Frequency measures global feature importance in a tree ensemble by counting how many split operations use each feature across all trees. In XGBoost, it helps identify which molecular descriptors contribute most to distinguishing RNA-binding from protein-binding activity. section 6.4.2
Unlike gain, which averages split improvement, or coverage, which averages affected samples, frequency counts usage. These measures summarize average behavior across the dataset rather than explaining one compound’s prediction; local interpretation is needed for hit-to-lead optimization. section 6.4.2
Figure 6.14 Feature importance of our XGBoost model across gain, coverage, and frequency metrics.

Compare frequency with gain and coverage in the XGBoost feature-importance visualization. section 6.4.2








