End-to-end listwise machine learning ranking models for information retrieval systems

Machine learned models are prevalent in a number of recommender system and information retrieval use cases. Often the models are trained using single data points (pointwise), or using pairs of instances of different quality (pairwise). However, previous research has showed that such pointwise or pairwise loss functions can be inferior to listwise loss functions, which aim to get an entire ranking correct. Yet, many recent deep neural techniques, such as BERT text ranking models, continue to be trained in pairwise manner. In this research, the student will investigate ways to adapt listwise loss functions to modern deep learned frameworks; Evaluation can be conducted using existing recommendation and search datasets.