Projects
My PhD Research
Date: September 2015 – Current
Recommender Systems (RS) in fast-paced dynamic scenarios must learn to adapt in response to interactive user evaluative feedback. In these settings, the RS faces an online learning problem and must balance two competing goals: exploit the known user model, or explore other preferences the user might have. My research aims to define the role of exploitation and exploration in RS, and propose mechanisms to balance and control this trade-off. The work takes inspiration from the Reinforcement Learning field where balancing the exploration-exploitation trade-off is key.
Keywords: Recommendation Systems, Reinforcement Learning, Multi-Armed Bandits, Information Adaptation.
IBM Research Africa
Data-Driven Policy Recommendations for Doing Business
Date: September 2017 – December 2017 (3 months)
I worked towards developing a data-driven solution to generate policy recommendations that can improve a government’s score in Doing Business.
Keywords: Markov Decision Process, Recommendation Systems, Human-in-the-Loop Machine Learning, Natural Language Processing.
RTÉ Industry/Research Project
Date: January 2015 – August 2015 (8 months)
RTÉ is the national television and radio broadcaster of Ireland. This project focused on building solutions for adapted content navigation and social community awareness to enrich the RTÉ Player service.
Supervision of Undergraduate Dissertation
Date: 2015
I had the pleasure to co-supervise Dulakshi Vihanga for her undergraduate dissertation tittled Experience-based Personalized Diversification of Recommendations.
Dulakshi received the Computing Project Award 2016 offered by the Sri Lanka Association for the Advancement of Science (SLAAS) as a result of this research project. The collaboration also resulted in two publications.
Master Dissertation
Diversification technique for Recommender Systems that Controls the trade-off between Exploitation of the User Profile and Exploration of Novel Products
Date: June 2014 – December 2014 (7 months)
Recommender Systems have emerged to guide users in the task of efficiently browsing/exploring a large product space, helping users to quickly identify interesting products. However, suggestions generated with traditional Recommender Systems usually do not produce diverse results, though it has been argued that diversity is a desirable feature. The study of diversity aware Recommender Systems has become an important research challenge in recent years, drawing inspiration from diversification solutions for Information Retrieval. However, we argue it is not enough to adapt Information Retrieval techniques towards Recommender Systems, as they do not place the necessary importance to factors such as serendipity, novelty and discovery which are imperative to Recommender Systems. In this report, we propose a diversification technique for Recommender Systems that generates a diversified list of results which not only balances the trade-off between quality (in terms of accuracy) and diversity, but also considers the trade-off between exploitation of the user profile and exploration of novel products. Our experimental evaluation, composed of both qualitative and quantitative tests, shows that the proposed approach has comparable results to state of the art approaches. Moreover, through control parameters, our approach can be tuned towards more explorative or exploitative recommendations.
The project’s webpage can be found here.
My dissertation can be found here and here.
Keywords: Recommendation Systems, Diversity, Exploitation and Exploration trade-off.
Towards Healthcare Analytics for the San Ignacio Hospital
Date: January 2013 – December 2013 (12 months)
Design of a plan of action for the San Ignacio Hospital to move towards Electronic Health Records.
The Lion framework for Agile Sofware Integration
Date: February 2012 – February 2013 (12 months)
Extend and consolidate the Lion framework for software generation. The new framework reduced the time required to integrate components to a code base from weeks/days to a few hours/minutes.
BESA Mobile: Towards a Mobile Multi-Agent Framework
Date: May 2011 – December 2012 (19 months)
Develop a Mobile Solution for the Multi-Agent BESA Framework by creating an interface component that allows for transparent communication between agents in a mobile device and a non-mobile device.
Undergraduate Dissertation
Design of a Generic Multidimensional System that provides Personalized Recommendation Service
Recommender Systems have emerged to help support, augment and systematize the everyday natural social process of creating and sharing recommendations by developing tools that can be used to quickly identify interesting products, and therefore, reduce a search space of alternatives. This project aims to construct a mechanism, under a generic approach, that can offer services to Information Retrieval applications so these may offer product recommendations that consider several Adaptation/Personalization dimensions (e.g., user dimension, context, among others). To begin with, a list of key design decisions for any type of Recommendation System solution that can also consider Adaptation/Personalization criteria is specified. Based on this list, the General Vizier Model was designed to support developers in the construction of any type of Multidimensional Generic recommendation solution. Finally, based on previous contributions, the Multi-Agent Vizier Recommendation Framework (Vizier) is proposed; on one hand, to assist those entities that currently develop Information Retrieval applications and wish to add recommendations to their services (e.g., E-Commerce applications); on the other hand, in order to offer a solution that hopefully provides better adapted/personalized results than current solutions by considering the multidimensionality of users, items and context. In order to validate Vizier, the initial Vizier prototype and initial ZoundBeat prototype were implemented.
The project’s webpage can be found here.
My dissertation can be found here.