Current students


Section: Computer Science and Engineering

Major Research topic:
Impressions-Based Recommender Systems: Algorithms and Evaluation

In today's digital world, users are accustomed to consuming digital services, such as online movie renting, music streaming, and online shopping. Several businesses have adapted their model to include themselves in the online market. Moreover, consumers have also changed their habits to do most of their tasks in an online-centered fashion. For instance, they do the expenses online, order food from home, and receive recommendations for new music via mobile applications.

Companies have started to offer personalized services on their platforms to grow their user base and captivate existing users. For instance, they recommend the content users like based on their information and behavior. Recommender Systems (RS) are software tools that provide personalized recommendations to users based on different factors, such as item descriptions, previous purchases, and more.

Research in Recommender Systems mainly focuses on a specific taxonomy of recommendation models. Those are called Collaborative Filtering (CF) recommenders. These recommenders model users' preferences toward items by leveraging their feedback with the system. Traditionally, this feedback is implicit. For instance, it is composed of clicks, purchases, likes, watch time, and other actions of users that may denote a certain level of preference towards the products from an online catalog.

Impressions-based recommender systems (IBRS) are a novel taxonomy in RS, with increasing research interest in recent years. IBRS leverages interactions and impressions, i.e., previously shown items on the screen. Hence, they can model more complex preference relationships between users and items as they learn from interactions and the possible options users had before the interaction (impressions). Currently, recommenders are unable to model this type of relationship. Furthermore, IBRS also enables the study of modeling of other topics of interest in RS research, such as user fatigue, biases in RS, feedback loops, echo chambers, and relevance of served recommendations. For instance, previous attempts have been made to model popularity biases.

This research work is focused on investigating different topics on IBRS from two related research directions: algorithms and evaluation. Mainly, from the algorithmic direction, the research is focused on reviewing past and proposing new IBRS recommendation models. The goal is that these recommendation models outperform traditional CF and IBRS recommenders. From the evaluation direction, the research is focused on the definition of sound and proper evaluation methodologies of IBRS. This characterization is lacking in the current literature and has been proven to be a fundamental and challenging task in RS research.