We present the Research Project: Actor2Critic RL & Attention on item&user embeddings.
For the MovieLens1M dataset, we adhere to a conventional approach: a warm start where 80% of the data is allocated for training, with the remainder to validation. We consider only ratings above 3 (i.e., 4 and 5) and include users who have viewed a minimum of 20 movies, as this prevents the model from becoming stagnant.
To score our models we used hitrate@10 and nDCG@10 metrics.