Spectral techniques for measuring bipartivity and producing partitions

Complex networks can often exhibit a high degree of bipartivity. There are many well-known ways for testing this, and in this article, we give a systematic analysis of characterizations based on the spectra of the adjacency matrix and various graph Laplacians. We show that measures based on these characterizations can be drastically different results and leads us to distinguish between local and global loss of bipartivity. We test several methods for finding approximate bipartitions based on analysing eigenvectors and show that several alternatives seem to work well (and can work better than more complex methods) when augmented with local improvement.