If a decision tree is constructed through a series of locally optimal solutions, such as the Greedy method, overfitting to the data is likely to occur. In order to avoid overfitting, many previous research have attempted to collectively optimize the structure of a decision tree by using evolutionary computation. However, if attributes of each split and their thresholds are searched simultaneously, the evaluation function becomes intermittent; thus, optimization methods assuming continuous distribution cannot be used. In this study, in order to enable efficient search assuming continuous distribution even for complicated data that contains a lot of noise and extraordinary values, such as financial time series data, the inner level search that optimizes each threshold value collectively given a specific attribute for each split in a model tree and the outer level search that optimizes the attributes of each split were performed by separate evolutionary computing. As a result, we obtained high prediction accuracy that far exceeded the performance of the conventional method.

如果通过一系列局部最优解来构建决策树，例如贪心法，很可能会出现对数据的过拟合。为了避免过拟合，之前的很多研究都试图通过进化计算来共同优化决策树的结构。但是，如果同时搜索每个分割的属性及其阈值，则评估函数会变得断断续续；因此，不能使用假设连续分布的优化方法。在本研究中，为了即使对于包含大量噪声和非凡值的复杂数据（例如金融时间序列数据），也能在假设连续分布的情况下进行高效搜索，优化每个阈值的内层搜索为模型树中的每个拆分共同给出特定属性，而优化每个拆分的属性的外层搜索由单独的进化计算执行。结果，我们获得了远远超过传统方法性能的高预测精度。