ASME - [A] [S]equential Recommendation [M]odel [E]valuator

Getting Started

The configuration files for "Enhancing Sequential Next-Item Prediction through Modelling Non-Item Pages in Transformer-Based Recommender Systems" can be found under non-items-paper-configs for the public datasets.

To train, install the dependencies as described below. Add the src folder to your pythonpath and run the main, e.g.

poetry run python -m plp_paper.plp_main train config_file

You can train a model with

train config_file

and evaluate with

evaluate --config-file config_file --checkpoint-file checkpoint

The ML-20m dataset will be downloaded automatically, the Coveo dataset (https://github.com/coveooss/SIGIR-ecom-data-challenge) has to be downloaded manually. To (re-)generate the coveo files for the first run, set perform_convert_to_csv: true in the config.

Experiments

The config files for "ml-20m" are split into the folders "ml-20m" for the models using items only, and "ml-20m-extended" for models using artificial non-iten pages. For the Coveo-Pageview dataset you can find the configs for items-only in "coveo-nd" and with non-item pages in "coveo-nd-pageview". For the Coveo-Search dataset you can find the configs for items-only in "coveo-slnd" and with non-item pages in "coveo-slnd-search".

Install Locally

• Install Poetry
• Clone the repository
• Build the development virtual environment: poetry install
• Enter the virtual environment: poetry shell

Docker

The docker file in /k8/dev can be used to create an image which automatically pulls and updates the specified branch from gitlab and uses poetry to execute asme.

Environment Variables

• PREPARE_SCRIPT: If set the script will be executed before the framework command is run. This is useful for copying data to a faster drive
• GIT_TOKEN: Token used to checkout the repository
• REPO_USER: User used to checkout the repository
• REPO_BRANCH: Branch to checkout
• PROJECT_DIR: A writeable path within the container, that will be used to store the repository

The development container does not come with a preinstalled version of the framework. Instead, the entrypoint will clone the specified branch and setup an environment using poetry. Your command will be executed inside this environment.

Generate Gitlab access token

Even though the password is supplied via an environment variable that can be populated by a kubernetes secret, you probably don't want to share it with the admins. Lucky for us, we can also generate a gitlab access token for that purpose.

1. Go to your Profile and select Settings->Access Tokens
2. Give it a name, e.g. k8s
3. Select read_repository
4. Save the final token, because it won't be accessible afterwards

Create kubernetes secret

kubectl -n <namespace> create secret generic gitlab-token --from-literal=token=xxx