Federated Bayesian Personalized Ranking

warning: some code improvements in progress

Reproduce experiments

Requirements

To use this Python script, install the requirements with:

pip install -r requirements.txt

To prepare the datasets, run the bash script download_raw_datasets.sh, so that the datasets are placed in the raw_datasets directory.

Then, use the Python script generate_dataset.py to create training, validation and test sets from raw datasets. In detail, use the arguments like in the following example:

python generate_dataset.py \
  --datasets Brazil Milan \
  --user_cut 20 \
  --item_cut 0 \
  --test_size 0.2 \
  --validation_size 0.2 \
  --parse_dates

Run the federated recommender

The script in this repository simulates in a single machine a federation of clients coordinated by a central server. As an example:

python main.py \
  --datasets MovieLens1M \
  --eval_every 5 \
  -F 50 \
  -lr 0.05 \
  -U 0.3 \
  -T single \
  -E 100

All the above-mentioned arguments and more options are completely described in the help python main.py -h.

Visualize results

For each experiment, the results are saved as raw recommendations in the folder ./results/<dataset>/recs/. The federated recommender saves a recommendation file in this folder each eval_every epochs. The structure of the files is the following:

user_id item_id predicted_rating