Online Learning Adaptive to Dynamic and Adversarial Environments

The present contribution deals with online learning of functions, where multi-kernel approaches, among other popular methods, have well-documented merits, but also face major challenges such as scalability and adaptivity. Leveraging the random feature approximation, an online multi-kernel learning scheme is developed to infer the intended nonlinear function. To account for dynamic and possibly adversarial environments, an adaptive and scalable multi-kernel learning scheme is also introduced at affordable complexity and memory requirements. Performance guarantees are provided in terms of dynamic regret analysis, while numerical tests on a Twitter dataset are carried out to showcase the effectiveness of our approach.