Numerical similarity-aware data partitioning for recommendations as a service

At present, recommendations become an acceptable choice to replace annoying and widespread advertisements. Recommender systems (RS) are mainly used by large scale e-businesses, such as Amazon [1] and Netflix [4], because implementing and deploying RS can involve substantial investments. Many e-commerce businesses prefer to outsource the recommendation services. Therefore, Recommendation as a Service (RaaS) becomes a newly emerging trend for providing a feasible RS alternative. The providers of these services, the RS providers, need to pay the fee of cloud computing services, which is proportional to the amount of time, memory requirement, and computation resources. In addition, the RS providers must support rapidly recommendation services to meet the requests of clients. In this paper, we propose a numerical similarity-aware data partitioning (NSDP) scheme that effectively uses the numeric and similarity of datasets to exactly estimate the memory and the computation requirements for distributing the workloads. The simulation results demonstrate that NSDP significantly improves the speedup performance and achieves high scalability in the RaaS distributed-memory environment.