ABSTRACT
One of the challenges in multi-label learning is how to effectively capture and utilize the correlation among different labels in classification. In this chapter, we present Hypergraph Spectral Learning (HSL), which employs a hypergraph [1] to capture the correlation among different labels for improved classification performance. A hypergraph is a generalization of the traditional graph in which the edges are arbitrary nonempty subsets of the vertex set. It has been applied for domains where higher-order relations such as co-authorship exist [1, 283]. In HSL, a hyperedge is constructed for each label and all instances annotated with a common label are included in one hyperedge. Following spectral graph embedding theory [54], HSL computes the lower-dimensional embedding through a linear transformation, which preserves the instance-label relations captured by the hypergraph. Thus, the projection is guided by the label information encoded in the hypergraph which captures the label correlations.
