UCAIR: Capturing and Exploiting Context for
Personalized Search
Xuehua Shen, Bin Tan, ChengXiang Zhai
Department of Computer Science
University of Illinois at Urbana-Champaign
ABSTRACT
Personalized search has much to do with capturing and exploiting
user-related context information to improve search accuracy. Exist-
ing retrieval systems can not support personalized search well for
ignoring a user’s search context. In this paper, we describe our on-
going work on the User-Centered Adaptive Information Retrieval
(UCAIR) project, which aims at capturing and exploiting naturally
available user context for personalized search.
1. INTRODUCTION
Precise understanding of a user’s information need is essential
for achieving optimal retrieval performance. Most existing retrieval
systems take a user’s query as the sole source of knowledge about
the user’s information need. However, a query usually only consists
of a few short keywords, which are generally insufficient for giving
a complete and accurate picture about what the user is really look-
ing for. Thus using more context information about the user and
the query is necessary for improving the retrieval performance. In-
deed, personalized search essentially boils down to capturing and
exploiting related user context information of a query to improve
search accuracy.
Short term (dynamic) Long term (static)
Implicit immediately viewed document past query log
Explicit judged relevant documents occupation, hobbies
Table 1: Typology and examples of user context
As shown in Table 1, many kinds of user context information can
be potentially exploited [2]. Explicit context consists of informa-
tion given by a user explicitly, whereas implicit context refers to
any context information naturally available while a user interacts
with a retrieval system. While explicit context information is more
reliable than implicit context, it is often not available to us because
it requires extra effort from the user. Implicit context information
is thus more interesting to exploit [4, 5, 9].
The goal of the User-Centered Adaptive Information Retrieval
Copyright is held by the author/owner.
(UCAIR) project at the University of Illinois at Urbana-Champaign
is to capture and exploit such implicit context, especially short-
term context, to optimize retrieval results for a specific user to
achieve personalized search1. While the project is still in its early
stage, we have already achieved some interesting results: (1) We
have developed a general decision-theoretic framework for context-
sensitive retrieval. (2) We have developed specific retrieval models
for exploiting immediate search context based on statistical lan-
guage model; experiments show that such models achieve better
retrieval performance than those not using the context [6]. (3) We
have developed a client-side search agent that implements our pro-
posed models and algorithms for personalized web search. A user
study shows that the UCAIR search agent performs consistently
better than a popular search engine (Google), on which UCAIR
search agent is based. (4) We have obtained some experience with
evaluating context-sensitive IR. Below we summarize our current
work in these directions.
2. A DECISION-THEORETIC FRAMEWORK
FOR CONTEXT-SENSITIVE IR
To exploit context for personalized search in a general way, we
view the retrieval problem as a decision problem, in which all con-
textual information and the normally available query and docu-
ments should be considered together to optimize the retrieval deci-
sion. In general, in response to every user action, the system would
choose an optimal system action to take. For example, a user’s ac-
tion may be submitting a query and the system’s response may be
returning a list of 10 document summaries.
An advantage of treating retrieval generally as a decision-making
problem is that we may also treat a user’s viewing a document in
the search results as a user action, to which the system can respond
with updating its own user model about the user’s information need.
Although, in this case, such a response does not affect the user im-
mediately, we may imagine that after the user views the document
and returns to see more search results, the system can choose to
rerank any unseen search results based on the updated user model.
Indeed, to bring maximum benefit of context to the user, we would
like to exploit context as soon as it is available and respond imme-
diately based on any new piece of context information. Such “eager
feedback” is precisely what the UCAIR project is aiming at.
We have developed a decision-theoretic framework for optimiz-
ing interactive information retrieval based on eager user model up-
dating [7], in which the system responds to every user action by
choosing some system action to optimize a utility function. Specif-
ically, as soon as we observe any new piece of evidence from the
user, the system would attempt to perform two tasks: (1) compute
1UCAIR project web site: http://sifaka.cs.uiuc.edu/ucair/

the current user model to update its belief about the user’s informa-
tion need (2) choose a response that minimizes a loss function. For
example, immediately after the user views a document, we could
use the knowledge that the viewed document summary is proba-
bly relevant to rerank the unseen results so as to minimize a loss
function that favors a decision to rank relevant documents above
irrelevant ones.
In the traditional retrieval paradigm, the retrieval problem is cast
as matching a query with documents and rank documents according
to their relevance values. As a result, the whole retrieval process
is a simple independent cycle of “query submission” and “result
display”, which is inadequate for exploiting context. The decision-
theoretic framework we developed generalizes this traditional re-
trieval paradigm and allows us to exploit the user’s search context
in a quite general way.
3. LANGUAGE MODELS FOR CONTEXT-
SENSITIVE IR
When instantiating the general decision-theoretic framework de-
scribed above with specific retrieval methods, we obtain specific
retrieval models that can rank documents based on search context.
