<p>Multi-dimensional data exploration is a classic research topic in visualization. Unlike most existing studies targeting record patterns, we propose a new recordset-related pattern, called subset community. A subset community typically consists of a few tightly related subsets that refer to meaningful real-world objects, thus revealing rich data insights unavailable for record-level analysis. We present a graph-based approach to explore subset communities. The key is to organize subsets extracted from a dataset into a graph according to their common records. We notice the weak community sense in subset graphs, and thus designed an algorithm to improve the pattern significance by decomposing a subset graph into connected components. We integrate the algorithm into a visualization system, achieving flexible subset creation and pattern exploration. Cases on real-world datasets, quantitative comparisons of communities with and without the decomposition, and feedback from the domain experts prove the effectiveness and usability of the approach. </p> Graphical Abstract <p></p>