Scalable optimization algorithms for recommender systems

摘要

Recommender systems have now gained significant popularity and been widely used in many e-commerce applications. Predicting user preferences is a key step to providing high quality recommendations. In practice, however, suggestions made to users must not only consider user preferences in isolation; a good recommendation engine also needs to account for certain constraints. For instance, an online video rental that suggests multimedia items (e.g., DVDs) to its customers should consider the availability of DVDs in stock to reduce customer waiting times for accepted recommendations. Moreover, every user should receive a small but sufficient number of suggestions that the user is likely to be interested in.This thesis aims to develop and implement scalable optimization algorithms that can be used (but are not restricted) to generate recommendations satisfying certain objectives and constraints like the ones above. State-of-the-art approaches lack efficiency and/or scalability in coping with large real-world instances, which may involve millions of users and items. First, we study large-scale matrix completion in the context of collaborative filtering in recommender systems. For such problems, we propose a set of novel shared-nothing algorithms which are designed to run on a small cluster of commodity nodes and outperform alternative approaches in terms of efficiency, scalability, and memory footprint. Next, we view our recommendation task as a generalized matching problem, and propose the first distributed solution for solving such problems at scale. Our algorithm is designed to run on a small cluster of commodity nodes (or in a MapReduce environment) and has strong approximation guarantees. Our matching algorithm relies on linear programming. To this end, we present an efficient distributed approximation algorithm for mixed packing-covering linear programs, a simple but expressive subclass of linear programs. Our approximation algorithm requires a poly-logarithmic number of passes over the input, is simple, and well-suited for parallel processing on GPUs, in shared-memory architectures, as well as on a small cluster of commodity nodes.