Traffic prediction plays an important role in evaluating the performance oftelecommunication networks and attracts intense research interests. Asignificant number of algorithms and models have been proposed to learnknowledge from traffic data and improve the prediction accuracy. In the recentbig data era, the relevant research enthusiasm remains and deep learning hasbeen exploited to extract the useful information in depth. In particular, LongShort-Term Memory (LSTM), one kind of Recurrent Neural Network (RNN) schemes,has attracted significant attentions due to the long-range dependency embeddedin the sequential traffic data. However, LSTM has considerable computationalcost, which can not be tolerated in tasks with stringent latency requirement.In this paper, we propose a deep learning model based on LSTM, called RandomConnectivity LSTM (RCLSTM). Compared to the conventional LSTM, RCLSTM achievesa significant breakthrough in the architecture formation of neural network,whose connectivity is determined in a stochastic manner rather than fullconnected. So, the neural network in RCLSTM can exhibit certain sparsity, whichmeans many neural connections are absent (distinguished from the fullconnectivity) and thus the number of parameters to be trained is reduced andmuch fewer computations are required. We apply the RCLSTM solution to predicttraffic and validate that the RCLSTM with even 35% neural connectivity stillshows a strong capability in traffic prediction. Also, along with increasingthe number of training samples, the performance of RCLSTM becomes closer to theconventional LSTM. Moreover, the RCLSTM exhibits even superior predictionaccuracy than the conventional LSTM when the length of input traffic sequencesincreases.
展开▼
