Research On Interpretability Multi-objective Evaluation Of Clustering Algorithms

With the arrival of the era of big data,artificial intelligence technology has leapt to a new level,promoting the intelligent process of human production and life.People’s demand for computational power in algorithmic models continues to increase,but in practical applications,they face many challenges,such as inevitable deviations in decision-making based on black box algorithmic models,and their decision-making processes and results are difficult to understand and trust.Only when people understand the decision-making mechanism of algorithm models can they trust their decisions and apply them to some important fields.Therefore,this thesis studies the interpretability of clustering algorithms based on traditional multi-objective evaluation methods,proposes three quantitative indicators of interpretability of clustering algorithms,and establishes a multi-objective evaluation system for the interpretability of clustering algorithms.In complex decision-making processes,the most satisfactory clustering algorithm is selected by referring to the decision-making preferences of all individual participants.This article first conducts in-depth research and analysis on the connotation,category,practical significance,and other related theories of clustering algorithm interpretability,determines the main objectives of clustering algorithm interpretability evaluation,and introduces in detail the five most classic clustering algorithms,four MCDM methods,hierarchical analysis,80-20 principles,and the experimental environment Weka platform,providing a theoretical basis for evaluation.Secondly,based on the objective evaluation method,the interpretability evaluation index of clustering algorithm is constructed,and the specific methods for quantifying the interpretability index of clustering algorithm are analyzed.It is pointed out that the subjective evaluation method and the objective evaluation method,which are two interpretable research methods,should choose the latter.Through the literature research method,it is pointed out that interpretable quantitative research can be carried out from seven aspects.This article presents the quantitative characteristics of six objective evaluations,laying a solid foundation for the establishment of indicators.Based on this,this thesis proposes clustering purity from the perspective of accuracy,PSI coefficient from the perspective of sensitivity,and Jaccard coefficient from the perspective of stability.These three indicators are used to quantify the interpretability of clustering algorithms.Finally,empirical analysis was conducted on 20 UCI datasets.Firstly,based on the Weka platform,the clustering solutions of each algorithm are obtained for the dataset,and then the clustering solutions are converted into understandable values using three interpretable quantitative indicators.Combining expert opinions,the weight of each indicator is determined using analytic hierarchy process.Based on the weighted interpretability index values,using four classic MCDM methods,each dataset will obtain an interpretable algorithm ranking,generate multiple contingency table rankings,and combine the 80-20 principle to obtain a final algorithm priority list to select the clustering algorithm with the best interpretability,and coordinate individual differences between their evaluations to obtain a ranking that can be used for interpretability evaluation.This system can coordinate different or even conflicting evaluation performance,thereby achieving group consensus in a complex decisionmaking environment.