A deep dive into collaborative filtering: sparse matrices, cosine similarity and k-NN with a heap