Visualizing Bipartite Network Data using SAS® Visualization Tools

Bipartite networks refer to the networks created from affiliation relationships, such as patient-provider relationships in healthcare data. Analyzing such networks allows us to gain additional insights on groups that share the same member and members that belong to the same group. This paper develops methods to visualize bipartite network data in healthcare using the Annotate facility. Link analysis is used to explore the connections between healthcare providers and patients, and hence connections between peer providers. In addition, this paper also takes advantage of the spatial information in the patient-provider data to gain additional insights. In particular, the connections between providers are projected on a Google map using Google Map Generator. INTRODUCTION Prior work has been published on using SAS to visualize and to compute network statistics of classical social networks (Hoyle 2006, Hornibrook 2009, Ellis 2009). This paper focuses on the visualization of bipartite networks, which are the networks that consist of two disjoint sets of nodes and all edges must be between the two sets. Bipartite networks are often used to model affiliation relationships, such as the patient-provider network or the consumer-seller network. In this paper, special attention is paid to uncovering the hidden connections in patientprovider bipartite networks. Instead of directly applying social network analysis and visualization techniques on the original bipartite network, we first convert the original bipartite network into a classical (unipartite) network of providers that is defined by patient-sharing. Once the network of providers is established, of interests are those unnaturally dense cliques within the network. The clustered providers are displayed distinctively and suspicious patterns flagged for further study. Large networks posts one of the greatest challenges in visualizing social networks. In the patient-provider network, for example, there are over 13,000 health care providers registered under NPI (National Provider Identification) in the tri state area of Maryland, Virginia and DC, and 230,000+ FFS Medicare Part B patients in calendar year 2009. When dealing with the big data set such as this, the first visualization challenge comes from laying out the nodes. Multidimensional scaling (MDS) is often used to produce spatial layout of the nodes based on their “similarity.” However, MDS does not scale well with increasing number of nodes. So far one method has been used to overcome the difficulties in laying out a very large network with MDS. The first method (Hoyle 2009) uses a variation of MDS that was used to generate coordinates for the node dataset. When spatial information (such as address or zip code) is available, one can also skip the MDS entirely by projecting all nodes on a Google map using Google Map Generator provided by SAS/GRAPH. This paper discusses these two methods as well as a third method, component analysis, to dissect the original network into many smaller, manageable sub-networks. THE REPRESENTATION OF SOCIAL NETWORK DATA IN SAS Social network data can be represented in one of two formats in SAS, the matrix format and the edge-list format. The matrix format stores social network data in an n-by-n square matrix with all nodes listed across the top and down the side. The value of cell(i,j) in the matrix indicates whether the node i and the node j is connected (0 – no connection; Posters NESUG 2010