关键词: 智能交通, 深度聚类, 自编码器, 公交出行需求, 迁移学习

Abstract: To enhance the accuracy of urban public transit travel demand clustering and overcome the limitations of traditional clustering methods in capturing non-convex patterns, handling noise, and reducing parameter dependency, this paper proposes a general framework for public transit travel demand clustering supporting cross-domain applications and develops a Geographic AutoEncoder-based Hierarchical Clustering (GeoAE-HC) algorithm. The sine-cosine positional encoding is introduced and a geographical self-attention mechanism is designed to capture both feature similarity and geographical proximity among transit travel demand data. A hierarchical clustering method integrating density-based clustering and mean-based clustering is developed and applied in the latent feature space extracted by the autoencoder, enabling accurate clustering of travel demand distribution. To improve the cross-domain generalization ability of the model across different cities, a transfer learning strategy combining domain adversarial training with a frozen encoder and fine-tuned decoder is designed to strip away city-specific noise. Simulation results show that in the Chengdu dataset with 375 cluster centers, all evaluation metrics of the GeoAE-HC algorithm are superior to those of the comparison algorithms. Specifically, the Silhouette Coefficient is improved by 11.82%, 23.14%, and 45.01% compared to Deep Embedded Clustering(DEC), DK-means, and k-means, respectively. The Calinski-Harabasz (CH) Index is improved by 6.14%, 5.62%, and 14.44%, respectively. After cross-domain transfer to the Beijing dataset, the performance of GeoAE-HC also outperforms those of comparative algorithms, validating its effectiveness and generalization capability.

Key words: intelligent transportation, deep clustering, autoencoder, public transit demand, transfer learning