Figure 5.

The task similarity derived from MoTSE is generalizable across models with different architectures and datasets with different data distributions

(A-C) The comparison results between MoTSE and baseline methods on the QM9 and PCBA datasets (measured in terms of R² and AUPRC, respectively), using the graph attention network (GAT), fully connected network (FCN), and recurrent neural network (RNN) models, respectively.

(D) The comparison results between MoTSE and baseline methods on the Alchemy dataset, measured in terms of R².

(E) The similarity trees constructed based on the task similarity estimated by MoTSE on the QM9 and Alchemy datasets, respectively.