Cluster　analysis　plays　an　important　role　in　machine　learning.　In　cluster　analysis,　there　still　exist　many　classical　problems,　like　how　to　determine　the　number　of　clusters　and　how　to　find　arbitraryshaped　clusters.　We　focus　on　these　problems　in　this　paper.　We　propose　a　link　growing　method　by　using　the　fact　that　the　similarity　within　a　cluster　is　usually　high　than　the　similarity　in　other　clusters.　The　method　first　constructs　multiple　links,　then　merge　links　into　clusters,　finally　a　dataset　is　divided　into　multiple　clusters.　Experiments　demonstrate　the　effectiveness　of　the　proposed　algorithm.　The　proposed　algorithm　has　good　performance　on　nonspherical　datasets.　©　2019　Association　for　Computing　Machinery.