Relaxed Softmax for learning from Positive and Unlabeled data
This work addresses inefficiencies in multi-class prediction for language modeling and recommendation, though it appears incremental as it builds on existing softmax methods.
The paper tackled the problem of applying softmax losses and sampling schemes to Positive and Unlabeled learning, proposing a Relaxed Softmax loss and Boltzmann-based negative sampling that improved performance on tasks like density estimation and next-event prediction.
In recent years, the softmax model and its fast approximations have become the de-facto loss functions for deep neural networks when dealing with multi-class prediction. This loss has been extended to language modeling and recommendation, two fields that fall into the framework of learning from Positive and Unlabeled data. In this paper, we stress the different drawbacks of the current family of softmax losses and sampling schemes when applied in a Positive and Unlabeled learning setup. We propose both a Relaxed Softmax loss (RS) and a new negative sampling scheme based on Boltzmann formulation. We show that the new training objective is better suited for the tasks of density estimation, item similarity and next-event prediction by driving uplifts in performance on textual and recommendation datasets against classical softmax.