Neural network ranking for personalized recommendations

This work presents the implementation of a universal open-source framework for convenient work with neural network models written in PyTorch, with an emphasis on neural network ranking models. The library provides a Scikit-learn API for working with models, which is a "fit-predict" interface for learning and prediction, as well as support for Pandas DataFrame as an input data format. The framework simplifies the launch of training and inference, without requiring the creation of PyTorch-specific entities manually, as well as taking over part of the input data processing, for example, encoding categorical features and working with categories missing from the training dataset. The library allows you to work with arbitrary datasets without additional code writing, without requiring any configurations or feature descriptions, everything is done inside the framework. An important part is also the extensibility of the written library. To add new architectures, you only need to implement several abstract methods of the same type in addition to the standard methods of PyTorch models. Currently, the DCNv2 and FinalNet architectures are implemented. This framework will simplify the work for researchers who want to quickly test another hypothesis, as well as for engineers who can use any neural network models in a convenient "fit-predict" interface and without having to write specific training and inference cycles. Also, in the future, such a framework may become something like a centralized hub for neural network ranking models, which can be used to compare the quality of different architectures. The main limitation in this case is the amount of data supplied to the input models, which in the case of inference is solved by its implementation by batches.