An example of training a triplet network for personalized and non-personalized recommendations using BPR loss.
Derived originally from an example by Maciej Kula with a number of changes.
See the main Jupyter notebook for full details. You can view this on GitHub or re-run it from scratch to reproduce the results.
In addition to the notebook, there are also three library files providing helper functions: data.py, metrics.py and net_helpers.py.
Tested in Python 2.7.13 from conda 4.3.22 distribution, plus:
- tensorflow=1.2.0rc0 built from source
- tensorflow-tensorboard=0.1.2 (pip)
- keras=2.0.4 (pip)
- pydot-ng=1.0.0.15 (conda)
- graphviz=2.38.0 (conda)
- graphviz=0.7.1 (pip)
Why are there two packages called graphviz? Well, confusingly, conda install graphviz only installs some binaries, not the actual Python graphviz package. pip install graphviz installs this. More details here.
TensorBoard is optional, you don't need this to run the notebook. But depending on how you install TensorFlow, you might have it already, without needing to install it separately.