基于后验异常检测的GRU网络在线模型

A GRU Network Online Learning Model Based on Posterior Anomaly Detection

在规律变化迅速的类似股票的时间序列的预测问题上,传统离线模型无法即时对自身进行调整而导致模型准确度降低,而传统在线模型也会因为异常数据导致模型不稳定,为提高神经网络模型对时间序列的预测准确度和模型的稳定性,文章提出了一种基于后验异常检测的门控循环单元(Gated Recurrent Unit GRU)在线学习模型。该模型在网络在线学习前加入基于贝叶斯的后验异常检测算法,从而避免在线网络模型利用异常值迭代。经过30次实验的平均结果显示,加入后验异常检测的GRU网络在线学习模型要优于离线模型。而没有异常检测算法的模型则在异常值出现后会短期失效,虽然在随后正常数据的到来会逐渐修正模型,但最终的准确度反而不如离线模型,因此模型有效地提高了准确性和稳定性。

On the prediction of time series of similar stocks with fast changing laws,the traditional offline model can not adjust itself in time,which leads to the decrease of model accuracy,while the traditional online model is also unstable due to abnormal data.In order to improve the prediction accuracy of the neural network model to time series and increase the stability of the model,this paper presents an online learning model of gated recurrent unit(GRU)based on posterior anomaly detection.In this model,Bayesian posterior anomaly detection algorithm is added before online learning,so as to avoid online network model using outliers iteration.The average results of 30 experiments show that the GRU network online learning model with posterior anomaly detection is better than the offline model.However,the model without anomaly detection algorithm will fail for a short time after the occurrence of outliers.Although the model will be gradually revised with the subsequent arrival of normal data,the resulting accuracy is not as good as that of the offline model.Therefore,the proposed model effectively improves the accuracy and stability.