Feature selection plays a significant role in the field of data mining and machine learning to reduce the data dimension, speed up the model building process and improve algorithm performance. Tree growth algorithm (TGA) is a recent proposed populationbased metaheuristic, which shows great power of search ability in solving optimization of continuous problems. However, TGA cannot be directly applied to feature selection problems. Also, we find that its efficiency still leave room for improvement. To tackle this problem, in this study, a novel improved TGA (iTGA) is proposed, which can resolve the feature selection problem efficiently. The main contribution includes, (1) a binary TGA is proposed to tackle the feature selection problems, (2) a linearly increasing parameter tuning mechanism is proposed to tune the parameter in TGA, (3) the evolutionary population dynamics (EPD) strategy is applied to improve the exploration and exploitation capabilities of TGA, (4) the efficiency of iTGA is evaluated on fifteen UCI benchmark datasets, the comprehensive results indicate that iTGA can resolve feature selection problems efficiently. Furthermore, the results of comparative experiments also verify the superiority of iTGA compared with other stateof-the-art methods.