Implementation of baseline model for recommendation system with positive and unknown ratings.

A proposed approach is compared with the approaches described in Collaborative Recurrent Autoencoder: Recommend while Learning to Fill in the Blanks and in Large-Scale Off-Target Identification Using Fast and Accurate Dual Regularized One-Class Collaborative Filtering and Its Application to Drug Repurposing.

Setup

1. First, install docker using these instructions.
2. Build the docker image by running make build. This make take several minutes.
3. To run experiments, type make notebook. This starts a jupyter notebook server on http://localhost:9999

Sample usage

from preprocessing import TextPreprocessor
from models import BlockCoordinateAscent
from metrics import recall_lift

prep = TextPreprocessor(glove_components=300, min_df=5, max_df=0.4)
processed_articles = prep.fit(articles, window=10, epochs=100)
embedded_articles = prep.idf_embed(processed_articles)

model = BlockCoordinateAscent(train_ones, embedded_articles_idf)
model.fit(n_epochs=5)
predictions = model.predict()  # filling the missing values in the recommendation matrix

test_metrics = recall_lift(predictions, train_ones, test_ones, 300)
print(test_metrics.recall)

TODO

1. Unit tests