Multi-view data depicts objects from different perspectives, with features in different views exhibiting correlations, complementary, and diverse information. Therefore, it is crucial to make full use of this information for the processing of multi-view data. However, the processing and analysis of multi-view data will be difficult due to the inherent challenges of dealing with a vast number of features and the presence of noise features in multi-view data. Unsupervised multi-view feature selection, emerging as a critical component in multi-view data learning, efficiently learns more accurate and compact representations from the original high-dimensional multi-view data without relying on label information to remarkably improve the performance of data analysis. This study reviews and categorizes these models based on the similarities and differences in the working mechanisms of existing unsupervised multi-view feature selection models, while also detailing their limitations. Furthermore, this study points out promising future research directions in the field of unsupervised feature selection.