Congratulations to winners and participants! It was an interesting competition with highly imbalanced classes. Not helped for me: statistical features, model stacking - my stacked models worked better on public LB and worse on private LB. But the solo model showed the opposite. So this is a simple Logistic regression model fitted on Tfidf character level and word level features. Because of a very small number of texts the rarest occurrences of n grams has been taken. Tfidf features were united with fasttext 60-length averaged word vectors for texts. Cross Validation has shown a rather good score, but it was unstable, so its public LB results were poor.

Github link:

https://github.com/CostiaB/AI4D-Malawi-News-Classification-Challenge