I want to evaluate some algorithms for the directed graph classification task. Therefore, I'm looking for directed graphs data sets (preferably without node or edge features) as benchmarks. Do you have any ideas for datasets that could fit?
3
-
1I think it would be much easier if you could first identify some applications of this type of classification, and then look for the datasets corresponding to these applications.– ErwanCommented Jun 14, 2021 at 21:50
-
Have you gone to the literature on directed graph classification? What are researchers using?– sboyselCommented Jun 21, 2021 at 15:25
-
Mostly datasets for clustering– OrFeCommented Jun 23, 2021 at 13:35
Add a comment
|
2 Answers
The Stanford Network Analysis Platform (SNAP) share a variety of widely used graph datasets. There are datasets for Large Networks and Biomedical Networks. You can see that there are many that are directed and most do not include features (if I remember correctly).
-
1Correct me if I'm wrong, but all the graphs there for the graph classification task are undirected.– OrFeCommented Jun 20, 2021 at 7:51
-
You are right, my mistake. I will leave the answer in case others are interested in adapting the other directed graph datasets for a classification task.– sboyselCommented Jun 21, 2021 at 15:25
In Chapter 11 - Transitivity, structural, balance, and hierarchy, of the free online book Methods for Network Analysis, there is a picture how triads of nodes are be classified. Scroll down to section "11.4 - Calculating a triad census" to see the figure and the names that I mean.
(the first part, the number before the "-" is a counter.)
Obviously, this classification is different from a machine-learning classification (=prediction) task.
However, if ML classification is your use-case, then I have no dataset examples. But the book in the same chapter, but section "11.5 - Random graphs galore!" also shows how you can create your own synthetic directed graphs with R and igraph.
Then you can control and know the size and the properties of the graph in advance. Perfect for prediction tasks and algorithm evaluation!
