In this example, we implement HyperML (Lucas V. Tran, et al., HyperML: A Boosting Metric Learning Approach in Hyperbolic Space for Recommender Systems, WSDM'20) by curvlearn over the Kindle-Store dataset from Amazon recommender datasets.
We process the data to extract the 5+ cores, such that each of the remaining users and items have at least 5 ratings each. The data is then re-indexed and formatted by users into list of python list, i.e.,
[
[1, 2, 3, ..., 30], # Item list of user 1
[31, 32, 33, ..., 37], # Item list of user 2
...
[136372, 31588, ..., 29315] # Item list of the last user
]
The configurations of training are listed in config.py, leading to the following performance.
| Manifold | HR@10 |
|---|---|
| Euclidean | 69.89 |
| PoincareBall | 69.21 |
| Stereographic | 77.18 |
The entry of the training is train.py. Launch the training by
python examples/hyperml/train.py
and have fun!
The code has been tested under the following environment settings:
Hardware:
Tesla P100 - 16GB (Actual consumption: 1.4GB)
Intel Xeon E5-2682 v4 @ 2.50GHz
Python dependencies:
tensorflow-gpu==1.15.0
numpy==1.16.5