Midterm colloquium Rohan Chandrashekar

18 May 2022 14:00 till 14:30 - Location: Bernd Schierbeek room, IDE - By: DCSC | Add to my calendar

"Graph-based Tensor Completion for Recommender Systems"

Increased volumes of data being processed to provide users with recommendations of their preference have paved the way to build Recommender Systems (RS) that can achieve this task. RS models are deployed in various scenarios ranging from e-commerce to music and movie recommendation engines. Building scalable and accurate RS models have been increasingly sought after, with data becoming exponentially sparse for decades. Current recommendation models predominantly use Collaborative Filtering (CF) algorithms due to their effectiveness in providing recommendations. However, traditional CF algorithms comprising Matrix Factorization and kNN models have drawbacks while handling either sparsity of the present data or scalability of the given model. Thus, to develop better RS models, we chose two specific structures: tensors and graphs, that can represent the available data and capture higher-order interactions that might be present between them.  

The use of tensors is profoundly motivated because matrices cannot capture data with higher-order relations. The transition to using tensors has been made promising with the advent of efficient tensor decomposition methods. Graphs are another powerful data structure that can capture relations between various entities that the available data might not explicitly dictate. The use of graph signal processing makes it more lucrative to extract information that is intrinsic to the underlying graph structures. In particular, it highlights that using graph information as regularizing coefficients could benefit tensor decomposition methods while building scalable and robust RS models. Combining these two heterogeneous data sources will facilitate maintaining any structural information that the data contains while alleviating issues of sparsity and scalability that might be seen with traditional methods.

Supervisors:
Dr. K. Batselier and E. Isufi

/* */