"Mining at scale with latent factor models for matrix completion."
Sapienza University of Rome, 2016.
Abstract: "Predicting which relationships are likely to occur between real-world objects is a key task for several applications. For instance, recommender systems aim at predicting the existence of unknown relationships between users and items, and exploit this information to provide personalized suggestions for items to be of use to a specific user. Matrix completion techniques aim at solving this task, identifying and leveraging the latent factors that triggered the the creation of known relationships to infer missing ones.
This problem, however, is made challenging by the size of today’s datasets. One way to handle such large-scale data, in a reasonable amount of time, is to distribute the matrix completion procedure over a cluster of commodity machines. However, current approaches lack of efficiency and scalability, since, for instance, they do not minimize the communication or ensure a balance workload in the cluster.
A further aspect of matrix completion techniques we investigate is how to improve their prediction performance. This can be done, for instance, considering the context in which relationships have been captured. However, incorporating generic contextual information within a matrix completion algorithm is a challenging task.
In the first part of this thesis, we study distributed matrix completion solutions, and address the above issues by examining input slicing techniques based on graph partitioning algorithms. In the second part of the thesis, we focus on context-aware matrix completion techniques, providing solutions that can work both (i) when the revealed entries in the matrix have multiple values and (ii) all the same value."