Minimum Spectral Connectivity Projection Pursuit for Unsupervised Classification

We study the problem of determining the optimal univariate subspace for maximising the separability of a binary partition of unlabeled data, as measured by spectral graph theory. This is achieved by finding projections which minimise the second eigenvalue of the Laplacian matrices of the projected data, which corresponds to a non-convex, non-smooth optimisation problem. We show that the optimal projection based on spectral connectivity converges to the vector normal to the maximum margin hyperplane through the data, as the scaling parameter is reduced to zero. This establishes a connection between connectivity as measured by spectral graph theory and maximal Euclidean separation. It also allows us to apply our methodology to the problem of finding large margin linear separators. The computational cost associated with each eigen-problem is quadratic in the number of data. To mitigate this problem, we propose an approximation method using microclusters with provable approximation error bounds. We evaluate the performance of the proposed method on simulated and publicly available data sets and find that it compares favourably with existing methods for projection pursuit and dimension reduction for unsupervised data partitioning.