Realistic Benchmark Datasets for Team Formation Problem in Social Networks

Many heuristic algorithms have been proposed in the literature to solve the team formation problem. The researchers considered a project as a set of skills selected randomly from the given pool of skills. But this leads to a skewed distribution of skills in the projects with many skills having very few experts, which we term as rare skills. In this work, we create a realistic bench-mark dataset for this problem. In general, any project/task in the industry can be seen to have a good mix of popular as well as rare skills. We first conduct an empirical study of the distribution of popular skills vs rare skills in the well-known DBLP (Digital Bibliography & Library Project) data set. The distribution of popularity of skills is shown to satisfy a power law with a heavy tail, indicating the presence of a large number of skills with very few experts and a small number of highly popular skills. We build a realistic a benchmark dataset using stratified random sampling to form tasks with various distributions of popular and rare skills. The classical team formation algorithms are evaluated using this new benchmark dataset. The evaluation is done with respect to the available communication costs in the literature as well as the execution time incurred by the algorithms. It has been observed from the experiments that all the measures show lower values of communication cost for tasks having higher proportion of popular skills.