A Family of Algorithms to Characterize Association in Click Model for Web Search

PhD Thesis Proposal Defence


Title: "A Family of Algorithms to Characterize Association in Click Model for Web
Search"

by

Mr. Weizhu Chen


ABSTRACT:

One of the major changes in the past decade is the heavy use of search 
engines which generates large-scale user activity data in Web search. 
These data have been in turn contributing to many critical Web tasks, such 
as optimizing search or sponsored results. As the most informative and 
reliable source of user action, click is believed to be the most important 
user activity in the data. Understanding the motivation behind a click or 
the decision making process to trigger it becomes the key to uncover the 
myth encoded in the data.

Yet, user click behaviour is complex, varying with users and implicit 
under various contexts. This poses challenges to characterize a click 
comprehensively. Many recent research works have attempted to model user 
click behaviour in a structural manner and positioned it as a click model 
problem, with the intention to better exploit user click behaviour so as 
to predict user click or estimate a user-perceived relevance for each 
query-document pair.

Despite of their success, most existing click models treat the modelled 
objects, such as queries, users, sessions, in isolation, disregarding 
their relationships. This may bring a simplification to the model but 
simultaneously sacrifices much valuable information, and hence interpret 
user click data in an inaccurate or biased way. This thesis proposal puts 
forward a family of algorithms to address these limitations. Our object is 
to characterize multiple associations among objects as well as design 
novel collective click models, where multiple modelled objects and their 
relationships are involved and associated together.

This algorithm family will depict the associations from three facets: 
region-based, query-based and user-based associations. Region-based models 
focus on the interplaying between organic search and sponsored search, so 
that it can depict user behaviour in the whole page thoroughly. 
Query-based models first collectively investigate the multiple queries 
with their corresponding clicks in a same session by designing a 
session-based click model. Then, it scrutinizes and uses the rich 
information in the high-frequent queries to alleviate the sparseness of 
long-tailed queries. User-based models characterize the user-centric click 
behaviour to design a personalized click model to entertain each 
individual user. These user-based models can also be tailored to better 
solve the sparseness challenges in the long-tailed queries. Finally, an 
ongoing feasibility study with extensive investigations showcases the 
practicality of the associative click models and future works are proposed 
to solve the problem to a next level.


Date:                   Friday, 25 November 2011

Time:                   1:00pm - 3:00pm

Venue:                  Room 3304
                         lifts 17/18

Committee Members:      Prof. Qiang Yang (Supervisor)
                         Prof. Dit-Yan Yeung (Chairperson)
 			Prof. Dik-Lun Lee
 			Dr. Raymond Wong


**** ALL are Welcome ****