Project idea · ungraded

Bandit-Based Recommender

Build a recommender that picks which article to show next, learning from simulated clicks. Compare random choice, epsilon-greedy and UCB by their cumulative regret, and see what happens when reader interests change over time.

You will use

Milestones

Build the smallest working version first. Ticks are saved in this browser only.

01Simulate 10 articles with hidden click probabilities and a reader who clicks with those probabilities.
02Implement random, epsilon-greedy and UCB recommenders with the same interface.
03Run each for 10,000 recommendations over 20 seeds and plot average cumulative regret.
04Make the click probabilities drift slowly and add a sliding-window version of UCB.
05Write up which strategy you would deploy and why.

Stretch goals

Implement Thompson sampling with Beta distributions.
Add reader features (contextual bandits) with one linear model per article.
Serve recommendations from a FastAPI endpoint that updates on click events.

Data

Everything is simulated with seeded NumPy generators, so experiments are reproducible.

Run it yourself

pip install numpy matplotlib
python bandits.py --strategy ucb --steps 10000