基于偏相关性测试的递归式因果推断算法

Recursive Causal Inference Algorithm Based on Partial Correlation Test

因果推断是挖掘事物间联系的一种重要方式,但在高维数据场景下,利用因果推断算法进行条件独立性(CI)测试存在冗余测试多和测试效率低的问题,这限制了因果推断在高维数据集上的应用。提出一种基于偏相关性测试的递归式因果推断算法。采用“分治”的方法对变量集进行递归式因果分割,得到更易于处理的低维子数据集,提高对数据集的处理效率。在每个子数据集上进行局部因果推断,减少每次因果推断的计算量并提升算法的运行速度。在此基础上,通过比较显著性值的合并策略整合所有子结果并得到完整的因果关系,保证总体因果结构的准确性。在“分治”过程中,采用高效的偏相关性测试避免高复杂度的核密度估算,进一步提升算法效率。基于10个经典数据集的实验结果表明,在准确率与经典推断算法CAPA持平的情况下,该算法的运算速度提升了2~10倍,且在样本量越大的数据集中提升效果越明显,证明递归式因果推断算法可以有效处理高维数据集,在保证准确率的同时提高运算效率。

Causal inference is an important tool for mining relationships between observed data points.The causal inference algorithm encounters the problems of redundant tests and low test efficiency in high-dimensional cases,which limits the application of causal inference in high-dimensional datasets.This study proposes a recursive causal inference algorithm based on partial correlation test.The strategy of‘divide and conquer’is used to perform the recursive causal segmentation of the variable set to obtain the low-dimensional sub-dataset,which is easier to handle and improves the processing efficiency of the dataset.Local causal inference is performed on each subset to reduce the computation amount for each causal inference and improve the running speed of the algorithm.Thereafter,the significant values of the merger strategy are compared to integrate all subresults and obtain a complete causal relationship to ensure the accuracy of the overall causal structure.By‘dividing and conquering’,an efficient partial correlation test is used to avoid the high complexity of kernel density estimation and further improve the efficiency of the algorithm.Experiments are performed on ten classical data sets.The results show that when the accuracy is the same as that of the classical inference algorithm,CAPA,the operation speed of this algorithm improved by two to ten times.The improvement effect is more obvious on the dataset with a larger sample size,which proves that the recursive causal inference algorithm can effectively handle high-dimensional datasets,ensure a good accuracy,and improve the operational efficiency.