BABYL OPTIONS: -*- rmail -*-
Version: 5
Labels:
Note:   This is the header of an rmail file.
Note:   If you are seeing it in rmail,
Note:    it means the file has no messages in it.

1, filed,,
Summary-line: 26-Oct       thomaz@media.mit.edu [237] #MAS965 - Project Paper Preview
Received: from ml.media.mit.edu (ml.media.mit.edu [18.85.13.107])
	by aleve.media.mit.edu (8.9.3/8.9.3/+ALEVE) with ESMTP id NAA23089
	for <lieber@media.mit.edu>; Thu, 26 Oct 2000 13:29:20 -0400 (EDT)
Received: from localhost (thomaz@localhost)
	by ml.media.mit.edu (8.8.8/8.8.7) with ESMTP id NAA14675
	for <lieber@media.mit.edu>; Thu, 26 Oct 2000 13:29:19 -0400 (EDT)
Date: Thu, 26 Oct 2000 13:29:19 -0400 (EDT)
From: Edison Thomaz <thomaz@media.mit.edu>
To: lieber@media.mit.edu
Subject: MAS965 - Project Paper Preview
Message-ID: <Pine.OSF.4.21.0010261328240.32530-100000@ml.media.mit.edu>
MIME-Version: 1.0
Content-Type: TEXT/PLAIN; charset=US-ASCII
X-UIDL: a4b8c4bd4b80dfd899c8283204b21a8e

*** EOOH ***
Date: Thu, 26 Oct 2000 13:29:19 -0400 (EDT)
From: Edison Thomaz <thomaz@media.mit.edu>
To: lieber@media.mit.edu
Subject: MAS965 - Project Paper Preview


Hi Henry,

As requested, here's my project paper preview.

Thanks,

Thomaz

-----------------------------------------------------------

MAS965 - Fall 2000 
Edison Thomaz Jr. 


Tempo: A Context-Sensitive Search Engine 
Publication Preview 1


1. Abstract

I describe the design, implementation and analysis of Tempo,
a search engine that relies primarily on user context and
user behaviour to classify and categorize web pages. I
propose the utilization of context variables such as (1)
time and (2) user system state information to select the web
pages that are most likely to contain the content that users
are looking for.


2. Introduction

Without any doubt, one of the biggest problems of the
Internet today is that due to the exponential growth of
published information on the web in the past few years, it
is extremely difficult to find information on the net. Not
only are current search mechanism hard to use for novice and
experienced users, but they are also ineffective and
painfully frustrating most of the time. Several kinds of
search engines have been developed recently to address this
problem and with the exception of a very few, most of them
fall short of their goals. I would like to present an
approach to searching for information on the Internet that
relies on user context and is significantly novel and
unique.


3. Method

Instead of a seach engine that analyzes the lexical and
syntactical structure of pages and their contents, I propose
the creation of a search system that recommends web pages to
users according to a temporal and contextual analysis of how
previous users reacted to the results of similar search
queries. In other words, the search mechanism relies
primarily on how much time users spent looking for
information in web pages in the past to make link
recommendations for future users. In addition to that, other
contextual cues such as how the user navigated from one
search result page to another and the state of the users
computer during the search cycle can help determine which
pages the user found interesting and useful. The fundamental
idea behind this search mechanism is to use time and user
behaviour as the context for web page classification and
recommendation.


4. Implementation

The entire search engine consists on 4 pieces: a stand-alone
application, an application server, a database and a
conventional search engine that makes recommendations
according to the associativity of web pages or keywords such
as Google, for example.

4.1 Stand-Alone Application

The stand-alone application is the visual cortex of the
system. It is completely independent of web browsers and
contains the edit field where users will effectively type
their search queries. The app has to be completely browser
independent because they need to track the amount of time
that users spend going from one web page to another. The
search results might be displayed in the web browser or in a
list box within the stand alone app itself. I plan to write
this app as a native Macintosh application.

4.2 Application Server

The application server communicates with the stand-alone app
by means of an XML-derived language, possibly SOAP, and is
basically the middle-man between the database and the app. I
am investigating the possibility of using an open-source
application server called Enhydra as this component of the
system.

