A K-means Clustering Method for Fuzzy Neighborhood Granules

K-means algorithm is a typical distance-based clustering analysis algorithm, and the neighborhood granular K-means clustering algorithm uses the neighborhood granulation technique on its basis to solve the problem of poor clustering of the K-means clustering algorithm on non-linear separable datasets. However, the use of neighborhood relations in the granulation lead to the inability of the algorithm to deal with outliers effectively, resulting in some samples being misclassified. By introducing the concept of fuzzy set and combining granular computing theory, a new neighborhood granular K-means clustering method is proposed, where samples are constructed using fuzzy neighborhood granularity technique to construct fuzzy neighborhood granules, and the membership between samples is used for neighborhood granularity instead of clear neighborhood relations, which reduces the impact of outliers. A new distance measure for fuzzy neighborhood granules is proposed based on the size and form of the fuzzy neighborhood granules, which reduces the computational complexity. In addition, the algorithm combines the idea of K Nearest Neighbor to define a neighborhood threshold metric that adaptively generates neighborhood thresholds for any data set. It considers the size of the data set objectively and reduces the subjective influence of manual selection of neighborhood threshold. Finally, the Accuracy and Fowlkes Mallows Index values of the algorithm are increased by 6.4 %, 9.21%, 3.5%, and 7.89% on the synthetic data set and UCI data set, respectively, and increased by 8.65% and 9.43% on the seismic facies lithology identification data set, respectively. The experimental results prove the feasibility and effectiveness of the fuzzy neighborhood granular K-means clustering method.