Sampling-Bias-Corrected Neural Modeling for Large Corpus Item Recommendations
I first learned about this paper when Google Brain released the Tensorflow Recommenders library last month. I focused on it because Google, which operates a massive recommendation system like YouTube, was releasing recommendation system-related code. The overall content is more detailed in the Tensorflow Blog, so please read it. The goals of TFRS (TensorFlow Recommenders) are as follows: Build recommendation candidates quickly and flexibly Structure that freely uses Item, User, Context information Multi-task structure that learns various objectives simultaneously Learned models are efficiently served through TF Serving Actually, the code itself doesn’t have much diverse content, but what was most impressive was the Two Tower Model introduced as the basic model in the code. It’s about training User and Item completely independently and only predicting click/unclick with dot product at the final stage. The more I think about it, the better the structure seems. Although it’s unknown whether it will show tremendous performance since user tower and item tower can’t interact during training, the structure itself has no constraints on input features, so you can freely add possible features, and during inference, you can serve efficiently by having user embeddings and item embeddings and calculating similarity only with dot product, so compatibility with ANN (Approximate Nearest Neighbors) libraries also looks good. ...