Release 1.1.0
Update
-
Add CUDA accelerated BPR-MF
sample usage- When tested against ml-20m and Kakao Brunch Dataset, CUDA version of BPR-MF was about 5x times faster than CPU implementation of BPR-MF in training time.
-
Add metric truncated AUC
-
Update documents
Bug fixes
- Fix SGD update of BPR
- Fix to use bias when calculating top-k recommendation in BPR-MF
Misc
- Clean code(linting, refactoring)
- Add internal data type check and state check
- Change validation format to be CSR matrix
- Validation took about 40% faster than the previous version.
- but this might be harmful because it requires larger memory during validation.