Multilabel Text Classification of Unbalanced Datasets: Two-Pass NNMF

The natural distribution of textual data used in text classification is often imbalanced. Categories with fewer examples are under-represented and their classifiers trained on the datasets transformed to bag-of-words representations or basic topic modeling transformations often perform far below a satisfactory level. We tackle this problem using a two-pass non-negative matrix factorization algorithm. This approach finds topics for each category independently allowing to better define topics for underrepresented categories. The results are analyzed from multiple goal perspectives - H-loss, accuracy, F-measure, precision, and recall, from the micro, macro and example-based aspect since each is appropriate in different situations. Through experimental validation, it is shown that the two-pass matrix factorization improves classification results achieved using bag-of-words representations.