HSTU did not achieve better results than sasrec on billion-scale industrial datasets

[ZhenYangIACAS]

Hi, authors:
Good jobs. HSTU usually achieves much better results than sasrec on the publicly available datasets which have hundreds of thousands samples. However, when we test HSTU on our billion-scale industrial datasets, it achieves poorer or same results on sasrec. Could you please give some suggestions?

[jiaqizhai]

Thanks for your interest in our work. There are many factors that could affect the effectiveness of an architecture. For instance, HSTU's performance is significantly improved vs vanilla Transformers in one-pass streaming setting (Table 2), but you may not need the one-pass streaming setting when the dataset is small (our internal dataset has hundreds of billion examples over a few days as discussed in Section 4.1.2) or you happen to deal with a relatively static vocabulary (eg POI recommendations or product recommendations). Detailed model preprocessing/postprocessing steps (esp in the ranking setting), including supervision, will likely also have large impacts on the results.

For similar questions, as long as people are open to disclosing their full dataset characteristics and training code/setups we'd be happy to take a look.

[ZhenYangIACAS]

@jiaqizhai Thanks for your comments. As you mentioned, could you please give some detailed suggestions about the preprocessing/postprocessing steps you find very important for the final performance?