Permittivity determines the practical applications of microwave dielectric materials, so accurate prediction of permittivity is an important task of the materials design on demand. Both classical dielectric theory and recent machine learning studies indicate that the ion dielectric polarizability ( D) is the decisive parameter for permittivity prediction. However, the most influential D database, put forward by Shannon in 1993, is facing three challenges in the 5G era: (1) Few data, (2) Simplistic relation and (3) Low frequency (kHz ~ MHz) oriented. Therefore, it cannot satisfy the accurate permittivity prediction of novel microwave dielectric materials for GHz band applications. Machine learning methods enable quantitative, accurate, efficient and low-cost prediction of material properties, and also shows advantages in mining hidden physical relationship. Thus, it opens up a new path for optimizing and extending the D database.A team led by Prof. Zhifu Liu from Shanghai Institute of Ceramics, Chinese Academy of Sciences, optimized and extended the ion dielectric polarizability database by multiple linear regression and support vector machine methods, achieving improved accuracy of D data, and greatly expanding the data amount from 61 to 915, covering 92 elements.In this work, based on the permittivity data of 334 microwave dielectric ceramics reported in literatures, the current D data was optimized according to the error degree by four-stage multiple linear regression method. Based on the optimized database, several machine learning models were used to try to establish the quantitative relationship between the D value and basic characteristics of ion. The optimal model, based on support vector machine algorithm and including atomic number (N), valence state (V), coordination number (CN) and ion radius (IR), was used to extend the D database. Combining dielectric theory and SHAP method, this work discussed the mechanism of basic characteristics of ion affecting the D value, and also found that the neglected importance of V and CN on the D value in previous studies. The previously found positive linear relation between the D value and the cube of ionic radius is valid for the ions with the same V and CN, meaning the change of IR caused by N. However, the law fails when the change of IR caused by CN, for ions with the same V and N. In addition, the D databases were integrated and publicly available online (https://qincas.gitee.io/idp-ml/), which serves as a reference for the research and development of microwave dielectric materials.