Creator2Vec: Creator Feature Embedding for Deep Learning Recommender System

The current embedding method neglects to deeply explore the associations between the items created by the same item creator, and directly puts the long-tail distributed categorical creator feature into the model for end-to-end training. Because the tail data provides too little information, it is difficult to train stable and meaningful embedding vectors. To this end, we propose an improved embedding method Creator2Vec used for recommender systems, which is based on the items produced by the creator to characterize the creator. Specifically, we first use Word2Vec to generate item embedding, and take the average of item embedding created by the creator as creator embedding. Then, according to the quality of item comments, we design a feature extraction method weighted to mine the information of item comments and comments on item comments. Benefit from the proposed method, we not only use item ratings as features, but are also concern that item comments with the same rating and different qualities have different recommendation effects. Finally, we uses the number of likes of item reviews as a standard to measure the quality of item comments. The empirical results on a practical dataset show that Creator2Vec and weighted cumulative item rating features, as the input layer of common deep learning models, have good effects on binary classification recommendation tasks.