search

Skwiwel / uni-recommendation-system

A university project focused at engineering an ML solution to a product recommendation problem.
1 stars 0 forks source link
docker go ml python rest-api

readme

Recommendation system

A university project focused on engineering an ML solution to a problem. Chosen was the subject of an recommendation system for an online shop.

The publicly shared project repo contains no usable data. Sample data, cleaned of most of recognizable information, is located in data_raw/.

Contributors

Functionalities

Two simple recommendation models:

Additionally, a microservice serving the recommendations on a per-user basis was implemented. The service offers running of a transparent (to the user) A/B test functionality for comparison of the aforementioned models. The service offers a JSON based RESTful interface.

Compiling

The data handling parts of the project are made using loose python scripts. To install the required Python libraries:

python3 -m pip install -r requirements.txt

The microservice is implemented with go and internally uses an SQLite database for A/B testing. To compile the service:

go get github.com/mattn/go-sqlite3
cd ./microservice/go
go build -o ./../../service.exe ./cmd

This compiles the microservice into service.exe located in the root directory of the repo. The .exe extension is mainly for compatibility with Windows.

Running

The required data sessions.jsonl and products.jsonl needs to be located in data_raw.   For A/B testing the database in ab_testing/ needs to be manually reset by removing the trailing .init.

Running with A/B testing enabled:

./service.exe --mode abtest

Data relevant for A/B testing is saved to the database.

The service interface takes in request via HTTP GET at the location /recommendations. The user for whom we are requesting recommendations should be specified by a user_id and the optionam number of recommendations by num.
For example:

curl -i "<service_address>/recommendations?user_id=106&num=10"

Docker image

The microservice is also packed into a Docker image available at DockerHub.
Running it will require mounting the data_raw/ dir containing actual data as well as the ab_testing dir with a db for A/B testing.

To run:

docker pull skwiwel/uni-recommendation-system
docker run \
    --name recommendations \
    -p 80:80 \
    --mount type=bind,source="$(pwd)"/data_raw,target=/uni-recommendation-system/data_raw,readonly \
    --mount type=bind,source="$(pwd)"/ab_testing,target=/uni-recommendation-system/ab_testing \
    skwiwel/uni-recommendation-system --mode abtest

The last --mode abtest argument is passed to the service executable and runs the service in A/B testing mode.

System demo

The service is (hopefully still) currently running on a cloud container under the public ip 130.61.188.211.
You can check out how it's working by requesting recommendations, e.g.:

curl -i "130.61.188.211/recommendations?user_id=106&num=10"