基于无监督特征提取的局部相似度保序投影方法

Locally similarity-preserving ordinal projection method based on unsupervised feature extraction

为了对机器学习和模式识别等领域中的高维数据进行特征降维,在传统的局部保持投影方法上,提出基于无监督特征提取的局部相似度保序投影方法,使用低阶与高阶的局部关系探索数据局部结构,同时引入平衡参数调整样本低阶结构和高阶结构的重要性,利用图嵌入框架,通过特征值分解得到了最优的子空间投影矩阵。在3个公共数据集上与一些经典的图像降维算法进行对比实验,并给出该算法的参数敏感性分析。实验结果表明:在相同的实验条件下,采用局部相似度保序投影方法处理后的数据具有较好的数据判别特征,在图像分类任务中表现最佳,与一些经典的降维方法相比,该方法具有更好的降维效果。

In order to perform feature dimensionality reduction on high-dimensional data in fields such as machine learning and pattern recognition,an order-preserving projection method of local similarity is proposed on the traditional local hold projection method,which uses local relations of low and high order to explore the local structure of the data,while introducing balance parameters to adjust the importance of the low-order structure and high-order structure of the samples,and using the graph embedding framework,the optimal subspace is obtained by eigenvalue decomposition projection matrix.The experiments are compared with some classical image dimensionality reduction algorithms on three public datasets,and the parameter sensitivity analysis of the algorithms is given.The experimental results show that under the same experimental conditions,the data processed by the local similarity-preserving projection method has better data discriminative features and performs best in image classification tasks,and the method has better dimensionality reduction effect compared with some classical dimensionality reduction methods.