Rankfm Versions Save

Added

• working PyPI and GitHub pip installs on both OSX and Linux
• wrapped the external Mersenne Twister C library to generate better random numbers for BPR/WARP training
• added a MANIFEST.in to include all C source and headers in the sdist archive

Changed

• changed the logic in setup.py to favor building extensions from the generated C source rather than re-cythonizing the .pyx files. This is best practice according to the Cython docs.
• removed Cython as a formal dependency as the generated C code will be included in the package sdist from now on.

Changed

• needed to instruct Python to compile the created .c file instead of the .pyx file as the latter doesn't get added to the sdist
• build tested and working now on both OSX and Linux

no changes, just syncing things up.

Added

• Cython back-end for _fit(), _predict(), _recommend() - the Cython _fit() function is 5X-10X faster than the original Numba version, and predict()/recommend() are about the same speed.

Changed

• split regularization into two parameters: alpha to control the L2 regularization for user/item indicators and beta to control the regularization for user-features/item-features. In testing user-features/item-features tended to have exploding gradients/overwhelm utility scores unless more strongly regularized, especially with fairly dense side features. Typically beta should be set fairly high (e.g. 0.1) to avoid numerical instability.

Changed

• pull the string loss param out of the private Numba internals and into the public fit() function
• change _init_interactions to extend rather than replace the user_items dictionary item sets
• added conditional logic to skip expensive user-feature/item-feature dot products if user and/or item features were not provided in the call to fit(). This reduces training time by over 50% if just using the base interaction matrix (no additional user/item features).

Fixed

• bug where similar_users(), similar_items() were performing validation checks on the original ID versus the zero-based index (wrong) instead of original values (correct) - this was causes a bunch of bogus assertion errors saying that the item_id wasn't in the training set

Added

• WARP loss - while slower to train this yields slightly better performance on dense interaction data and much better performance on highly sparse interaction data relative to BPR
• new hyperparameters loss and max_samples
• re-wrote the numba _fit() function to elegantly (IMHO) handle both BPR and WARP loss

Added

• added support for sample weights - you can now pass importance weights in addition to interactions
• automatically determine the input data class (np.ndarray vs. pd.dataframe/pd.series)
• assert/ensure that all model weights are finite after each training epoch to fail fast for exploding weights

Fixed

• bug where pd.dataframe interactions with columns not named [user_id, item_id] were not getting loaded/indexed correctly - fixed by using the input class determination utility created

Changed

• more efficient loops for updating item feature and user/item feature factor weights - this cuts training time by around 30% with no auxiliary features, and by 50%+ in the presence of auxiliary features

Added

• core package functionality
• example notebook: quickstart.ipynb
• source distribution and package wheel
• basic test suite
• CircleCI build, lint, test CI workflows