Research On Discriminative Feature Learning Based On Multidimensional Scaling

As an important data analysis method in machine learning and data mining,feature learning has been widely used in various fields.Feature learning aims to automatically learn new data representations from the original data,making it easier for subsequent learners to uncover the required information from the new data representations and thus improving the performance of subsequent learners.Since discriminative features tend to further improve the performance of subsequent machine learning tasks,discriminative feature learning has important research implications.Aiming at the problem that traditional multidimensional scaling method ignores the discriminability of the learned low dimensional embedding,this thesis first proposes a discriminative feature learning method based on the multidimensional scaling method,which is referred to as Discriminative Multidimensional Scaling for Feature Learning(DMDS).DMDS not only maintains the local topological information in the original data points.Meanwhile,it also automatically discovers the structure of clusters in the samples by fuzzy k-means,so that the new data representation corresponding to the same cluster is close to the cluster center during the feature learning process,which increases the compactness of the new data representation in the same cluster,thus improving the discriminability of the learned data representation.To further enhance the discriminability of the learned data representation,a discriminative multidimensional scaling model based on pairwise constraints is also proposed in this thesis,which is referred to as Discriminative Multidimensional Scaling Based on Pairwise Constraints for Feature Learning(pc DMDS).Given the pairwise constraint information between some samples,the extended pairwise constraint information between more samples is firstly obtained via the constraint propagation algorithm.Then,in the feature learning process,both the topology of the samples in the original space and the cluster structure in the new space are considered,and the extended pairwise constraint information in the samples is also incorporated to further improve the ability of the proposed model to obtain discriminative data representation.Extensive comparative experiments are conducted on the two feature learning models proposed in this thesis on public datasets.The experimental results show that the proposed model not only learns more discriminative features than the traditional multidimensional scaling method,but also has a greater advantage over other feature learning algorithms.