Benchmark of ARI and NMI with Leiden and GraphST on DLPFC.

[linjing-lab]

The code implementation of stMMC looks like similar to DeepST and GraphST, the former uses fully connected layers as the encoder, while the latter uses Bilinear to process mutual information and uses permutation function. The benchmark of ARI (shows in Fig. 3 on paper) and NMI (shows in Fig. 4 on paper) with Leiden and GraphST causes the differences, Leiden did not consider multimodal information as one of the reasons for low ARI, and what other factors contribute to GraphST being slightly lower than stMMC. DeepST and GraphST both choose deep graph network for model architecture, Is adding multimodal information the only reason why stMMC performs better on the DLPFC dataset.