AoBPRRecomender execution is too inefficient

[Kiteflyingee]

I have a data set of 60,000 data, the data form is user-item, only the user-item relationship, no score, but this data set “rec.iterator.maximum=30” on the AoBPR program takes 30 minutes, I think the efficiency is too low .It's Is there any place to optimize?

[SunYatong]

Hi, if the dataset doesn't contain rating column, the converting process should throw exceptions. You'd better debug the recommender to check its trainMatrix and testMatrix.

On Mon, Nov 27, 2017 at 8:53 PM, Kiteflyingee notifications@github.com wrote:

I have a data set of 60,000 data, the data form is user-item, only the user-item relationship, no score, but this data set “rec.iterator.maximum=30” on the AoBPR program takes 30 minutes, I think the efficiency is too low .It's Is there any place to optimize?

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/guoguibing/librec/issues/220, or mute the thread https://github.com/notifications/unsubscribe-auth/AQWU7uumr6tGbcgJdsZ_mkYXKwvRRr-tks5s6rC9gaJpZM4Qrl0c .

[Kiteflyingee]

I have debug the program.I find librec populated my dataset with a rating of 1.0,this is exactly the rating I need.Is it possible that my rating matrix is ​​too sparse and has a good solution?

[SunYatong]

Yes, if the user number is too large, AoBPR could be really slow. In each epoch it samples 100*numUser positive/negative pairs.

[Kiteflyingee]

ok，I understand，if i samples 50*numUser positive/negative pairs，whether it will have a great impact on the precision of the recommendation process?

[SunYatong]

It will cost less time for each epoch. But it might need more epoch to converge. And the result could be different. You can try it yourself.

[Kiteflyingee]

Thank you.