This paper deals with two key issuses of the artificial intelligence （AI） -based transient stability assess- ment （TSA） of power system, namely the selection of kernel input features and the stability-related evaluation mo- del. In the investigation, first, data-driven feature selection method and rule extraction algorithm are proposed. Then, the key features are evaluated and the transient stability rules are made from the training samples. During the feature selection, a genetic algorithm-based k-nearest neighbor （GA-knn） is used to assess the input features. Du- ring the rule extraction, a mining algorithm of classification and association rules is followed to form the rules of transient stability assessment. The proposed method is then applied to both the New England 10-machine 39-bus and the 3-machine 9-bus systems, and the results are compared and analyzed. It is found out that the selected ker- nel features from 53 candidates and the obtained rules are adapted for the two test power systems. However in the stability boundary, evaluation rules are complex and specific.