Contextual-Bandit-Regret-Risk-Dataset

Simulated cumulative regret trajectories and regret-blowup risk for 2800 contextual and multi-armed bandit deployments across seven algorithms (epsilon-greedy, UCB1, Thompson Sampling, LinUCB, contextual TS, GLM bandit, regression oracle), five India-relevant tasks, and reward-drift conditions, grounded in the classical and modern bandit regret-bound literature. Forward-looking task predicts regret blowup from a read at just 2.5 percent of the deployment horizon.

About Dataset

This dataset simulates cumulative regret growth for bandit algorithms under stationary and drifting reward conditions. The simulation spans 2800 configurations across five India-relevant tasks (vernacular content recommendation, dynamic pricing, agri-advisory messaging, adaptive clinical trials, ad placement), seven algorithms with published regret-bound guarantees, arm or context scale, and drift rate, with an explicit flag for whether a drift-aware mechanism is used. Ten grounding sources are documented in bandit_theory.py and the README Research basis, including UCB1, Thompson Sampling, LinUCB, and Besbes, Gur and Zeevi's non-stationary bandit theory.

Purpose of Dataset

Lets Teams Running Live Recommendation, Pricing, Or Ad-serving Bandits Estimate, From A Very Early Read, Whether Their Deployment's Regret Is Headed For A Blowup Due To Unhandled Reward Drift, Before Enough Rounds Pass To Confirm It Visually.