COMID is an online algorithm which can ensure the structure of L1 regularization.Its stochastic convergence rate can be obtained directly from the regret bound in online settings.However

the derived final solution has poor sparsity because it only takes the form of averaging all previous

T

iterates.Naturally

the individual solution has nice sparisity.So it becomes more and more important to discuss individual convergence rates in the stochastic learning.In this paper

we focus on the regularized non-smooth loss problems.When the regularizer are L1 and L1+L2

we prove the individual convergence rates of COMID respectively.The extensive experiments on large-scale datasets demonstrate that the individual solution consistently improves the spa

rsity while keeping almost the same accuracy.For the datasets with poor sparse structure

the sparsity of solution is improved even up to four times.