Optimal training cadence for the HGNN and Two-Tower components in 2T-HGNN

Determine the optimal training cadence for the Heterogeneous Graph Neural Network and the Two-Tower components in 2T-HGNN (e.g., training the HGNN weekly while training the Two-Tower model daily), and quantify how different cadences impact recommendation performance and computational cost.

Background

The authors emphasize the modularity of 2T-HGNN, which permits training the HGNN and Two-Tower components at different frequencies to balance freshness of user representations against training cost.

They provide an example (HGNN weekly, 2T daily) but explicitly leave the systematic exploration of cadences and their performance implications for future investigation.