Abstract: Aiming at the problems of time-consuming, low efficiency and inability to select the optimal base learners in the traditional Stacking algorithm to manually select the base learner, a Stacking algorithm based on dynamic clustering was proposed and applied to the sales forecasting task. Firstly, the silhouette coefficient values were calculated for the outputs of multiple initial base learners with different numbers of clusters by the silhouette coefficient method. Then, the cluster number with the largest coefficient value was dynamically selected for k-means clustering. After each round of clustering, the reward value was given according to the error between each cluster center and the label value. Finally, the base learner contained in the cluster with the largest reward value was selected as the optimal base learner. The experimental results show that the proposed algorithm reduces the Root Mean Square Percentage Error (RMSPE) by 1.3 percentage points and the Mean Absolute Percentage Error (MAPE) by 1.0 percentage points compare with the Stacking algorithm based on feature fusion; the RMSPE is reduced by 1.1 percentage points, and the MAPE is reduced by 0.8 percentage points compare with the Stacking algorithm based on Analytic Hierarchy Process(AHP), which improves the accuracy of the algorithm.

Key words: ensemble learning, dynamic clustering, time series data, deep learning, predictive models

摘要： 针对传统Stacking算法手动选择基学习器存在耗时长、效率低和无法选择最优基学习器的问题，提出了一种基于动态聚类的Stacking算法并将其应用于销量预测任务中。首先通过轮廓系数法对多个初始基学习器的输出以不同的簇数计算其轮廓系数值；然后动态选择系数值最大时的簇数进行k-means聚类，每轮聚类后根据各簇心与标签值的误差给予回报值奖励；最后选择回报值最大的簇所包含的基学习器作为最优基学习器。实验结果表明，所提算法与基于特征融合的Stacking算法相比,均方根百分比误差(RMSPE)减少了1.3个百分点，平均绝对百分比误差(MAPE)减少了1.0个百分点；与基于层次分析的Stacking算法相比，RMSPE减少了1.1个百分点，MAPE减少了0.8个百分点，提升了算法的准确度。

关键词: 集成学习, 动态聚类, 时序数据, 深度学习, 预测模型