Bipartite network projection

Content sourced from Wikipedia, licensed under CC BY-SA 3.0.

Bipartite networks have two kinds of nodes, X and Y, and only cross-type links between them. To study relationships among one type of node, we often create a one-mode projection: connect two X nodes if they share at least one Y neighbor (or connect two Y nodes if they share an X neighbor). The plain, unweighted projection is simple but usually loses a lot of information about the original structure, so it’s common to assign weights to the projected links.

Weighting methods vary with the network and the goal, and the right choice helps keep as much original information as possible. The weight on a link often depends on the degrees of the nodes in the bipartite graph. For example, in a co-authorship network, how strong two authors’ connection is depends on how many papers each author wrote and how many authors there are on each paper.

Backbone algorithms for bipartite projections use this information to identify edges that are statistically stronger (or weaker) than expected. Keeping only these significant edges yields a sparse, easier-to-analyze backbone network that still captures the important structure. This approach differs from other weighted-backbone methods, such as the disparity filter.