The Research Of The Optimization Method For Large Scale Convex Problem In Recommender System

Recommender System is an effective technology for information filtering,which finds the valuable information for users based on their previous information,which dis-tinguished them in many applications,such as e-commerce and the news paltformIn this paper,the development and the application of recommender sysytem are intro-duced in three aspects,that are,user's information based recommendation,content based recommendation,and the collaborative filtering.Moreover,the challenge of large scale convex problem in recommender system is presentedLater,this paper reviewed some popular optimization method and their adjustment for the challenge of data intense problem,as well as their strength and weakness.Based on Cutting Plane Method,this paper proposed two new algorithm to solve large scale convex problem:(i)For the single-machine computing environment,an algorithm called Mini-Batch Cutting Plane Method(MBCPM)is proposed,which uses a small batch of data to update the model parameters iteratively and achieved a appealing iterating speed.In order to guarantee the convergence,MBCPM adopts an operation to eliminate abnormal cutting plane.The theoretical analysis shows that MBCPM obtains a convergence rate similar to that of the standard cut plane method.Experiments based on MovieLens 20M,a real movie scoring data set,show that the objective function decreases faster in the proposed MBCPM than the standard CPM method(ii)For cluster environment,this paper proposes a distributed asynchronous opti-mization algorithm called Asynchronous Bundle Method for Large Regularized Risk Min-imization(Async BMRM).By using asynchronous window,this method allows nodes to iterate asynchronously,so it can effectively solve the problem of unbalanced comput-ing load caused by unbalanced scoring data.The theoretical analysis of Async BMRM shows that the convergence of the algorithm can be guaranteed.Experiments based on MovieLens 20M show that Async BMRM can reduce communication time and make the objective function decline faster.