In this study, we analyse two mobile phone activity datasets to predict the future traffic of mobile base stations in urban areas. The predicted time series can be used to reflect the trend of human activity flow. Although common methods such as recurrent neural network and long short-term memory (LSTM) network often achieve a high precision, they have the short back of time-consuming. So, we present the improved gradient-boosted decision tree algorithm based on Kalman filter (GBDT-KF) due to the noise in the original time series, because the decrease in the performance of GBDT is usually caused by overfitting the noise in the signal. According to our experiments, although the RMSE of the predicted values of our GBDT-KF and the ground truth is only 12–14% worse than that of the LSTM model, the proposed GBDT-KF algorithm makes a trade-off between the precision and time complexity and achieves over 100-time training time reduction compared with the LSTM model. By implementing the result of our work, service providers could predict where and when a network congestion would happen; therefore, they could take actions ahead of time. Such applications are useful especially in the era of 5G.

中文翻译：

---

基于卡尔曼滤波器（GBDT-KF）的改进梯度提升决策树算法在时间序列预测中的应用

在这项研究中，我们分析了两个手机活动数据集，以预测城市地区移动基站的未来流量。预测的时间序列可以用来反映人类活动流的趋势。虽然循环神经网络和长短期记忆（LSTM）网络等常用方法往往能达到很高的精度，但它们都存在耗时的缺点。因此，由于原始时间序列中的噪声，我们提出了基于卡尔曼滤波器（GBDT-KF）的改进的梯度提升决策树算法，因为GBDT性能的下降通常是由于信号中的噪声过度拟合引起的。根据我们的实验，虽然我们的 GBDT-KF 和地面实况的预测值的 RMSE 仅比 LSTM 模型差 12-14%，所提出的 GBDT-KF 算法在精度和时间复杂度之间进行了权衡，与 LSTM 模型相比，训练时间减少了 100 多次。通过实施我们的工作结果，服务提供商可以预测网络拥塞将在何时何地发生；因此，他们可以提前采取行动。此类应用在 5G 时代尤其有用。