Instructions to run experiments

Preparing datasets

To initiate datasets preparation process, one needs to call the corresponding script with the following command:

python scripts/get_data.py --dataset=<dataset name> [--pcore]

Supported datasets:

• Movielens (https://grouplens.org/datasets/movielens):
  • ml-1m, ml-10m, ml-20m, ml-25m, ml-latest
• Amazon (https://nijianmo.github.io/amazon/index.html):
  • Amazon_Fashion, All_Beauty, Appliances, Arts_Crafts_and_Sewing, Automotive, Books, CDs_and_Vinyl, Cell_Phones_and_Accessories, Clothing_Shoes_and_Jewelry, Digital_Music, Electronics, Gift_Cards, Grocery_and_Gourmet_Food, Home_and_Kitchen, Industrial_and_Scientific, Kindle_Store, Luxury_Beauty, Magazine_Subscriptions, Movies_and_TV, Musical_Instruments, Office_Products, Patio_Lawn_and_Garden, Pet_Supplies, Prime_Pantry, Software, Sports_and_Outdoors, Tools_and_Home_Improvement, Toys_and_Games, Video_Games

An example of the corresponding command for the Movielens-1M dataset:

python scripts/get_data.py --dataset=ml-1m

Tuning hyper-parameters

To start hyper-parameter tuning, use the command python tune.py with the following input arguments, either in the format of --parameter_name=parameter_value or simply --parameter_name for boolean parameters.

The list of accepted input arguments is the following:

• model The name of the model.
• dataset The name of the dataset. Note that when the --pcore setting is provided (default for Amazon datasets as they are 5-core in the source storage), the name of the dataset must be appended with the pcore value (i.e., 5 by default). For example, the Amazon Books datasets must be referenced as Books_5, if a different pcore value was not explicitly provided.
• time_offset A real number between zero and one. Used for splitting sequential observations into train and test. The time interval of observations is split in proportion to the time offset, which is simply a quantile of observations up to the splitting timepoint.
• config_path The path to the config file. Sample configs for various models are provided in the grid folder. For example, to use the default config for hypsasrec, use the option --config_path="./grids/hypsasrec.py".
• grid_steps A positive integer number. Determines the total number of configurations to be tried during the grid search.
• check_best Include this parameter, without any parameter values, to see the final results on the test data for the best configuration found.

Note that the model naming in the code is different form the one used in the paper:

• sasrec is the BCE-based Euclidean model
• sasrecb is the CE-based Euclidean model
• hypsasrecb is the BCE-based hyperbolic model
• hypsasrec is the CE-based hyperbolic model

An example of a full command:

python tune.py --model=sasrec --dataset=ml-1m --time_offset=0.95  --config_path="./grids/sasrec.py" --grid_steps=60

Reproducibility

Please see the conda.yml file for the vesrions of the Python packages used in the experiments. The best configs for several datasets are provided in best folder.