【作者】 吴磊；

【导师】 吕恕；

【作者基本信息】 电子科技大学 ， 统计学， 2024， 硕士

【摘要】 随着大数据时代的到来,高维数据在带来丰富信息的同时,也进一步增加了分析问题的复杂性。特征选择作为维度约减的重要手段,近年来,在数据预处理阶段发挥着关键的作用。由于传统特征选择仅考虑相关性,导致模型存在不可解释性。因此,因果特征选择算法又称为马尔科夫毯发现算法应运而生,提升模型的可解释性和鲁棒性是该类算法的关注重点,因为它们可以考虑数据的生成机制。本文针对已有的因果特征选择算法中出现的问题和不足,进行了多方面的改进工作,主要的研究内容如下:(1)对因果特征选择领域进行理论研究,整理归纳现有算法研究框架,通过大量文献总结算法研究现状,大多数现有算法仅考虑提升精度或提高速度的单方面改进,因此本文提出一种同时改进精度和速度的两全算法,将直接法速度快的优势和分治法精度高的优势进行结合,提升马尔可夫毯(Markov blanket,MB)发现的效率。(2)提出将传统特征选择算法与因果特征选择算法融合的思路,充分利用传统特征选择算法的优势,高效快速地识别MB变量。具体做法如下:一种基于相关的快速过滤算法(Fast Correlation-Based Filter,FCBF)具有识别父子变量(parents and children,PC)的特性,为了充分发挥其作用,本文首先通过实验选择合适的阈值使得FCBF可以实现最小误差,并且通过实验验证,FCBF具有比传统PC变量识别算法更好的性能。(3)提出了改进的因果特征选择算法(Markov Blanket Discovery Based on Fast Filtering,MB-BF),该算法利用增量关联马尔科夫毯(Incremental Association Markov Blanket,IAMB)、FCBF进行数据的预处理工作,IAMB用于快速识别初始但不完全的MB集合,FCBF用于识别PC变量且过滤不相关和冗余的变量。并在算法执行过程中,利用“相关性越大成为PC的可能性越高”的思想执行排序策略。最终加入基于OR规则的对称性检验,提升算法在实际中的效能。实验结果表明,MB-BF在没有显著降低速度的前提下,具有比现存因果特征选择算法更高的精度,并且可以更好的应用于特征选择任务中。更多还原

【Abstract】 With the advent of the big data era,high-dimensional data,while bringing abundant information,has further increased the complexity of analyzing problems.Feature selection,as an important means of dimension reduction,has played a crucial role in the data preprocessing stage in recent years.Traditional feature selection only considers correlation,leading to the uninterpretability of the model.Therefore,causal feature selection algorithms,also known as Markov blanket discovery algorithms(MB discovery),have emerged as the times require.These algorithms consider the underlying generation mechanism of data,improving the interpretability and robustness of the model.Thesis carries out various improvement works aimed at addressing the problems in existing causal feature selection algorithms.The main work is as follows:(1)The theory of causal feature selection is reviewed.This thesis firstly summaries the definition and theoretical basis of feature selection.Then through a large number of literature research,it is found that most existing algorithms only consider improvements in accuracy or speed individually.Therefore,this thesis proposes a comprehensive algorithm that simultaneously improves both accuracy and speed by combining the advantages of simultaneous methods in terms of speed and divide-and-conquer methods in terms of accuracy,thereby enhancing the efficiency of MB discovery.(2)The idea of integrating traditional feature selection algorithms into causal feature selection algorithms is proposed.The specific approach which fully utilize the advantages of traditional feature selection algorithms and efficiently identify MB variables is as follows: A fast correlation-based filter algorithm(FCBF)has the property of identifying PC variables.To fully leverage its capabilities,thesis first selects an appropriate threshold through experiments to enable FCBF to achieve minimal error.Experimental verification demonstrates that FCBF performs better than traditional PC identification algorithms.(3)An improved causal feature selection algorithm(MB-BF)is proposed.It utilizes IAMB and FCBF for data preprocessing.IAMB is used to quickly identify an initial but incomplete MB set,while FCBF is used to identify PC variables and filter out irrelevant and redundant variables.During the execution of the algorithm,a sorting strategy is implemented based on the idea that “the greater the correlation,the higher the likelihood of being a PC”.Finally,a symmetry test based on OR rules is added to enhance the algorithm’s practical effectiveness.Experimental results show that MB-BF achieves higher accuracy than existing causal feature selection algorithms without significantly reducing speed,and it can be better applied to feature selection tasks.更多还原