SimDex: Exploiting Model Similarity in Exact Matrix Factorization Recommendations

We present SimDex, a new technique for serving exact top-K recommendations on matrix factorization models that measures and optimizes for the similarity between users in the model. Previous serving techniques presume a high degree of similarity (e.g., L2 or cosine distance) among users and/or items in MF models; however, as we demonstrate, the most accurate models are not guaranteed to exhibit high similarity. As a result, brute-force matrix multiply outperforms recent proposals for top-K serving on several collaborative filtering tasks. Based on this observation, we develop SimDex, a new technique for serving matrix factorization models that automatically optimizes serving based on the degree of similarity between users, and outperforms existing methods in both the high-similarity and low-similarity regimes. SimDexfirst measures the degree of similarity among users via clustering and uses a cost-based optimizer to either construct an index on the model or defer to blocked matrix multiply. It leverages highly efficient linear algebra primitives in both cases to deliver predictions either from its index or from brute-force multiply. Overall, SimDex runs an average of 2x and up to 6x faster than highly optimized baselines for the most accurate models on several popular collaborative filtering datasets.