【摘要】 针对基于深度神经网络填补多变量时序数据缺失值时存在误差引入的问题，将强化学习和图神经网络结合，提出了基于强化学习的多变量时序数据缺失值补全方法(reinforcement learning based missing value completion method for multivariate time series data, RL-CFMTS)。具体地，设计了一个图神经网络模块，用于预测最终的缺失值。同时，该模块还为强化学习模块提供状态向量和奖励，指导强化学习为目标时序节点挑选其认为最相似的b个节点来建立时序关系，以减少误差的引入。在3个真实世界数据集上的实验结果表明，RL-CFMTS在插补准确性方面优于近年来的多变量时序数据缺失值填补方法，在天气、股票和疫情数据集中平均绝对误差(mean absolute error, MAE)指标分别提升了12.78%、15.80%和55.73%。更多还原

【Abstract】 Aiming at the problem of introducing errors in filling missing values of multivariate time series data based on deep neural networks, reinforcement learning was combined with graph neural networks to propose a reinforcement learning based missing value completion method for multivariate time series data(RL-CFMTS) model. Specifically, a graph neural network module was designed to predict the final missing value. At the same time, this module could also provide the reinf orcement learning module with state vectors and rewards to guide reinforcement learning to select the most similar b nodes for the target timing node and establish a timing relationship to reduce the introduction of errors. The experimental data from three real-world sets show that RL-CFMTS outperforms the most advanced methods available in terms of interpolation accuracy, achieving performance improvements of 12.78%, 15.80%, and 55.73% in mean absolute error(MAE) indicators in weather, stock, and epidemic data sets, respectively.更多还原