【摘要】：針對(duì)心臟瓣膜疾病模糊分類問題,提出基于多標(biāo)簽Ada Boost的模糊分類改進(jìn)算法。結(jié)合模糊集理論,采用隸屬函數(shù)將疾病的嚴(yán)重程度映射到區(qū)間[0,1]內(nèi)的實(shí)數(shù)值,將超聲診斷結(jié)果用模糊標(biāo)簽向量表示。利用余弦相似性分析疾病之間的復(fù)雜關(guān)系,計(jì)算標(biāo)簽相關(guān)性矩陣并對(duì)模糊標(biāo)簽向量進(jìn)行補(bǔ)充。結(jié)合實(shí)際問題選取合適的閾值,將標(biāo)簽空間劃分為標(biāo)簽集、標(biāo)簽相關(guān)集和標(biāo)簽無關(guān)集。本文算法以最小化排序損失為目標(biāo),針對(duì)不同的標(biāo)簽給予不同的權(quán)值調(diào)整因子,調(diào)整樣本權(quán)重更新速度,強(qiáng)迫弱分類器關(guān)注與樣本標(biāo)簽相關(guān)性較高的標(biāo)簽。在臨床超聲心動(dòng)圖(TTE)測量數(shù)據(jù)集上的實(shí)驗(yàn)結(jié)果表明:在對(duì)超聲診斷結(jié)果模糊化時(shí),通過隸屬函數(shù)將疾病嚴(yán)重程度中的"無病"映射為0,"輕度"映射到區(qū)間[0.8,0.85],"中度"映射到區(qū)間[0.85,0.9],"重度"映射到區(qū)間[0.9,1],構(gòu)造模糊標(biāo)簽矩陣,并通過標(biāo)簽相關(guān)性矩陣對(duì)其進(jìn)行補(bǔ)充,此時(shí)所構(gòu)造的分類器性能達(dá)到最優(yōu)。將本文算法與Ada Boost.MLR算法、Ada Boost.MR算法、BPMLL算法、Rank SVM算法和ML-KNN算法進(jìn)行對(duì)比分析,在多標(biāo)簽分類的5種評(píng)價(jià)指標(biāo)上,本文算法的分類性能均優(yōu)于其他對(duì)比算法,分類結(jié)果更接近超聲診斷結(jié)果。
[Abstract]:An improved fuzzy classification algorithm based on multi-label Ada Boost is proposed for fuzzy classification of heart valve diseases. Based on the fuzzy set theory, the degree of disease severity is mapped to the real value in the interval by membership function, and the results of ultrasonic diagnosis are represented by fuzzy label vector. Using cosine similarity to analyze the complex relationship between diseases, the label correlation matrix is calculated and the fuzzy label vector is supplemented. The label space is divided into tag set, tag correlation set and label independent set. This algorithm aims at minimizing the ranking loss, gives different weight adjustment factors for different labels, adjusts the update speed of sample weights, and forces the weak classifier to pay attention to the labels with high correlation with the sample labels. The experimental results on the clinical echocardiographic (TTE) data set show that: when the results of ultrasonic diagnosis are blurred, The "disease-free" degree of disease severity is mapped to zero, "mild" to interval [0.88 0.85], "moderate" to interval [0.850.90], "heavy" to interval [0.99 ~ 1] by membership function, and fuzzy label matrix is constructed, which is supplemented by label correlation matrix. The performance of the proposed classifier is optimal. This paper compares this algorithm with that of Ada Boost.MLR algorithm, Ada Boost.MR algorithm and ML-KNN algorithm. The classification performance of this algorithm is better than that of other comparison algorithms on five evaluation indexes of multi-label classification. The classification results are closer to the ultrasonic diagnosis results.
【作者單位】： 中國科學(xué)院成都計(jì)算機(jī)應(yīng)用研究所;中國科學(xué)院大學(xué);
【基金】：四川省科技支撐計(jì)劃資助項(xiàng)目(2016JZ0035) 中科院西部之光人才培養(yǎng)計(jì)劃項(xiàng)目資助
【分類號(hào)】：R542.5;TP301.6
，

本文編號(hào)：2213260