Non-popular Items for Accurate & Diverse Linear Auto-Encoding Recommenders.

High-dimensional linear models are some of the best performing collaborative filtering models today. They learn full-rank item embeddings by inverting the Gram matrix calculated from the input user-item matrix. The entries of this Gram matrix are item co-occurrence counts which are unbounded. As a result, the Gram matrix is dominated by larger co-occurrence counts of popular items. In this paper, we propose to alleviate this issue by incorporating cosine similarity with the co-occurrence counts. We show that this increases the recommender diversity and more non-popular items are recommended. We also show that this increase in diversity correlates with an increase in accuracy signifying that the newly recommended non-popular items are relevant. Finally, we also present a more efficient procedure to obtain the parameters of linear auto-encoding recommender and show that it reduces the running time by at least half on the three standard publicly available datasets used for this line of research.