Recommendation systems have become an irreplaceable tool that helps users navigate through vast amounts of information available online by suggesting potential items of interest. Learning to Rank is a standard technique to implement recommendation systems where each item in the catalog is assigned a score, indicating how relevant the item is to the user, and displaying the Top-K items (called slate) with the highest scores. However, this method does not optimize the slate as a whole, and therefore, it fails to learn the dependencies between items, which can result in generating slates with similar items. This lack of diversity can cause fewer interactions with the slate, which might lead the user to lose interest.
We extend the work of https://arxiv.org/pdf/1803.01682.pdf where the authors proposed a paradigm shift from traditional greedy ranking to direct slate generation framework using a conditional variational autoencoder model. Our work focuses on extending the model to generate personalized recommendations using implicit feedback. We also introduce a new hyperparameter that acts as a trade-off between relevance and diversity. As the hyperparameter's value increases, the diversity of the slate increases, generally at the cost of relevance.
No publicly available dataset provides slates that are needed to train our model. Therefore we had to synthetically generate slates that reflect actual user views from the popular MovieLens dataset. We experimented with 3 types of slate generation processes, and the competitive results were only obtained when the model is conditioned to predict the users' future interactions. In all of our experiments, our model outperforms several strong baselines with a relative improvement of 28% and 62% on average in Precision@5 and Recall@5, respectively. Our model's recommendations are generally more diverse, on average, having 16% more categories than the best performing baseline.