/Multi-aspect-Reinforcement-Recommendation

Course project for https://deeppavlov.ai/rl_course_2020

Primary LanguageJupyter Notebook

Reinforcement learning for Recommendation Systems

Implementation of Deep Reinforcement Learning based Recommendation with Explicit User-Item Interactions Modeling.

Some of the code were reused from Catalyst demo notebook and higgsfield's RL-Adventure, it helps a lot.

TODO:

  • Implement validation
  • Change training process (switch to sessions)
  • Add Prioritized Experience Replay

Special TODO:

  • Add Ornstein–Uhlenbeck noise for better exploration

Movielens (1M) results

Model nDCG@10 hit_rate@10
DDPG with OU noise 0.280 0.502
DDPG 0.254 0.454
Neural Collaborative Filtering 0.238 0.430
Random (for comparison) ~0.05 ~0.1

viz