Research On Classification Algorithm Based On Manifold Learning

In recent years,as the amount of data has grown exponentially,we have lived in a world with a large amount of high-dimensional complex data.Manifold learning is a framework for recovering low-dimensional manifold structures from high-dimensional data.Facing high-dimensional complex data manifold learning has gradually become one of the core technologies in face recognition and spectral clustering.At present,most spectral clustering(SC)algorithms divide data by calculating the Laplacian eigenvalues of data,such as PCA and Isomap.Such a traditional dimensionality reduction algorithm can only process data with a single manifold structure,but when the data is at or near multiple smooth low-dimensional manifolds,it is difficult to describe the true shape with the intersection surface and cause incorrect clustering.Extreme learning machine(ELM)has fast training speed and high classification rate.However,due to the high dimensionality of data in practical problems,the traditional dimension reduction method for data ignores the manifold information and small data between complex data.Sample problems can cause the ELM to over-fitting in the case of limited training samples.We have studied the above mentioned problems,the main research results of this paper are as follows:1)A new manifold clustering using tilt path algorithm is proposed.The algorithm randomly selects several landmark points and then checks whether there is a path bound by the curvature between each landmark point and other points.It is also proved that the lack of curvature limitation for shortest path algorithm can cause the path distance between the points on different surfaces,which solves the clustering problem of cross manifold topology.The algorithm is successfully applied to artificial data sets and real data.And compared with other clustering algorithms,it has achieved good clustering accuracy.2)A Force-Locality Preserving Projections(FLPP)algorithm is proposed to maintain the global and local geometry of samples,and the inter-class dispersion matrix is added to the manifold information,thus avoiding the troubles of sample point overlap,and small sample problems.Experimental results show the proposed algorithm can effectively improve the generalization ability and classification accuracy of the extreme learning machine(ELM).