Recommendations with uncertainty

Abstract

The goal of this project is to develop an advanced sequential recommendation system by making it aware of its uncertainty when making predictions. In this thesis, we attempt to use Reconstruction Uncertainty Estimation (RUE) technique to estimate the prediction uncertainty of the Self-Attentive Sequential Recommendation (SASRec) model. We then use the uncertainty information to perform post-hoc training in an attempt to calibrate SASRec. Throughout this thesis, we will be using the MovieLens-1M, a dataset that has been frequently used in recommendation systems research, including the original SASRec paper. Additionally, this paper will explore challenges encountered and the difficulty in using RUE to estimate uncertainty in sequential recommendation systems.