Distributed Modeling in a MapReduce Framework for Data-Driven Traffic Flow Forecasting

Abstract

With the availability of increasingly more new data sources collected for transportation in recent years, the computational effort for traffic flow forecasting in standalone modes has become increasingly demanding for large-scale networks. Distributed modeling strategies can be utilized to reduce the computational effort. In this paper, we present a MapReduce-based approach to processing distributed data to design a MapReduce framework of a traffic forecasting system, including its system architecture and data-processing algorithms. The work presented here can be applied to many traffic forecasting systems with models requiring a learning process (e.g., the neural network approach). We show that the learning process of the forecasting model under our framework can be accelerated from a computational perspective. Meanwhile, model fusion, which is the key problem of distributed modeling, is explicitly treated in this paper to enhance the capability of the forecasting system in data processing and storage.