See section 7.12.2 of Newman book.

It is the number of common neighbours minus the expected number of common neighbours if edges were random ($k_i k_j /n$), normalized in a certain way. It is also the covariance between two rows of adjacency matrix divided by the product of their standard deviations: