Sparse Ordinal Regression via Factorization Machines

Most existing ordinal regression methods are adapted from traditional supervised learning algorithms (e.g., support vector machines and neural networks) which have shown to work well mostly on dense data. However, the use of existing ordinal regression methods on sparse data has received less scrutiny. This paper proposes to address the sparsity issue arose in many real-world ordinal regression applications by leveraging the feature interaction modeling techniques. Following the popular threshold methodology in ordinal regression studies, we extend Factorization Machines, an effective solution to modeling pairwise feature interactions in sparse feature space, to ordinal regression. The proposed model, namely Factorization Machines for Ordinal Regression (FMOR), combines the ability of threshold methodology in predicting targets of ordinal scale with the advantages of factorization models in handling high-dimensional sparse data. Through extensive experimental studies, we show that the proposed FMOR is both effective and efficient against state-of-the-art baselines.