Novel statistical learning methods to model environmental exposures on human health
Conference
66th ISI World Statistics Congress
Format: WSC 2027: IPS Abstract
Keywords: graph, lasso, svm
Abstract
Regularized support vector machines (SVMs) are widely used for classification and feature selection in high-dimensional data. However, conventional SVM-based methods often fail to exploit the underlying structure among features, particularly when variables are highly correlated or belong to multiple groups. To address this limitation, we propose a novel Sparse Overlapping Graph Support Vector Machine (SOG-SVM) that incorporates graph-structured information to improve both feature selection and classification performance. The proposed method accommodates overlapping feature groups and promotes sparsity in the latent feature space rather than directly on the original features. To facilitate efficient optimization, we adopt the Huberized hinge loss, which provides a smooth objective function and improves the convergence properties of the optimization algorithm. The performance of the proposed SOG-SVM is evaluated through extensive simulation studies and real-world data analyses. Results demonstrate that SOG-SVM achieves competitive classification performance while providing more effective feature selection than existing methods. We further apply the proposed model to data from the Canadian Health Measures Survey (CHMS) Cycles 1–6 (2007–2019), where it successfully predicts health outcomes from chemical exposure profiles and demonstrates strong practical performance.