Abstract:With the deep development and popularization of Internet, new data types emerge in new application fields so that many classic clustering algorithms are no longer effectively adapted to new situations, so data mining is becoming thorny issues and research focus. Therefore the article proposes a novel clustering algorithm based on self-optimizing the centers and boundaries of classes. The algorithm contains the points' distance-radius-distribution matrix-R and the cumulative radius-distribution matrix-ΣR characterizing the degree of data aggregation. The data points with the minimum R and ΣR as the class centers are searched under the breadth-first. The algorithm also includes the partial derivative matrix-R' of the distance-radius distribution to describe the gradient change of the loose degree between different points. According to self-optimizing and breadth-first, the transition point of matrix-R', which its partial derivative is the biggest one in adjacent points, is found as the class boundary, inside which all points belong to the class. After emulating and testing the algorithm by typical clustering data sets of Aggregation, the result shows that the algorithm can effectively cluster the data sets with different shapes, sizes and different densities, identify the isolated points and noises, and also have better robustness and accuracy.