Dockerized Recommender system as REST API using Python, LightFM and Flask

This project was created as a follow-up wrapper for the paper that I prepared for the 2020 Lomonosov Conference "Economic overview of 2020s": https://lomonosov-msu.ru/rus/event/6500/page/1743.

My paper got selected for the "Digital transformation" section (please see p.23: https://www.econ.msu.ru/sys/raw.php?o=66528&p=attachment).

I currently do a lot regarding recommender systems at work (in the domain of food retail) so I decided to replicate some of the results using publicly available data.

To sum up, the end result of this research is a matrix factorisation-based recommender system for retail goods built on Flask and wrapped into a Docker container.

• Data source used: https://www.kaggle.com/carrie1/ecommerce-data

How to deploy the recommender

The model can be ran either in a regular mode or inside a Docker container.

All the main logic sits inside api.py file. It trains and serialises the model using pickle module.

To run the API run python api.py <port_number>

Endpoints

### /train (GET)
Train the recommender model, serialize it into a pickle dump, calculate model metrics and serialize user&item dictionaries.

### /wipe (GET)
Remove serialized model and all dirs with auxiliary data.

### /items_to_user (POST)
Recommend an array of items to a user. Accepts arguments:
    - user_id
    - nrec_items: number of items to recommend (gets fetch from model's 'predict')
    - show_known: show previously purchased/liked items by this user
    Might be used for personalised recommendations on the user's HOME SCREEN.

### /users_to_item (POST)
Recommend array of users to a particular item. Accepts arguments:
    - item_id
    - len_users: number of users to recommend
    Might be used for TARGETED PROMO OFFERS.

### /items_to_item (POST)
Recommend array of items to a particular item.  Accepts arguments:
    - item_id
    - n_items: number of items to recommnd
    Might be used for "YOU MAY ALSO LIKE"-type recommendations.

Working with Docker

One of the ways to make your Data Science (and, generally, all Software Engineering endeavours) project portable and easily deployable to production is packaging all the environment and your code into a Docker container.

It's somehing I often do at work and aim to achieve with this academic reasearch too.

1. Build docker image from Dockerfile

    `docker build -t "recommender" -f Dockerfile . `
   

2. Run docker container from image docker run -p 8891:8891 recommender

Sending real requests to recommender

You can send requests to the recommnder model either using Python requests module or straight from the command line using CURL. See files:

• mock_requests.sh
• mock_requests.py

Examples:

• POST:
  • curl -X POST "http://127.0.0.1:8891/items_to_user" -H "Content-Type: application/json" -d '{"user_id": "12947", "nrec_items": "10", "show_known": "True"}'
  • curl -X POST "http://127.0.0.1:8891/users_to_item" -H "Content-Type: application/json" -d '{"item_id": "22386", "len_users": "15"}'
  • curl -X POST "http://127.0.0.1:8891/items_to_item" -H "Content-Type: application/json" -d '{"item_id": "22466", "n_items": "11"}'
• GET:
  • curl -X GET "http://127.0.0.1:8891/wipe"
  • curl -X GET "http://127.0.0.1:8891/train"