This research describes an advanced workflow of an object-based geochemical graph learning approach, termed OGE, which includes five key steps: (1) conduct the mean removal operation on the multi-elemental geochemical data and then normalize them; (2) data gridding and multiresolution segmentation; (3) calculate the Moran’s I value and construct the geochemical topology graph; (4) unsupervised deep graph learning; (5) the within-object statistical analysis. The final product of OGE is an object-based anomaly score map. The performance of OGE was demonstrated by a case study involving eighteen ore-forming elements (Cu, Pb, Zn, W, Sn, Mo, F, Au, Fe₂O₃, etc.) in stream sediment samples in the Bayantala-Mingantu district, North China. The results showed that the OGE analysis performed at lower levels of scale greatly improved the quality of anomaly recognition: more than 80% of the known ore spots, no matter what their scales and mineral species, were predicted in less than 45% of the study area, and most of the ore spots falling outside the delineated anomalous regions occur nearby them. OGE can extract both the spatial features and compositional relationships of geochemical variables collected at irregularly distributed centroids in irregularly shaped image objects, and it outperforms other convolutional autoencoder models such as GAUGE in anomaly detection.