This is a production-ready, but very simple, content-based recommendation engine that computes similar items based on text descriptions. It comes with a sample data file (the headers of the input file are expected to be identical to the same file -- id, description) of 500 products so you can try it out.
It is a flas-based REST webservice designed to be deployed to Heroku and relies on Anaconda for installation of the scientific computing dependencies, and Redis to store precomputed similarities.
Read the comments in engine.py to see how it works. It's very simple!
web.py contains the two endpoints:
-
/train -- calls engine.train() which precomputes item similarities based on their descriptions in sample-data.csv using TF-IDF and cosine similarity.
-
/predict -- given an item_id, returns the precomputed 'most similar' items.
Create a new virtualenv with the needed dependencies. Note this
conda create -n crec --file conda.txt
Now, in the virtualenv (source activate crec):
python web.py
Then, in a separate terminal window, train the engine:
curl -X GET -H "X-API-TOKEN: FOOBAR1" -H "Content-Type: application/json; charset=utf-8" http://127.0.0.1:5000/train -d "{"data-url": "sample-data.csv"}"
And make a prediction!
curl -X POST -H "X-API-TOKEN: FOOBAR1" -H "Content-Type: application/json; charset=utf-8" http://127.0.0.1:5000/predict -d "{"item":18,"num":10}"
This engine is designed to be deployed to Heroku. First, create a new app:
heroku create
You'll then need to set the buildpack for the app to use Anaconda; a packaging system for scientific computing libraries in Python.
heroku buildpacks:set https://github.com/kennethreitz/conda-buildpack.git
Be sure to set your environmental variables (in settings.py) and provide your own training data. Then just:
git push heroku master
Well...technically it's running test, singular :)
python -m unittest tests