CAREER: On the identification of collections with complex objectives

点击翻译按钮获取中文摘要

In many domains, there is an increasing reliance on Recommender Systems for helping identify products, services and people that meet some user-specified criteria. Given a pool of entities (e.g., movies, books, experts) and an objective function such systems have to identify a collection (i.e., a subset) of entities from the pool that optimizes the objective function. For example, in movie-recommendation systems (e.g., Netflix) the goal is to identify subsets of movies to recommend to registered users. Analogous problems arise in social networks and social media (e.g., Twitter, Facebook), where advertisers need to identify a small set of targets for their advertisements. Finally, project management teams in large organizations often use expertise management systems to identify the subset of experts needed to complete a specific project.Current Recommender Systems suffer from severe limitations in settings where (i) the users multiple interactions with the system over time and the recommendations provided to a specific user at any given time need to take into account the past recommendations given to the same user or (ii) The entities that make up the recommended collections are rational entities, e.g., participants in a social network, or members of a project team, that have their own goals and preferences that influence their behavior as members of the collection. This project aims to address these two shortcomings of current Recommender Systems by designing, implementing, and evaluating combinatorial algorithms for identifying (a) sequences of collections, rather than a single collection and (b) collections of rational entities with individual goals, preferences, or objectives.In addition to developing a suite of novel combinatorial algorithms and heuristics for recommending sequences of collections and collections of rational entities, the project aims to develop and deploy two application-specific testbeds: (i) a personalized meal planner provides to its users weekly meal recommendations to guide them towards healthy eating choices; and (ii) A crowdsourcing platform with support for virtual team formation to allow students registered in some of the courses at Boston University, to form teams online to collaborate on class projects (when appropriate). Broader impacts of this research include: new models and methods that signficantly advance the current state of the art in Recommender Systems, with broad applications in a number of domains including social networks (e.g., LinkedIn, Facebook, etc.), online recommendation systems (e.g., Amazon, Netflix, etc.), and daily-deal sites (e.g., Groupon, LivingSocial, etc.). The project contributes to the education and advanced research-based training of graduate and undergraduate students in Computer Science at Boston University. Wide dissemination of software implementations of the algorithms can be expected to benefit the larger research community.Additional information about the project, including links to project personnel, publications, and software can be found at: http://www.cs.bu.edu/~evimaria/recommendations.html