Recommender Systems play a significant part in filtering and efficiently prioritizing relevant information to alleviate the information overload problem and maximize user engagement. Traditional recommender systems employ a static approach towards learning the user's preferences, relying on logged previous interactions with the system, disregarding the sequential nature of the recommendation task and consequently, the user preference shifts occurring across interactions. In this study, we formulate the recommendation task as a slate Markov Decision Process (slate-MDP) and leverage deep reinforcement learning (DRL) to learn recommendation policies through sequential interactions and maximize user engagement over extended horizons in non-stationary environments. We construct the simulated environment with various degrees of preferential dynamics and benchmark two DRL-based algorithms: FullSlateQ, a non-decomposed full slate Q-learning based on a DQN agent, and SlateQ, which implements DQN using slate decomposition. Our findings suggest that SlateQ outperforms by 10.57% FullSlateQ in non-stationary environments and that with a moderate discount factor, the algorithms behave myopically and fail to make an appropriate tradeoff to maximize long-term user engagement.
Please refer to the paper for an in-depth treatment on the subject matter.
Main Contributions: recsim/environments, experiments.py, main.py
This project of is a part of the 2022 Edition of CSE3000 Research Project course at Delft University of Technology.
Cd into the parent directory.
> pip install -r requirements.txtCd into the parent directory. This works on my windows machine. Could potentially differ on other OSs. Example command:
> $ python -m recsim.experiments --agent_name=full_slate_q --episode_log_file='episode_logs.tfrecord' --base_dir=/recsim/experiments/ To view the results of the experiment, first cd into the recsim sub-directory, and decide on the log path. Example command:
> $ tensorboard --logdir=recsim/tmp/recsim/ --host localhost --port 8088Due to the high amounts of computation necessary for generating statistically significant data, the result files could not be published because of storage limit.
However, should one be interested in accessing the results, send an email to luca.p.pantea@gmail.com, and I will give access to the WandB project where all the run logs are stored.
To best observe the effects of such dynamics on recommendation performance, we turn to RecSim, a configurable simulation environment, which allows for the study of RL in the context of stylized recommender settings. This repository contains all the necessary source code for the preference dynamics environment. For more examples of environments and further work in RecSim, we point the reader to the RecSim GitHub repository.
To get started, please check out our Colab tutorials. In RecSim: Overview, we give a brief overview about RecSim. We then talk about each configurable component: environment and recommender agent.