The document presents the CoFactor model, which jointly factorizes the user-item click matrix and item-item co-occurrence matrix to improve recommender system performance. CoFactor is motivated by word embedding models like word2vec that learn embeddings from word co-occurrence. It outperforms weighted matrix factorization on several datasets based on quantitative and qualitative evaluations. The authors analyze the model fits and show its benefits from accounting for item co-occurrence patterns in user data.