poprox-recommender

This repository contains the POPROX recommender code — the end-to-end logic for producing recommendations using article data, user histories and profiles, and trained models.

• Installation
• Local Development
• Editor Setup
• License

Installation for Development

Software environments for this repository are managed with pixi, and model and data files are managed using dvc. The pixi.lock file provides a locked dependency set for reproducibly running the recommender code with all dependencies, on Linux, macOS, and Windows (including with CUDA on Linux).

To install Pixi on Linux or Windows, see the install instructions. To install on Mac, either do the same or use Homebrew (brew install pixi).

Once Pixi is installed, to install the dependencies needed for development work:

pixi install -e dev

Once you have installed the dependencies, there are 3 easy ways to run code in the environment:

1. Run a defined task, like test, with pixi run:

   pixi run -e dev test

2. Run individual commands with pixi run, e.g.:

   pixi run -e dev pytest tests

3. Run a Pixi shell, which activates the environment and adds the appropriate Python to your PATH:

   pixi shell -e dev

Note

If you have a CUDA-enabled Linux system, you can use the dev-cuda and eval-cuda environments to use your GPU for POPROX batch inference.

Note

pixi shell starts a new, nested shell with the Pixi environment active. If you type exit in this shell, it will exit the nested shell and return you to you original shell session without the environment active.

Finally, set up pre-commit to make sure that code formatting rules are applied as you make changes:

pre-commit install

Note

If you will be working with git outside of the pixi shell, you may want to install pre-commit separately. You can do this with Brew or your preferred system or user package manager, or with pixi global install pre-commit.

To get the data and models, there are two steps:

1. Obtain the credentials for the S3 bucket and put them in .env (the environment variables AWS_ACCESS_KEY_ID and AWS_SECRET_ACCESS_KEY)
2. dvc pull

Dependency Updates

If you update the dependencies in poprox-recommender, or add code that requires a newer version of poprox-concepts, you need to regenerate the lock file with pixi update. To update just poprox-concepts, run:

pixi update poprox_concepts

To update all dependencies, run:

pixi update

Note

We use Pixi for all dependency management. If you need to add a new dependency for this code, add it to the appropriate feature(s) in pixi.toml. If it is a dependency of the recommendation components themselves, add it both to the top-level dependencies table in pixi.toml and in pyproject.toml.

Note

Currently, dependencies can only be updated on Linux.

Local Endpoint Development

Local endpoint development also requires some Node-based tools in addition to the tools above — this install is automated with npm and pixi:

pixi run -e dev install-serverless

To run the API endpoint locally:

pixi run -e dev start-serverless

Implementation

Under the hood, those tasks run the following Node commands within the dev environment:

npm ci
npx serverless offline start --reloadHandler

Once the local server is running, you can send requests to localhost:3000. A request with this JSON body:

{
  "past_articles": [
    {
      "article_id": "e7605f12-a37a-4326-bf3c-3f9b72d0738d",
      "headline": "headline 1",
      "subhead": "subhead 1",
      "url": "url 1"
    },
    {
      "article_id": "a0266b75-2873-4a40-9373-4e216e88c2f7",
      "headline": "headline 2",
      "subhead": "subhead 2",
      "url": "url 2"
    },
    {
      "article_id": "d88ee3b6-2b5e-4821-98a8-ffd702f571de",
      "headline": "headline 3",
      "subhead": "subhead 3",
      "url": "url 3"
    }
  ],
  "todays_articles": [
    {
      "article_id": "7e5e0f12-d563-4a60-b90a-1737839389ff",
      "headline": "headline 4",
      "subhead": "subhead 4",
      "url": "url 4"
    },
    {
      "article_id": "7e5e0f12-d563-4a60-b90a-1737839389ff",
      "headline": "headline 5",
      "subhead": "subhead 5",
      "url": "url 5"
    },
    {
      "article_id": "2d5a25ba-0963-474a-8da6-5b312c87bb82",
      "headline": "headline 6",
      "subhead": "subhead 6",
      "url": "url 6"
    }
  ],
  "interest_profile": {
    "profile_id": "28838f05-23f5-4f23-bea2-30b51f67c538",
    "click_history": [
      {
        "article_id": "e7605f12-a37a-4326-bf3c-3f9b72d0738d"
      },
      {
        "article_id": "a0266b75-2873-4a40-9373-4e216e88c2f7"
      },
      {
        "article_id": "d88ee3b6-2b5e-4821-98a8-ffd702f571de"
      }
    ],
    "onboarding_topics": []
  },
  "num_recs": 1
}

should receive this response:

{
    "recommendations": {
        "28838f05-23f5-4f23-bea2-30b51f67c538": [
            {
                "article_id": "7e5e0f12-d563-4a60-b90a-1737839389ff",
                "headline": "headline 5",
                "subhead": "subhead 5",
                "url": "url 5",
                "preview_image_id": null,
                "published_at": "1970-01-01T00:00:00Z",
                "mentions": [],
                "source": null,
                "external_id": null,
                "raw_data": null
            },
            ...
        ]
    },
    "recommender": {
        "name": "plain-NRMS",
        "version": null,
        "hash": "bd076520fa51dc70d8f74dfc9c0e0169236478d342b08f35b95399437f012563"
    }
}

You can test this by sending a request with curl:

$ curl -X POST -H "Content-Type: application/json" -d @tests/request_data/basic-request.json localhost:3000

Running the Evaluation

The default setup for this package is CPU-only, which works for basic testing and for deployment, but is very inefficient for evaluation. The current set of models work on both CUDA (on Linux with NVidia cards) and MPS (macOS on Apple Silicon). To make use of a GPU, do the following:

1. Set the POPROX_REC_DEVICE environment variable to cuda or mps.

2. Run dvc repro under the eval or dev environment (using either pixi run or pixi shell).

Timing information for generating recommendations with the MIND validation set:

Machine	CPU	GPU	Rec. Time	Rec. Power	Eval Time
DXC	EPYC 7662 (2GHz)	A40 (CUDA)	45m¹	418.5 Wh	24m
MBP	Apple M2 Pro	-	<20hr²		30m²
MBP	Apple M2 Pro	M2 (MPS)	<12hr²

Footnotes:

1. Using 12 worker processes
2. Estimated based on early progress, not run to completion.

Editor Setup

If you are using VSCode, you should install the following plugins for best success with this repository:

• EditorConfig
• Python
• Ruff

When you open the repository, they should be automatically recommended.