在线深度神经网络的弱监督概念漂移检测方法

Weakly Supervised Concept Drift Detection Method for Online Deep Neural Networks

流数据作为大数据的重要形式广泛存在于实际问题中,由于流数据中数据分布变化产生概念漂移,容易导致模型的泛化性能下降,且在实际应用问题中,数据标记成本较高,难以获得强监督的信息.针对以上问题,本文提出一种基于在线深度神经网络的弱监督概念漂移检测(Weakly supervised conceptual drift detection method based on online deep neural network,WSCDD)方法.该方法设计了一种在线深度神经网络模型,采用Hedge反向传播方法在线学习网络深度,并通过设计Dropout层在模型预测时引入随机性,利用蒙特卡罗方法量化深度神经网络模型的预测不确定性,通过自适应滑动窗口技术检测弱监督环境下概念漂移的发生,并使模型适应新的概念.实验结果表明,该方法可以准确检测数据流中概念漂移的发生,在漂移发生后能够快速收敛到新的数据分布,提高了学习模型的泛化性能.

As an important form of big data,streaming data is widely present in practical problems,due to the change of data distribution in streaming data,conceptual drift occurs,which easily leads to the decline of the generalization performance of the model,and in practical application problems,the cost of data labeling is high,and it is difficult to obtain strongly supervised information.To solve the above problems,this paper proposes a Weakly supervised conceptual drift detection based on online deep network(WSCDD)method.This method designs an online deep neural network model,uses the Hedge backpropagation method to learn the network depth online,introduces randomness into the model prediction by designing the Dropout layer,quantifies the prediction uncertainty of the deep neural network model by Monte Carlo method,detects the occurrence of concept drift in the weakly supervised environment through the adaptive sliding window technology,and adapts the model to the new concept.Experimental results show that the proposed method can accurately detect the occurrence of conceptual drift in the data flow,and can quickly converge to a new data distribution after the drift occurs,which improves the generalization performance of the learning model.