Detailed Descriptions of Evaluation 

BTC-Alpha and BTC-OTC are trust networks within bitcoin trading platforms where users establish trust or distrust relationships. Each edge in these networks not only has a sign and timestamp but also carries a weight indicating the degree of trust or distrust, ranging from -10 to +10. 

Wiki-RfA is a voting network where Wikipedia members cast either supportive or opposing votes regarding requests for administrative privileges.

Epinions is a trust network among consumers on an online product review site, where relationships of trust or distrust formed.

The transductive and inductive settings evaluate each NE method by using the testseen and testunseen sets: the former comprise pairs of nodes that were testseen during training, while the latter includes those with at least one node that was testunseen, respectively. The hybrid setting uses a combination of the testseen and testunseen sets.

The experiments ran on NVIDIA RTX A6000 GPUs with 48GB memory and 256GB RAM, using PyTorch 2.0.1 on Ubuntu 22.04 OS. To adapt SNE methods to a dynamic setting, we arranged the learning sequence according to the order in which edges occur, following [42]. Also, for DNE methods, we incorporated edge sign as attributes to enable a signed setting. All experiments used five different seed settings, reporting the average accuracy. Embedding dimensionality was set to 64, following [14, 30, 42]. The optimal hyperparameters for the competitors were identified through a grid search with the validation set. For POLARDSN, we set 𝛼 = 64 (BTC-Alpha and Wiki-RfA), 128 (BTC-OTC), and 32 (Epinions); 𝛽 = 3 (BTC-Alpha), 5 (BTC-OTC), and 2 (Wiki-RfA and (Epinions); 𝛾 = 1e-6 (BTC-Alpha, BTC-OTC, and Epinions) and 1e-7 (Wiki-RfA). Additionally, a learning rate of 0.001 was used across all datasets.

For all methods, we employ an early stopping strategy within 100 epochs to select the best epoch to halt training. 

For every dataset across all methods, the batch size was generally set to 64.