Published:
TL;DR
This project explored using Semantic IDs (SIDs) with the HSTU model on MovieLens-1M and -20M. I show that simply replacing raw IDs with SIDs reduced model performance, but combining them showed hints of faster convergence. Analysis suggests SIDs may have limited values where collaborative behavioral signals dominate, but could remain promising for large-scale and content-driven domains.
Published:
TL;DR
This project explored using Semantic IDs (SIDs) with the HSTU model on MovieLens-1M and -20M. I show that simply replacing raw IDs with SIDs reduced model performance, but combining them showed hints of faster convergence. Analysis suggests SIDs may have limited values where collaborative behavioral signals dominate, but could remain promising for large-scale and content-driven domains.
Published:
TL;DR
This project explored using Semantic IDs (SIDs) with the HSTU model on MovieLens-1M and -20M. I show that simply replacing raw IDs with SIDs reduced model performance, but combining them showed hints of faster convergence. Analysis suggests SIDs may have limited values where collaborative behavioral signals dominate, but could remain promising for large-scale and content-driven domains.