BanditLP: Large-Scale Stochastic Optimization for Personalized Recommendations