Title: Ka Cheung Sia, Shenghuo Zhu , Yun Chi , Koji Hino , and Belle L' Tseng
1Capturing User Interest by Both Exploitation and
Exploration
Ka Cheung Sia, Shenghuo Zhu, Yun Chi, Koji
Hino, and Belle L. Tseng University of
California, Los Angeles, NEC Laboratories America
Not used
Application Scenario
NEC Laboratories America
A personal information manager that recommends
news articles based on users browsing history
and implicit feedbacks
- Unobtrusive and quick detection save users time
to train the system - Focused and diversified recommendation strike a
balance between relevance and novelty of
recommendations - Adapt to drift of user interests systems should
update users profile over the time
User study 45 categories from dmoz.org Arts/Archit
ecture Computers/E-books Science/Biology etc. Surv
ey of user interest before experiment Present 7
items each time Interleave 3 strategies randomly
- Features of the personal information manager
- Built as a browser plug-in user spend less
effort to familiarize with the interface - Interactive recommendation process faster
detection of user interest and require less
tuning by user - Privacy preservation
- users browsing histories are not revealed to
3rd party
Exploitation and Exploration (EE) framework
Ranking function of topics
Optimization goal click utility (user
satisfaction)
Click probability
Read probability
Exploitation
Exploration
Notations and assumptions ri ranking of topic
i ?i click probability of topic i, parameters
of Bernoulli distribution g(.) read probability
as a function of ranking K number of
topics Prior of ?i follows a Beta distribution
with parameters ai and ßi
ai number of times user clicked topic i ßi
number of times user not clicking topic i ?
weight of exploration (tuning parameter)
Model parameters update ai ai 1 if topic i
is clicked ßi ßi g(ri) if topic i is not
clicked
Simulations
User studies
EE generates 23 more clicks than the greedy
method
EE gives higher click utility and lower
estimation error than exploitation alone (greedy)
method
- Random samples of 45 topics from dmoz.org
- 3 strategies are interleaved randomly to avoid
potential bias
- 45 different topics, 7 items displayed at a time
- Simulate interest drift every 30 iterations
Accuracy of modelparameter estimation
Click utility under normal scenario
Click utility underinterest drift scenario
Click utility undernormal scenario
Click utility under interest drift scenario
User interface of study experiment