58
v1v2 (latest)

Logistic Regression Regret: What's the Catch?

Abstract

We address the problem of the achievable regret rates with online logistic regression. We derive lower bounds with logarithmic regret under L1L_1, L2L_2, and LL_\infty constraints on the parameter values. The bounds are dominated by d/2logTd/2 \log T, where TT is the horizon and dd is the dimensionality of the parameter space. We show their achievability for d=o(T1/3)d=o(T^{1/3}) in all these cases with Bayesian methods, that achieve them up to a d/2logdd/2 \log d term. Interesting different behaviors are shown for larger dimensionality. Specifically, on the negative side, if d=Ω(T)d = \Omega(\sqrt{T}), any algorithm is guaranteed regret of Ω(dlogT)\Omega(d \log T) (greater than Ω(T)\Omega(\sqrt{T})) under LL_\infty constraints on the parameters (and the example features). On the positive side, under L1L_1 constraints on the parameters, there exist algorithms that can achieve regret that is sub-linear in dd for the asymptotically larger values of dd. For L2L_2 constraints, it is shown that for large enough dd, the regret remains linear in dd but no longer logarithmic in TT. Adapting the redundancy-capacity theorem from information theory, we demonstrate a principled methodology based on grids of parameters to derive lower bounds. Grids are also utilized to derive some upper bounds. Our results strengthen results by Kakade and Ng (2005) and Foster et al. (2018) for upper bounds for this problem, introduce novel lower bounds, and adapt a methodology that can be used to obtain such bounds for other related problems. They also give a novel characterization of the asymptotic behavior when the dimension of the parameter space is allowed to grow with TT. They additionally establish connections to the information theory literature, demonstrating that the actual regret for logistic regression depends on the richness of the parameter class, where even within this problem, richer classes lead to greater regret.

View on arXiv
Comments on this paper