Toward an Incremental Classification Process of Document Stream Using a Cascade of Systems.

In the context of imbalanced classification, deep neural networks suffer from the lack of samples provided by low represented classes. They can’t train enough their weights with a statistically reliable set. All solutions in the state of the art that could offer better performance for those classes, sacrifice in return a huge part of their precision on bigger classes. In this paper, we propose a solution to this problem by introducing a system cascade concept that could integrate deep neural network. This system is designed to keep as mush as possible the original network performance while it reinforces the classification of the minor classes by the addition of stages with more specialised systems. This cascade offers the possibility to integrate few-shot learning or incremental architecture following the deep neural network without major restrictions on system internal architecture. Our method keeps intact or slightly improves the performances of a deep neural network (used as first stage) in conventional cases and improves performances in strongly imbalanced cases by around +8% accuracy.