MINING DISCRIMINATIVE ITEMS IN MULTIPLE DATA STREAMS by Zhenhua

How can we maintain a dynamic profile capturing a user’s reading interest against the common interest? What are the queries that have been asked 1, 000 times more frequently to a search engine from users in Asia than in North America? What are the keywords (or tags) that are 1, 000 times more frequent in the blog stream on computer games than in the blog stream on Hollywood movies? To answer such interesting questions, we need to find discriminative items in multiple data streams. Each data source, such as Web search queries in a region and blog postings on a topic, can be modeled as a data stream due to the fast growing volume of the source. Motivated by the extensive applications, in this thesis, we study the problem of mining discriminative items in multiple data streams. We show that, to exactly find all discriminative items in stream S1 against stream S2 by one scan, the space lower bound is Ω(|Σ| log n1 |Σ|), where Σ is the alphabet of items and n1 is the current size of S1. To tackle the space challenge, we develop three heuristic algorithms that can achieve high precision and recall using sub-linear space and sub-linear processing time per item with respect to |Σ|. The complexity of all algorithms are independent of the size of the two streams. An extensive empirical study using both real data sets and synthetic data sets verifies our design.