Adaptive RL Recommender simulating boredom, novelty, and satiation via Q-learning. Dynamically balances feature diversity and appeal, avoiding over-repetition. Visualizes genre fatigue over time. Generalized for any item-feature CSV (e.g., movies, music, products).
Adaptive RL Satiation Recommender This project demonstrates an Adaptive Reinforcement Learning (Q-learning) engine for boredom-aware, novelty-seeking recommendations, modeling both individual item appeal and genre/feature-level satiation. Originally shown on Netflix movie data, the model is applicable to any dataset with items and feature tags (e.g. genres, topics, categories, moods).
Features
Adaptive Q-learning engine for maximizing appeal and minimizing boredom. Dynamically balances feature (genre) diversity and item re-use. Models “cooldown” and genre-level fatigue/recovery. Fully interactive [Plotly] visualization of genre satiation score dynamics. Easily extends to any item-feature CSV.
Quickstart
Place your CSV data (e.g., netflix_titles.csv) in the project directory.
Edit the FILENAME variable if your data file has a different name.
Install dependencies: pip install numpy plotly
This works for music, books, news, products, and any item-tag data.
Output Console log: Recommendation sequence, policy and satiation stats for each run. Plot: Interactive visualization showing how genre satiation/fatigue changes over time.
License Licensed under MIT License.