School of Computing and Information Sciences
Ramesh Baral is a Ph.D. candidate in the School of Computing and Information Sciences at FIU under the supervision of Professor Sitharama S. Iyengar. Ramesh received his B.E. in Software Engineering from Pokhara University, Nepal in 2007. Before joining FIU in Fall 2013, he was a Senior Software Engineer at Verisk Health Inc. Nepal for about six years. His research interests include social network analysis, data mining, natural language processing, and machine learning. He has published a book chapter and several journal and conference papers in renowned conferences and journals, such as ACM RecSys, IEEE IRI, ACM UMAP, IEEE MobileSoft, ACM SIGSPATIAL, ACM SIGKDD, AAAI, and Journal of Data Mining and Knowledge Discovery. He is a member of IEEE and ACM. He is a recipient of FIU Dissertation Year Fellowship.
The increasing volume of information has created overwhelming challenges to extract the relevant items manually. Fortunately, the online systems, such as e-commerce (e.g., Amazon), location-based social networks (LBSNs) (e.g., Facebook) among many others have the ability to track end users’ browsing and consumption experiences. The explicit experiences (e.g., ratings, reviews) and many implicit contexts (e.g., social, spatial, temporal, and categorical) are useful in preference elicitation and recommendation. As an emerging branch of information filtering, the recommendation systems are already popular in many domains, such as movies (e.g., YouTube), music (e.g., Pandora), and Point-of-Interest (POI) (e.g., Yelp).
The POI domain has many contextual challenges (e.g., spatial (preferences to near places), social (e.g., friend’s influence), temporal (e.g., popularity at certain time), categorical (similar preferences to places with same category), locality of POI, etc.) that can be crucial for an efficient recommendation. The reviews on POIs shared across different social networks provide more granularity in users’ consumption experience. From the data mining and machine learning perspective, following three research directions are identified and considered relevant to an efficient context-aware POI recommendation, (1) incorporation of major contexts into a single model and a detailed analysis of the impact of those contexts, (2) exploitation of user activity and location influence to model hierarchical preferences, and (3) exploitation of user reviews to formulate the aspect opinion relation and to generate explanation for recommendation.
This dissertation presents different machine learning and data mining-based solutions to address the above-mentioned research problems, including, (1) recommendation models inspired from contextualized ranking and matrix factorization that incorporate the major contexts and help in analysis of their importance, (2) hierarchical and matrix-factorization models that formulate users’ activity and POI influences on different localities that model hierarchical preferences and generate individual and sequence recommendations, and (3) graphical models inspired from natural language processing and neural networks to generate recommendations augmented with aspect-based explanations.