<p>Software defect prediction (SDP) is a technique that uses known software features and defect information to predict target software defects. It helps reduce software development costs, save development time, and improve software quality. The performance of SDP has always been the goal of this field. Many researchers have found that software features and their processing have a significant impact on the results of SDP. However, few reviews have summarized the features types, use, utility, processing and effect. So, we searched the six databases for SDP empirical papers. One hundred and forty-eight researches were obtained by using the inclusion and exclusion strategies. The information of feature types, processing methods, prediction schemes, modeling techniques, data sets, and performance evaluation metrics in these papers was collected and analyzed. We find that the rise of semantic and network features has broken the situation where only software metrics were used for SDP. These features demonstrate significant potential in prediction, especially semantic features. Deep learning is used increasingly in SDP, and deep learning feature extraction replaces the original feature selection based on machine learning. Although SDP performance continues to improve, issues such as data quality, feature stability, and scenes selection still deserve future research attention.</p>