On Identifying Collections with Complex Objectives

(2013 - )

Project Description

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.

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.).

NSF CAREER Award (# 1253393), PI, 500K

Abstract

Publications

Datasets and Code available here]

Datasets and Code are available here

Demos

B. Lawson E. Galbrun, E. Terzi: Demo on recipe recommendations that satisfy budget and dietary constraints.

demo

Teaching activities

One of the key achievements of this project is the development of a new data-science course by the Evimaria Terzi during spring 2015. The course is a data-science course that introduces the students to basic data-mining techniques using python. The PI developed the course from scratch using ipython notebooks, which allowed for the class to be very interactive. The class had one homework every approximately two weeks, and the students developed a project in the end of the class. The projects were presented in an open poster session. The course was a big success and the CS department is turing it into a regular course that is going to be taught every semester (i.e., both fall and spring). The material of this class is made publicly available here:

available on github