ascent - direct simultaneous gradient ascent on $U$ and $V$.

See demo_bilinearRegress.m for a speed comparison; the optimal method seems to depend on the choice of dimensions and rank.