Performance test on two graph mining systems: Fractal and Arabesque.
We usedReddit Hyperlink Social Networkdata graph in our evaluation. The hyperlink networkrepresents the directed connections between two subreddits (a subreddit is a community on Reddit). Eachhyperlink is annotated with three properties: the timestamp, the sentiment of the source community posttowards the target community post, and the text property vector of the source post. The network is directed,signed, temporal, and attributed
We evaluate the performance test on two graph mining systems: Fractal and Arabesque. Performancecomparison is based on the most popular graph mining problems:
- Motif Counting. A motif is any connected, unlabeled graph pattern. The problem involves countingthe occurrences of all motifs in G up to a certain size.
- Frequent Subgraph Mining (FSM). The problem involves listing all labeled patterns with k edgesthat are frequent in G (i.e., frequency of their matches in G exceed a threshold). The frequency of apattern (also called support) is measured in a variety of ways, but most systems choose the minimumnode image (MNI) support measure since it can be computed efficiently.
- Clique Counting. K-clique is a fully-connected graph with k vertices. The problem involves countingthe number of k-cliques in G. Variations of this problem include counting pseudo-cliques, i.e., patterns5 whose edges exceed some density threshold; maximal cliques, i.e., cliques that are not contained in anyother clique; and, frequent cliques, i.e., cliques that are frequent (exceeding a frequency threshold).
We evaluated Fractal and Arabesque on a wide array of applications: counting motifs with 3 and 4 vertices,counting k-cliques, for k ranging from 3 to 5; FSM with patterns of 3 edges on labeled datasets using varioussupports.3.3 LaunchWe have prepared docker images and scripts to run performance tests, available on github repository.
git clone https://github.com/jplewa/SUU-AGH.git
docker-compose up --build fractal
docker-compose up --build arabesque