4.3 Database

The database is the key component of the system because it
will contain the associations between the keywords typed by
previous users of the system and the web pages where
previous users theoretically found what they were looking
for. There is an open-source RDBMS called InstantDB that
looks very promising for such a task. I have used it
successfully in the past.

4.4 Conventional Search Engine

The conventional search engine is important because
throughout the lifetime of the system, users will always
make new queries that are not associated with any previous
query. As a result of that, the database of the system will
not be able to return any contextual information of where
previous users found the information that they were looking
for. So, with a conventional search engine, the system can
return conventional search results to the users under these
scenarios and then make new context-related inferences about
the user search and then gather the necessary information
for future inquiries. I would like to use Google as this
component of the system.


5. Discussion

If a user is looking for some piece of information X and
we associate a time-line with the entire search process, at
time t=0s, the user finishes typing his or her query and
presses the Search button. At the very first run, the
system database will be empty, so the system returns results
provided by the conventional search engine, in this example
scenario, Google.

At time t=10s, lets say, the user clicks on the first
result returned by Google. At this time, the stand alone
application starts counting how many seconds the user spends
in the first result returned by Google and stores that in
memory. After 10 seconds, at t=20s lets say, the user
selects the second link returned by Google and spends 15
seconds in the web page associated with this link. The stand
alone app, still tracking the behaviour of the user, now
knows that the user spent 10 seconds in the first web page
returned by Google and 15 seconds in the second page
returned by Google. If the user selects the third web page
returned by Google at t=35s and spends 5 minutes there, then
the stand alone app, realizing that the third page attracted
the users attention for a much longer period of time,
infers that the third web page is much closer in providing
the information that the user is looking for than the first
two visited web pages. The stand alone app then writes that
inference into the database by storing the amounts of time
that the user spent visiting each page and by giving the
third page much higher priority than the first two pages. It
it important to mention that we cannot guarantee that the
third page presented by Google is the one that has exactly
what the user was originally looking for, but it is clear
that it drew the users attention long enough to be
considered relevant in one way or another.

The next time a user searches for X, the system will
already have some reference pages associated with such a
keyword and will be able to return a much more precise set
of links to the user, as opposed to returning the results of
Google. This set of links will be derived by making a
calculation, determined empirically, according to the amount
of time that previous users spent on each one of the
associated pages in the database. In this example case, the
third page , the one that attracted the attention of the
first user for the longest period of time, would be most
likely shown to the second user at the top of the list.


6. Conclusions and Future Work

I strongly believe that we can improve the efficiency and
effectiveness of Internet searches greatly if we take into
consideration the overall context of users at the exact
moment that they look for information. As far as
improvements for the search engine are concerned, I believe
that its accuracy can be significantly enhanced by taking
into account more contextual degrees of freedom. By doing
that it can learn more about the user and his or her search
needs implicity, without any external intervention. For
instance, it can develop a better undertanding of what the
user is trying to accomplish by studying how mouse movements
relate to search keywords and navigation patterns. The
search system can also study the current state of the users
computer and check which processes are often running when
certain searches are performed. An example of this would be
returning lots of multimedia-related links as search results
if the user is running some audio or video player in the
background. In this case, the fact that the user is watching
videos or listening to audio might be an evidence that the
user is looking for a particular type of content, perhaps
multimedia files. If the user is running a program like
Photoshop in the background, it is more likely that the user
is looking for clip art or stock photography when performing
a search. By utilizing statistical algorithms when studying
the data collected by the stand-alone application, the
search engine might be able to make several useful
inferences about the user and his or her goals.


7. Reference

1.	Bates, M.J. The design of browsing and berrypicking
techniques for the online search 		interface. Online
Review 13, 5 (October 1989), 407-24.

2.	Koenemann, J., and Belkin, N. A case for interaction:
a study of interactive information 		retrieval
behavior and effectiveness, in Proceedings of CHI '96
(Vancouver Canada, May 		1996), ACM Press, 205-12.

3.	Baldonado, Q., and Winograd, T. SenseMaker: An
Information-Exploration Interface 		Supporting the
Contextual Evolution of a User's Interests, in Proceedings
of CHI97 		(Atlanta USA, March 1997),