PRTNets: Cold-Start Recommendations Using Pairwise Ranking and Transfer Networks

In collaborative filtering, matrix factorization, which decomposes the ratings matrix into low rank user and item latent matrices is widely used. The decomposition is based on the rating scores of users to item, with the user and item latent matrices sharing a common embedding space. A similarity function between the two represents the predicted rating of a user to an item. However, this matrix factorization approach falls short for cold-start recommendation where items have very few or no ratings. This paper puts forward a novel approach of doing cold-start recommendation by using a neural network, the Transfer Network, to learn a nonlinear mapping from item features to the item latent matrix. The item latent matrix is produced by another network, the Pairwise Ranking Network, which utilizes pairwise ranking functions. The Pairwise Ranking Network efficiently utilizes implicit feedback by optimizing the ranking of the recommendation list. We find the optimal architecture for the Pairwise Network and the Transfer Network through warm-start and cold-start evaluation. With the Transfer Network, we map the Tag Genome dataset to the item latent matrix and produce cold-start recommendations for a test set derived from the MovieLens 20M dataset. Our approach yielded a significant margin of improvement of 0.276 and 0.089 average precision at \(k=10\) over the baseline LightFM and neighborhood averaging methods respectively.