Flexible and efficient exploration of rated datasets

As users increasingly rely on collaborative rating sites to achieve mundane tasks such as purchasing a product or renting a movie, they are facing the data deluge of ratings and reviews. Traditionally, the exploration of rated data sets has been enabled by rating averages that allow user-centric, it...

Full description

Bibliographic Details
Main Author: Kolloju, Naresh Kumar
Language:English
Published: University of British Columbia 2013
Online Access:http://hdl.handle.net/2429/44028
Description
Summary:As users increasingly rely on collaborative rating sites to achieve mundane tasks such as purchasing a product or renting a movie, they are facing the data deluge of ratings and reviews. Traditionally, the exploration of rated data sets has been enabled by rating averages that allow user-centric, itemcentric and top-k exploration. More speci cally, canned queries on user demographics aggregate opinion for an item or a collection of items such as 18-29 year old males in CA rated the movie The Social Network at 8:2 on average. Combining ratings, demographics, and item attributes is a powerful exploration mechanism that allows operations such as comparing the opinion of the same users for two items, comparing two groups of users on their opinion for a given class of items, and nding a group whose rating distribution is nearly unanimous for an item. To enable those operations, it is necessary to (i) adopt the right measure to compare ratings, and to (ii) develop e cient algorithms to nd relevant <user,item,rating> groups. We argue that rating average is a weak measure for capturing such comparisons. We propose contrasting and querying rating distributions, instead, using the Earth Mover's Distance (EMD), a measure that intuitively re ects the amount of work needed to transform one distribution into another. We show that the problem of nding groups matching given rating distributions is NP-hard under di erent settings and develop appropriate heuristics. Our experiments on real and synthetic datasets validate the utility of our approach and demonstrate the scalability of our algorithms. === Science, Faculty of === Computer Science, Department of === Graduate