Nonnegative Matrix Factorization in the Component-Wise L1 Norm for Sparse Data

A team of researchers including Giovanni Seraghiti and Nicolas Gillis has published a significant advancement in Nonnegative Matrix Factorization (NMF), a core machine learning technique for finding interpretable, low-dimensional representations of data. Their new model, dubbed weighted L1-NMF (wL1-NMF), replaces the standard least-squares error measure with a component-wise L1 norm. This makes the model robust to heavy-tailed noise and outliers—common in real-world data—and strongly enforces sparsity in the resulting factors, which aids interpretability. A key innovation is a penalty parameter that controls sparsity, preventing the model from being misled by 'false zeros' in the data.

Their second major contribution is a novel Sparse Coordinate Descent (sCD) algorithm designed specifically for this model. sCD solves subproblems using a weighted median algorithm, and its computational complexity scales linearly with the number of nonzero entries in the input matrix. This is a breakthrough for efficiency, as it allows wL1-NMF to be applied to massive, sparse datasets—like recommendation systems, text corpora, or genomics data—where most entries are zero, without paying a computational cost for the empty space. Extensive experiments on synthetic and real-world data demonstrate the model's effectiveness and the algorithm's speed.

The paper also establishes important theoretical foundations, proving that L1-NMF is NP-hard even for a rank-1 factorization, highlighting the intrinsic complexity of the problem their algorithm tackles. By providing both a more robust statistical model and a computationally efficient solver, this work equips data scientists with a better tool for extracting clear, sparse patterns from messy, large-scale datasets where traditional NMF methods may struggle.