With the widespread adoption of IoT, IoT device fingerprinting is crucial for strengthening security and management in IoT ecosystems. Existing approaches often rely on predefined network protocols or simple correlation, lacking machine learning-based feature selection, which may overlook complex feature relationships and dependencies. Additionally, training and testing times are rarely reported, making it difficult to evaluate and compare different fingerprinting approaches. Furthermore, implementations are typically conducted on a single dataset composed of the same network of IoT devices, limiting generalization evaluation. This study aims to demonstrate the effectiveness of machine learning-based approaches, accompanied by feature selection techniques, in IoT device fingerprinting. It seeks to evaluate multiple fingerprinting approaches with a focus on performance and computational efficiency across various datasets. The proposed framework TreePrint employs various ensemble classifiers implemented on two public datasets. Three feature selection methods: Recursive Feature Elimination (RFE), Mutual Information (MI), and Principal Component Analysis (PCA) are employed. The MI-selected feature set significantly reduced training time by 38.1% without compromising accuracy. RFE improved test accuracy from 88.88% to 90.12%. These results underscore the effectiveness of machine learning-based feature selection in enhancing both performance and computational efficiency for IoT fingerprinting. Furthermore, TreePrint demonstrates strong generalization, consistently exceeding 90% accuracy on two datasets.