Abstract: Aiming at  the problem that the traditional BIRCH algorithm was prone to excessive intra-cluster variance or excessive merging when dealing with indicator data with strongly correlated features and uneven distribution, we proposed an improved BIRCH algorithm  that integrated an autoencoder and a dynamic threshold strategy. Firstly, this algorithm  utilized an autoencoder for nonlinear feature mapping and dimensionality reduction, weakening the influence of inter-feature correlation on the distance metric and improving the compactness and discriminability of the data representation. Secondly, we designed a dynamic threshold strategy to adaptively adjust the clustering radius based on local sample density and cluster size, enhancing the algorithm’s adaptability to unevenly distributed data. Finally, we  constructed a clustering feature tree by using the improved feature space and adaptive threshold strategy to achieve efficient and stable hierarchical clustering, and applied to intelligent clustering analysis of multidimensional data of university teachers. Experimental results show that the improved algorithm achieves superior performance on multiple clustering evaluation metrics, significantly improving  stability and accuracy of clustering.

Key words: BIRCH algorithm, autoencoder, dynamic threshold, teacher evaluation