关键词: 谱多流形聚类, 子空间聚类, 聚类参数调节, 比例-积分-微分（PID）

Abstract: The complexity and diversity of data demand more in big data processing and analysis. Manifold clustering techniques have shown remarkable success in numerous problems in data mining. Parameter adjustment is one of the difficulties in the research of clustering algorithm, which directly affects clustering performance. Traditional parameter adjustment methods of clustering algorithm depend on the experience, or the blindness and randomness of the parameter adjustment make the algorithm invalid or more complex. In this paper, a novel method to control the adjustment of spectral multi-manifold clustering parameters actively, based on proportional-integral-derivative (PID) constraints, is proposed. By constructing the similarity matrix, multiple principal component analyzers are used to estimate the local tangent space. The model approximation process is controlled by parameter transfer and PID adjustment. In PID adjustment, the three-dimensional ZN method is used to adjust model parameters and to extend the search space, so that the clustering process is controlled with feedback results, and thus the accuracy and complexity of the clustering algorithm are improved. Better clustering performance can be obtained by detecting different types of data feature sets in synthetic data and real data.

Key words: spectral multi-manifold clustering, subspace clustering, clustering parameter adjustment, proportional-integral-derivative (PID)