Non-Negative Matrix Factorization (NMF)

Non-Negative Matrix Factorization (NMF) is a way to break a large table of data into a small set of meaningful building blocks. The key idea is simple but powerful: everything is constrained to be non-negative, so the model can only combine parts by adding them, not by canceling one part with a negative amount of another.

Suppose your data matrix X contains only non-negative values, such as word counts in documents, pixel intensities in images, or product purchases by customers. NMF approximates it as X ≈ WH, where:

• W stores a small set of latent components or “parts”
• H stores how strongly each data point uses those parts

Because W and H must also stay non-negative, the factors are usually easier to interpret than methods like PCA. Instead of directions with positive and negative loadings, NMF gives additive patterns: a document can be built from topics, or a face image from parts like eyes, nose, and mouth. The model is fit by minimizing reconstruction error, commonly with coordinate descent or multiplicative update rules.

In unsupervised learning, NMF is valuable when you want a lower-dimensional representation that people can actually read and reason about. It is widely used for:

• Topic discovery: documents become mixtures of interpretable topics
• Recommendation: user-item behavior is summarized by latent preference factors
• Image compression: images are represented using reusable visual parts
• Bioinformatics: gene expression patterns are decomposed into biological signatures

NMF works best when the data is naturally non-negative and additive. Choosing the number of components is important: too few gives blurry, merged patterns; too many creates noisy or redundant parts. In practice, many people use sklearn.decomposition.NMF in scikit-learn, selecting a loss such as Frobenius norm or Kullback–Leibler divergence depending on the data. Its appeal is not just compression, but interpretable structure hidden inside large unlabeled datasets.