Matrix factorization methods, ranging from common and garden varieties (eg Principal Component Analysis; PCA, and Non-negative Matrix Factorization; NMF), to more exotic beasts,
are widely used to help summarize and interpret data.
Often a single method is used in any given analysis, although different methods may give complementary insights. This talk will review some of the factors that affect interpretability of matrix factorization results, including non-negativity, sparsity, orthogonality and data transformations. We illustrate the ideas using several applications arising in genomics, although we expect many of the ideas to apply more generally.