一种适用于高维网络的方向推断算法

An Approach for Inferring Causal Directions in High Dimensional Causal Networks

挖掘数据间蕴含的因果性是很多科学领域中的一个基础问题.然而现在很多方法只能从数据中构造出关联型网络,暂时还没有有效的途径对高维的关联型网络进行方向推断.这里提出一种采取分治策略的算法,该算法先将高维网络中边方向的学习问题分解成每一个节点的因果网络边方向的学习问题,然后根据每个子网络的具体结构利用条件独立性测试和加噪声模型推断出每一个子网络所有边的方向.所有节点都迭代完后得到数据集的一个完整因果网络图.数据实验表明,该算法应用在高维因果网络下,效果要优于目前其他算法.

Discovering causality from dataset is one of the basic problems in many scientific fields. However,there is still no efficient and effective method to discover causality from high-dimensional data. In this work, an approach for large scale causality discovery based on divide-and-conquer strategy is proposed. Firstly, The proposed algorithm decompose an n-dimensional network into n parts according to each node. Causalities of a certain sub-network can be discoverod by methods based on conditional independence test and additive noise model. By updating the causal skeleton of a target variable in each iteration,It can construct the whole skeleton of the high dimensional network at the end of the last iteration. Experimental results show that the proposal outperforms other methods.