Text classification with coupled matrix factorization

We introduce a coupled matrix factorization and logistic regression model for simultaneous feature extraction and binary classification in high dimensional data sets. We write the model starting from probabilistic notation of nonnegative matrix factorization and logistic regression. We derive the gradient descent multiplicative update rules from the log likelihood function. We show that the model yields better results than logistic regression on a recent Turkish text corpus of news articles, while maintaining comparable accuracy in the topic clustering problem.