Provided are systems, methods and techniques for classifying items. According to one preferred embodiment, initial feature sets are obtained for a current batch of items, and classification predictions are generated for the items based on their initial feature sets, using a set of existing classifiers. The classification predictions are then appended as additional features to the respective feature sets of the items, thereby obtaining enhanced feature sets, and a first classifier is trained, using a plurality of the items as training samples and using the enhanced feature sets of the training samples. Finally, items in the current batch are classified using their enhanced feature sets and the first classifier. According to this embodiment, the existing classifiers were trained on a plurality of different sets of items that are representative of corresponding different times.
