Paulina Tomaszewska
supervisor: Przemysław Biecek
Offline Deep Learning based solutions work well in production as long as the input at the prediction time has similar characteristics to data used for training. Otherwise, if the discrepancy is substantial, the model may need to be modified e.g. by fine-tuning using new data or the change of the model architecture that may require training from the scratch. Later, the previous model can be replaced by the new one. Another approach is training a model in a continuous manner on streaming data. In my work, I focus on a case where class-priors change over time. The results on the Twitter dataset show that the proposed LIMES algorithm improves the metric in a form of the minimum model accuracy within a day.