Abstract:An improved fuzzy algorithm based on multi-label AdaBoost was proposed for classifing heart valve diseases.To reflect the severity of the disease,a membership function was used to map ultrasonic diagnosis to a fuzzy value in ［0, 1］,and the ultrasonic diagnosis was described by a fuzzy label matrix.Then,cosine similarity was adopted to capture the complex correlations among heart valve diseases.Besides,a supplementary label matrix was incorporated,which augments the fuzzy label matrix by exploiting the label correlations.The label space was divided into three parts:label set,relevant label set and irrelevant label set.In order to minimize ranking loss,the sample weight update speed was adjusted and weak learner was forced to reward relevant labels by setting different labels with different weight adjustment factors.The experimental results on clinical transthoracic echocardiography (TTE) data sets illustrated that the proposed algorithm in which the fuzzy label matrix was quantified on ultrasonic diagnosis results got the best classification performance.Specially,the severity of the disease such as “non” was mapped as 0,“slight” was mapped in ［0.8, 0.85］, “moderate” was mapped in ［0.85, 0.9］,“severe” was mapped in ［0.9, 1］ by the membership function.The label correlation matrix was calculated through cosine similarity on crisp label space,where value 1 denoting sick and value 0 denoting non-sick.The performance of the proposed algorithm on five multi-label evaluation criteria were better than AdaBoost.MLR algorithm,AdaBoost.MR algorithm,BPMLL algorithm,RankSVM algorithm and ML-KNN algorithm,and the predict results of the proposed algorithm were closer to real ultrasonic diagnosis results.