Distributed Stochastic Learning for Composite Sharing Optimization in Consumer Electronics

As electronic technology advances, a growing number of consumer electronics products emerge. Accurate and trustworthy consumer electronics product recommendation can improve user experiences and thus attract more consumers. Thus, big data analytics for distributed learning for personalized consumer recommendation is crucial. Distributed learning enables fast processing and analysis of big data through a decentralized way. This manner enables consumer electronic devices to respond faster to user requests and provide more real-time and accurate information services. In this article, we study a distributed composite sharing learning problem in large-scale networked communication systems where the nodes in the whole system endeavor to minimize a finite sum of two local cost functions plus a common possibly non-smooth coupling composite term. One of the local cost functions is further constituted by the average of some component functions. Problems of this nature usually called the multi-node sharing problem emerge in a number of applications in consumer electronics, mostly evident in image processing and recognition classification, recommender systems, etc. To tackle this problem, we provide an efficient variance-reduced distributed proximal stochastic gradient algorithm (named VR-DPSG) to obtain a lower computation cost. Then, we furnish a concise demonstration that establishes its linear convergence under milder conditions in comparison to prevailing distributed stochastic works. Our results indicate that VR-DPSG matches the linear rate of the centralized ones in the absence of any additional conditions. Extensive simulations illustrate the efficacy of VR-DPSG and validate our theoretical discoveries.