Improving Future Network Traffic Prediction Accuracy Based on Multi-Feature Construction and An Attention Mechanism

Le résumé fourni par la source

Future network traffic prediction helps network operators detect possible future traffic changes and modify link capacities and routing strategies in advance. However, current network traffic prediction methods are unable to capture the short period and long period traffic change patterns simultaneously. This may lead to severe hysteresis phenomenon when predicting the link traffic in distant future, resulting in poor prediction accuracy. Therefore, this paper proposes a network traffic prediction method based on multi-feature construction and an attention mechanism. By transforming historical traffic data based on time cycles and incorporating them as new feature dimensions of the training data, the representativeness of the input data can be greatly improved. By applying return sequences on the recurrent neural network, traffic change features in different moments can be better preserved. The multiplication attention mechanism is used to weigh importance of the output of recurrent neural network, enabling the model to adaptively learn the best convergence parameters. Experimental results on GEANT and Abilene datasets showed that the proposed method achieved higher prediction accuracies when predicting traffic in the distant future and complicated traffic change patterns of network links, compared with traditional single feature recurrent neural network models.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.