Algorithms for sparsity constrained principal component analysis

The classical Principal Component Analysis problem consists of finding a linear transform that reduces the dimensionality of the original dataset while keeping most of the variation. Extra sparsity constraint sets most of the coefficients to zero which makes interpretation of the linear transform easier. We present two approaches to the sparsity constrained Principal Component Analysis. Firstly, we develop computationally cheap heuristics that can be deployed in very high-dimensional problems. Our heuristics are justified with linear algebra approximations and theoretical guarantees. Furthermore, we strengthen our algorithms by deploying the necessary conditions for the optimization model. Secondly, we use a non-convex log-sum penalty in the semidefinite space. We show a connection to the cardinality function and develop an algorithm, PCA Sparsified, to solve the problem locally via solving a sequence of convex optimization problems. We analyze the theoretical properties of this algorithm and comment on the numerical implementation. Moreover, we derive a pre-processing method that can be used with previous approaches. Finally, our findings from the numerical experiments we conducted show that our greedy algorithms scale to high dimensional problems easily while being highly competitive in many problems with state-of-art algorithms and even beating them uniformly in some cases. Additionally, we illustrate the effectiveness of PCA Sparsified on small dimensional problems in terms of variance explained. Although it is computationally very demanding, it consistently outperforms local and greedy approaches.