AutoML for Contextual Bandits

Contextual Bandits is one of the widely popular techniques used in applications such as personalization, recommendation systems, mobile health, causal marketing etc . As a dynamic approach, it can be more efficient than standard A/B testing in minimizing regret. We propose an end to end automated meta-learning pipeline to approximate the optimal Q function for contextual bandits problems. We see that our model is able to perform much better than random exploration, being more regret efficient and able to converge with a limited number of samples, while remaining very general and easy to use due to the meta-learning approach. We used a linearly annealed e-greedy exploration policy to define the exploration vs exploitation schedule. We tested the system on a synthetic environment to characterize it fully and we evaluated it on some open source datasets to benchmark against prior work. We see that our model outperforms or performs comparatively to other models while requiring no tuning nor feature engineering.