As a case study, we developed several different language models for
using implicit feedback information to improve retrieval accuracy
in interactive information retrieval [6]. We use the KL-divergence
retrieval model [10] as a basis and propose to treat context-sensitive
retrieval generally as estimating a query language model based on
the current query and any search context information. We proposed
and tested several statistical language models to incorporate query
and clickthrough history into the KL-divergence model, including
linear interpolation with fixed coefficients, Bayesian interpolation,
Online Bayesian updating and Batch Bayesian updating. In gen-
eral, the experiment results show that using implicit feedback in-
formation, especially the clickthrough data, can effectively and effi-
ciently improve retrieval performance without requiring additional
effort from the user at all [6].
4. A CONTEXT-SENSITIVE IR SYSTEM –
UCAIR SEARCH AGENT
We have developed a client-side search agent (called UCAIR)
embedded in a web browser which can capture a user’s search
context and perform implicit feedback [7]. The UCAIR search
agent incorporates models and algorithms proposed in section 3
to dynamically rerank the search results to reflect the most updated
knowledge of the user’s information need whenever any new piece
of implicit feedback becomes available.
We chose to do context-sensitive IR at the client side instead of
the server side as it has three remarkable advantages. First, the user
does not need to worry about privacy infringement, which is a big
concern for personalized search [8]. Second, a richer category of
user interactions such as mouse movement can be easily captured
for implicit feedback. Third, the computation needed for personal-
ization and the storage of the user profile are both done at the client
side, so the server is not burdened [3].
We implemented specific techniques to capture and exploit two
types of implicit feedback information: (1) identifying any related
immediately preceding query and using the query and its corre-
sponding search results to select appropriate terms to expand the
current query, and (2) exploiting the viewed document summaries
to dynamically rerank any document that has not yet been seen by
the user.
User studies show that the UCAIR search agent improves per-
formance over a popular search engine (Google), on which UCAIR
search agent is built.
5. EVALUATION OF CONTEXT-SENSITIVE
IR
Evaluation of context-sensitive IR poses special challenges due
to the difficulty in collecting appropriate user interaction data and
cleanly identifying baseline methods. For example, one challenge
in evaluating implicit feedback algorithms is that there does not
exist any suitable test collection for evaluation. In our study, we
used the TREC AP data to create a test collection with implicit
feedback information that can be used to quantitatively evaluate
implicit feedback algorithms. To the best of our knowledge, this is
the first test set for implicit feedback [6].
When evaluating the UCAIR search agent, we conducted a user
study involving 6 people. The participants are asked to do a web
search on selected query topics from TREC 2004 Terabyte track
and TREC 2003 Web track topic distillation task and then make rel-
evance judgments of the search results. By comparing our ranking
that incorporates context information and Google’s original rank-
ing, we can see whether the use of context information is bene-
ficial. Such a method [7] can be applicable to evaluating similar
context-sensitive retrieval systems.
6. FUTURE WORK
The current work can be extended in the following ways. (1)
We will further study the retrieval framework for sequential deci-
sion making in interactive information retrieval and study how to
optimize some of the parameters in the context-sensitive retrieval
algorithms. (2) We have only explored some very simple language
models for incorporating implicit feedback information. It would
be interesting to develop more sophisticated models to better ex-
ploit query history and clickthough data. For example, we may treat
a clicked document summary differently depending on whether the
current query is a generalization or refinement of the previous query.
(3) We will study other important user interactions. At the client
side, UCAIR search agent will capture and exploit many other user
actions such as mouse movement and dwelling time on a document,
which may have strong correlation with the document’s relevance
[1]. (4) Currently, the UCAIR search agent considers the server-
side retrieval system as a black box and therefore can not make use
of the server’s full-text indexing capability. We will study how to
make the client-side UCAIR search agent collaborate with the re-
mote retrieval system to provide more powerful contextual search.
7. ACKNOWLEDGEMENT
This work was supported in part by the National Science Foun-
dation grants CAREER-IIS-0347933 and ITR-IIS-0428472. Any
opinions, findings and conclusions or recommendations expressed
in this material are those of the author(s) and do not necessarily
reflect those of the National Science Foundation.
8. REFERENCES
[1] M. Claypool, P. Le, M. Waseda, and D. Brown. Implicit
interest indicators. In Proceedings of Intelligent User
Interfaces 2001, pages 33–40, 2001.
[2] P. Ingwersen and N. Belkin. Information retrieval in context
– IRiX. SIGIR Forum, 38(2), 2004.
[3] G. Jeh and J. Widom. Scaling personalized web search. In
Proceedings of WWW 2003, 2003.
[4] T. Joachims. Optimizing search engines using clickthrough
data. In Proceedings of SIGKDD 2002, 2002.