Three-way clustering based on the graph of local density trend

Three-way clustering demonstrates its unique advantages in dealing with the issues of information ambiguity and unclear boundaries present in real-world datasets. The core and boundary region in the data are identified as key features of cluster analysis. Typically, data is segmented into three regions based on a set of predetermined global thresholds, a common practice in three-way clustering. However, this method, which relies on global thresholds, often overlooks the intrinsic distribution patterns within the dataset and determining these thresholds a priori can be quite challenging. In this paper, we propose a three-way clustering method based on the graph of local density trend (3W-GLDT). Specifically, the algorithm first uses a density-decreasing strategy to build subgraphs and divide the core region data. Then, the unreasonable connection is corrected by using isolated forest, which increases the number of core points and enlarges the distribution range of core points. Next, a three-way allocation strategy is proposed, which fully considers the degree of local aggregation of subgraphs and the natural domain information of each data object to ensure the correct allocation. Finally, the proposed algorithm is compared with 8 different clustering methods on 8 synthetic datasets and 10 UCI real datasets. The experimental results show that the 3W-GLDT algorithm has good performance and clustering results.