Prediction of Network Traffic Using Local Auto-regressive Methods Based on Self-organizing Map Neural Network

With a focus on network traffic forecasting, we propose a class of local auto-regressive (AR) methods based on a self-organizing map (SOM) neural network. As a generalization of the temporal domain of the SOM associative memory technique, and using the vector-quantized temporal association memory (VQTAM) modeling technique as a basis, the AR-SOM builds multiple local linear AR models, the K-SOM computes the K first winning neurons to build a single time-variant local AR model with weight vectors instead of input vectors, and the local linear map (LLM)-SOM simultaneously updates the coefficient vectors of the local models with clustering data vectors. In contrast to the global model, the proposed SOM-neural-network-based local AR methods are then applied to various network traffic forecasting instances, and are flexible enough to present an effective superficial neural architecture as well as low computation complexity. Compared to existing prediction methods, our experimental results confirm that the proposed methods may significantly improve the accuracy of prediction and yield considerably better performance.