ClassiNet -- Predicting Missing Features for Short-Text Classification