Abstract: We consider the setting of online logistic regression and consider the regret with respect to the $\ell_2$-ball of radius $B$. It is known (see Hazan et al. (2014)) that any proper algorithm which has logarithmic regret in the number of samples (denoted $n$) necessarily suffers an exponential multiplicative constant in $B$. In this work, we design an efficient improper algorithm that avoids this exponential constant while preserving a logarithmic regret. \n\nIndeed, Foster et al. (2018) showed that the lower bound does not apply to improper algorithms and proposed a strategy based on exponential weights with prohibitive computational complexity. Our new algorithm based on regularized empirical risk minimization with surrogate losses satisfies a regret scaling as $O(B\log(n) + B^2)$ with a per-round time-complexity of order $O(d^2)$.

Summary presentation

Full presentation