A new mini-batch negative momentum proximal stochastic variance reduction method for nonconvex optimization

First-order stochastic optimization algorithms have been widely applied to large-scale machine learning tasks. We have proposed a new stochastic optimization algorithm, which is inspired by an accelerated stochastic variance reduced method originally developed for convex optimization. The proposed algorithm addresses nonconvex and nonsmooth problems via the proximal operator and mitigates overfitting through a mini-batch strategy that exploits richer gradient information. We established sublinear convergence under standard assumptions. Extensive experiments on synthetic classification tasks, real-world datasets, nonconvex matrix factorization problems, and time series prediction tasks showed that the proposed algorithm consistently achieves faster convergence and competitive test performance compared to state-of-the-art stochastic optimization algorithms, underscoring its potential in practical applications for large-scale, nonsmooth learning problems.