On Caching for Local Graph Clustering Algorithms

Abstract

In recent years, local graph clustering techniques have been utilized as devices to unveil the structured hidden of large networks. With the ever growing size of the data sets generated in domains of applications as diverse as biomedicine and natural language processing, time-efficiency has become a problem of growing importance. We address the improvement of the runtime of local graph clustering algorithms by presenting the novel caching approach SGD^⋆. This strategy combines the Segmented Least Recently Used and Greedy Dual strategies. By applying different caching strategies to the unprotected and protected segments of a cache, SGD^⋆ displays a superior hitrate and can therewith significantly reduce the runtime of clustering algorithms. We evaluate our approach on four real protein-protein-interaction graphs. Our evaluation shows that SGD^⋆ achieves a considerably higher hitrate than state-of-the-art approaches. In addition, we show how by combining caching strategies with a simple data reordering approach, we can significantly improves the hitrate of state-of-the-art caching strategies.