Network Classification and Inferencing
Arts and Sciences
0000-0002-6943-7310, 0000-0002-9523-0057, 0000-0002-1106-2704
Currently, there is no definitive method for classifying networks into distinct categories. The leading method in network classification involves using Support Vector Machines (SVM) to identify subgroups within a broader category, often a specific field of investigation. By looking at data mining classification methods, and feature analysis, this work seeks to classify networks into meta-categories with high accuracy.
Preliminary investigations have been conducted with the Network Repository data from BHOSLIB, DIMACS, DIMACS10, Retweet Networks, Social Networks, and Temporal Reachability networks. SVM has revealed promising results with a classification accuracy of 67.5%. This was achieved by excluding Total Triangles. In addition, Naive Bayes has shown good results with the exclusion of the attribute maximum triangles. It has produced a classification accuracy of 84.3%.
Identifying the best algorithm and the best features to consider will lead to a more procedural and efficient way of classifying graphs into these meta-categories. This will be useful to the wider scientific community by allowing them to more easily choose effective algorithms for graph mining and investigations.
Nowak-Wolff, Sammantha K.; Knapp, AnnaLee; and Morris, Charles Jr., "Network Classification and Inferencing" (2016). Symposium on Undergraduate Research and Creative Expression (SOURCE). 583.