相关工作mmc 两类m3fs 多类m3fs 实验结果.ppt

Multi-Class M3FS 多类SVM Multi-Class M3FS 类似两类问题，给每个特征定义一个比例因子，就可以得到多类的M3FS： p：第p类，k：第k个特征 提纲 相关工作 MMC 两类M3FS 多类M3FS 实验 实验 实验数据集 Real word data set DB： UCI：ionosphere，digits，letter，satellite Minst 说明 Clustering performance，generalization ability of M3FS Toy data sets DB： UCI：iris data set(4 feature/instance)+10 noise features 说明 Ability to detect relevant features 对比方法： Feature Selection based on Gaussian mixture model(FSGMM) Laplacian score +MMC/K-Means MMC without feature selection Ours 实验 评价指标 实验时，取有label的数据，将label去掉，进行聚类，聚类后每个cluster的label取得是cluster中样本数目最多label。 Clustering Accuracy(Acc): Cluster中正确分类的样本数目/对应class中样本的数目(the extent to which each cluster contained data points from the corresponding class) Rand Index(RI) R=(a+b)/(a+b+c+d) a: # of data pairs in the same set in both cluster and class b: # of data pairs in different set in both cluster and class c: # of data pairs in the same set in cluster, but different set in class d: # of data pairs in different set in cluster, but same set in class 实验 实验数据集 Real word data set DB： UCI：ionosphere，digits，letter，satellite Minst 说明 Clustering performance，generalization ability of M3FS Toy data sets DB： UCI：iris data set(4 feature/instance)+10 noise features 说明 Ability to detect relevant features 对比方法： Feature Selection based on Gaussian mixture model(FSGMM) Laplacian score +MMC/K-Means MMC without feature selection Real word data set 特征维数 原始特征维数 实验 实验数据集 Real word data set DB： UCI：ionosphere，digits，letter，satellite Minst 说明 Clustering performance，generalization ability of M3FS Toy data sets DB： UCI：iris data set(4 feature/instance)+10 noise features 说明 Ability to detect relevant features 对比方法： Feature Selection based on Gaussian mixture model(FSGMM) Laplacian score +MMC/K-Means MMC without feature selection Toy data sets Conclusion 提出了一种新的无监督的特征选择的方法Manifold-Based Maximum Margin Feature Selection (M3FS) 利用流形保持数据内在的结构 * Heng Huang副教授 Chris Ding,教授， DijunLuo，二人co-advising Computer ScienceEngineer Dep：做Machine Leani