recommendations of content tied to a user

• NumPy (1.14.3+)
• SciPy (1.1.0+)
• sh (1.12.14+)
• TensorFlow (1.8.0+)
• Pandas (0.22.0+)
• Scikit-Learn (0.19.1+)
• Matplotlib (2.2.2+)

Additionally, to run the api you need:

• Flask (1.0.2+)
• Redis-Py (2.10.5+)

There are essentially three steps involved:

1. Preprocessing: preprocess.py Accepts either a JSON or a CSV file as input and outputs two Scipy sparse matrices: one for training another for testing.
2. Training: task.py Accepts the Scipy sparse matrices and trains the model on them
3. Prediction: predict.py Will output the top n ratings for a specified user id.

You can run any of these scripts with the -h or --help option for more information on supported options.

First of all, you need a dataset. You can use any that catches your fancy. We will be working the 5-core amazon music dataset

Now, we need to take this JSON and transform it into a trainging matrix and a testing matrix.

$ python preprocess.py --data Digital_Music_5.json --format json --col-order reviewerID asin overall --lines True

There should now be two files in your directory: train.npz and test.npz. We now train the model.

$ python task.py --train-data train.npz --test-data test.npz

Now, there should be a new directory called model with two files: row.npy and col.npy. To get predictions, run

$ python predict.py --u model\row.npy --v model\col.npy --user-id 12
[ 990 1973 1255 2268  644]

12 here is the row idex of the user in our matrix and [ 990 1973 1255 2268 644] are the column indices of the recommended music in our matrix. To recover the actual user id and music id, you need to set --save-map in preprocess.py and supply the two maps to predict.py. See help for more details. (Accesible by running python preprocess.py --help and python task.py --help)

At present, the API is tightly coupled with the Collaborative Communities project. It is only useful for making recommendations based on a CC user's viewing history. To use the API, you will need to have the event logging module installed.

To get recommendations, make a GET request to the server with the user id and (optionally) the number of recommendations needed.

eg: http://localhost:3445/rec?user=12&nrecs=3

Before the API is able to generate recommendations, it must be trained. To train the API make a POST request to the server specifying the URI of the logs and optionally specify the parameters for preprocessing and training.

eg: curl -i -X POST -H 'Content-Type: application/json' -d '{"article-view": "http://localhost:8000/logapi/event/article/view/?after=1970-01-01T00:00:00"}' http://localhost:3445/train

To visualise recommendations, make a GET request to the server. Optionally, you may specify the user id and the percentage of items to display.

eg: http://localhost:3445/visual?user=1&r=3

• Install redis: sudo apt−get install redis−server)

• Create a virtual environment: virtualenv --system-site-packages -p python3 rec_api

• Activate the virtual environment: source ~/rec_api/bin/activate

• Clone this repo: git clone https://github.com/fresearchgroup/Community-Recommendation.git

• Change into the directory: cd Community-Recommendation

• Install dependencies: pip3 install -r requirements.txt

• Set up Flask: export FLASK_APP=flask_api.py and, optionally, export FLASK_ENV=development

• Set the token for event logs: export LOG_AUTH_TOKEN=Your_Token_Here

• Run the server (eg: flask run --host 0.0.0.0 --port 3445)

• Install Docker and Docker-Compose
• Clone this repo: git clone https://github.com/woodsy-sounding-wilful-sapwood/Community-Recommendation.git
• Change into the directory: cd Community-Recommendation
• Add the event logs token: echo ”Your Token Here” >> .env
• Build the system: sudo docker-compose build
• Run the system: sudo docker-compose up