Do second-order similarities provide added-value in a hybrid approach?

Recent studies on first- and second-order similarities have shown that the latter one outperforms the first one as input for document clustering or partitioning applications. First-order similarities based on bibliographic coupling or on lexical approaches come with specific methodological issues like sparse matrices, sensitive to spelling variances or context differences. Second-order similarities were proposed to tackle these problems and take the lexical context into account. But also a hybrid combination of both types of similarities proved an important improvement which integrates the strengths of the two approaches and diminishes their weaknesses. In this paper we extend the notion of second-order similarity by applying it in the context of the hybrid approach. We conclude that there is no added value for the clearly defined clusters but that the second-order similarity can provide an additional viewpoint for the more general clusters.