A regression-based matrix-factorization method (SA-ZIG) for high-dimensional, sparse, highly skewed nonnegative co-occurrence count matrices, e.g. word-word co-occurrence counts or item-item and user-item counts. Each row/column entity is a latent dense vector; observed counts are modeled as zero-inflated Gamma variables whose mean depends, through a canonical or log link, on the cosine similarity between the paired vectors. Vectors and parameters are fit by alternating zero-inflated Gamma regressions that share parameters across the two factors, with proven convergence and a learning-rate (Adam-style) adjustment that fixes Fisher-scoring stalls.
Introduces the shared-parameter alternating zero-inflated Gamma (SA-ZIG) model for factorizing high-dimensional, sparse, highly skewed nonnegative co-occurrence count matrices, such as word-word co-o…