neural-collaborative-filtering icon indicating copy to clipboard operation
neural-collaborative-filtering copied to clipboard

Published 20 hours ago • verified publisher icon yihong-chen

the performance of ncf in hr@10 didn't achieve 0.7

Open LiuHao-THU opened this issue 5 years ago • 4 comments

the results of hr@10 using ncf model is blow 0.7,however in paper "neural collaborative filtering" is over 0.72 using factor 32

LiuHao-THU avatar Jan 09 '20 07:01 LiuHao-THU

Thank you for reporting the issue. Could you add the exact hyperparameters settings for this ? NCF is quite sensitive to hyperparameters. I didn't spend a lot of time tuning them. I was suspecting all the weight decay(regularization) and finetuning. Would you like to help with this and report the results here?

yihong-chen avatar Jan 09 '20 15:01 yihong-chen

here is a link for NCF BPR and NFM https://github.com/guoyang9/NCF https://github.com/guoyang9/BPR-pytorch https://github.com/guoyang9/NFM-pyorch the scores are exactly the same even higher than origin papers!

LiuHao-THU avatar Jan 09 '20 15:01 LiuHao-THU

Nice points! I checked the repos and found I didn't do the weight initialization tuning. Maybe that's the reason. You could use those repos if you want. I will fix the issue later according to my schedule.

yihong-chen avatar Jan 09 '20 17:01 yihong-chen

@LiuHao-THU Could you elaborate a bit on factor ? Is it latent factor or predictive factor?

yihong-chen avatar Jan 12 '20 23:01 yihong-chen