Date and Time: March 13, 2017, 1:00 pm – 2:00 pm
Location: CDM 708
Abstract: In recent years, recommender systems and social graph mining techniques using data-driven machine learning algorithms have been successfully employed to overcome the information overload and to extract insightful information in many online applications. Among many algorithmic solutions, the matrix completion which aims at recovering a low-rank matrix from a partial sampling of its entries, has been proven as a successful method in collaborative filtering for recommender systems (e.g., the Netflix problem), missing data prediction, dimensionality reduction, and social graph mining (e.g., link prediction and network completion).
However, matrix completion methods perform poorly in practice especially when the observed matrix is sparse, some of the rows or columns are entirely missing– the so-called cold start problem or the observed ratings are not sampled uniformly at random. Recently, there has been an upsurge interest in utilizing other rich sources of side information about items/users, such as social or trust/distrust relationships between users and meta-data about items to compensate for the insufficiency of rating information and mitigate the cold-start users/items problem.
In this talk, we introduce a novel algorithmic framework for matrix completion that exploits the similarity information about users and items to alleviate the data sparsity issue and specifically the cold-start problem. In contrast to existing methods, our proposed algorithm decouples the following two aspects of the matrix completion to effectively exploit the side information: (i) the completion of a rating sub-matrix, which is generated by excluding cold-start users/items from the original rating matrix; and (ii) the transduction of knowledge from existing ratings to cold-start items/users using side information. We provide theoretical guarantees on the recovery error of the proposed decoupled algorithm and show through experiments on real-world data sets to demonstrate the merits of decoupled matrix completion.
We then discuss a unified framework to aggregate multiple sources of side information about users/items into a single distance metric that can be used in different recommendation methods. By modeling different types of side information as a similarity/dissimilarity constraint graph between entities, we cast the problem of learning from multiple sources as a distance metric learning (DML) problem from constraint graphs and introduce an efficient algorithm to learn such a metric. Aggregation of such information is of great importance, especially when a single view of the data is sparse or incomplete.
Bio: Rana Forsati is a post-doc researcher and instructor at Computer Science and Engineering department at Michigan State University since 2014. She obtained her Ph.D. (with honors) from Shahid Beheshti University (formerly known as The National University of Iran) in 2014. She also spent a year as a visiting research scholar at the University of Minnesota during her Ph.D. studies. Her research interests include applied machine learning, data mining, and optimization with applications in recommender systems, social graph analysis, and natural language processing.
DePaul Analytics Group