2017-3学习心得

无监督特征选择被认为是一个更难的问题，因为缺少引导相关信息搜索的类标记。

方差和可分离性

Variance score

暂且译为方差得分
可能是最简单的半监督特征评估

Laplacian Score

拉普拉斯得分
属于光谱特征选择家族
用于无监督特征选择

$L_r=\frac{\mathop{\mathring{a}}\limits_{i,j}(f_{ri}-f_{rj})^2S_{ij}}{\mathop{\mathring{a}}\limits_{i}(f_{ri}-m_r)^2D_{ii}}$

$C_r=\frac{\mathop{\mathring{a}}\limits_{(x_i,x_j)\hat{I}W_{ML}}(f_{ri}-f_{rj})^2}{\mathop{\mathring{a}}\limits_{(x_i,x_j)\hat{I}W_{CL}}(f_{ri}-f_{rj}^i)^2}$

$\mathop{\mathring{a}}\limits_{i,j}$

$\mathop{\arg\min}\limits_{v \in V}$

$\hat{I}$

SPEC method

CSFS framework

Constraint Selection

aforementioned 前述的
nevertheless 虽然
depends merely on 仅仅依赖于
derogate 减损
coherent 相关
informativeness and coherence 信息量和连贯性
vicinity 附近
exert 施加
projected overlap 投影重叠
generalize 概括、归纳
joint 关联
dissimilarity 差异
Diagonal matrices 对角矩阵
algebraic steps 代数步骤

$Coh_d(W)=\frac{\mathop{\mathring{a}}\limits_{m\hat{I}W_{ML},c\hat{I}W_{CL}}d(over_cm=0\grave{U}over_mc=0)}{|W_{ML}||W_{CL}|}$

$\mathop{\mathring{a}}\limits_{m\hat{I}W_{ML},c\hat{I}W_{CL}}d(over_cm=0\grave{U}over_mc=0)$

$\lvert W_{ML}\lvert$ $|W_{ML}||W_{CL}|$

SOM 自组织映射

Adaptive k-Neighborhood determination

低误差和更多保留局部的
Adaptive k-nearest-neighborhood 自适应k近邻用于聚类

Result

验证框架
Ascendant Hierarchical Clustering 上升的分层聚类
白血病、结肠癌数据集
和Laplacian, ReliefF,F2+r 4 and F3+r scores的比较

Results on Leukemia and Colon Cancer Data Sets

在白血病和结肠癌数据集的结果

Results on face images data sets

CONCLUSION

we proposed a framework for feature selection based on constraint selection for semi-supervised dimensionality reduction.

Future work may include the amelioration of the choice of labels set from which constraints are generated. Other perspectives may be the choice of clustering algorithm between the constraints.