基于K2算法的因果结构学习研究综述

A Survey of Causal Structure Learning Based on K2 Algorithm

贝叶斯网络是一种处理不确定性知识的有效工具,因为其能很大程度降低推理的复杂度,在数据挖掘、机器学习等领域得到了广泛的应用.贝叶斯网络结构学习是其重要研究内容之一,其中K2算法由于其能有效避免似然等价问题,以及在时间复杂度和准确度上都优于大部分经典算法,因此备受研究者关注.针对K2算法受节点序约束,采用贪婪搜索技术处理模型选择导致的寻优效率差的问题,研究者们提出了不同的改进策略,根据是否有领域知识可分为基于先验序和搜索先验序两类方法.对近几年改进的K2算法进行了调研,并对K2算法未来的研究改进做了总结和展望.

Bayesian network is an effective tool to deal with uncertain knowledge.Because it can greatly reduce the complexity of reasoning,it has been widely used in data mining,machine learning and other fields.Bayesian network structure learning is one of its important research contents.The K2 algorithm is currently attracting attention from researchers because of its ability to effectively avoid likelihood equivalence problems and its time complexity and accuracy.In view of the K2 algorithm being constrained by the variable ordering,the greedy search technology is used to deal with the problem of poor optimization efficiency caused by model selection.Researchers have proposed different improvement strategies.According to whether there is domain knowledge,it can be divided into prior order and search prior order two types of methods.This paper investigates the improved K2 algorithm in recent years,and prospects the future research improvement of the K2 algorithm.