Extended Data Fig. 1: Exploration of knowledge abundance in ElementKG.
From: Knowledge graph-enhanced molecular contrastive learning with functional prompt
a, Performance of KANO with different ElementKG components. Green denotes the removal of class hierarchy from ElementKG, which removes various classes (except for the lowest-level classes directly connected with entities), as well as axioms rdfs:subClassOf and owl:disjointWith. It consists only of entities, lowest-level classes, data properties, and object properties. Purple denotes the deletion of data properties of each entity. Yellow represents the removal of the entire functional group component, including class hierarchy and entities of functional groups, and their relations with element entities. Red indicates the complete ElementKG with all components. The results are reported as mean values +/- SD on three independent runs. The error bars represent the SD, while the dots represent three individual data points. b, Performance of KANO with different keeping rates of data properties in ElementKG. We vary the proportion of data properties of element entities retained in ElementKG and report the corresponding performance trends across datasets in various domains, represented by different colors. The horizontal axis represents the keeping rate, which refers to the proportion of knowledge introduced. The vertical axis represents the performance measured by ROC-AUC on classification tasks (higher is better) and RMSE and MAE on regression tasks (lower is better). The results are reported as mean values +/- SD on three independent runs. The mean is represented by the lines, the SD is depicted by the error bars, and individual data points are marked with dots.
