SWAMI: Searching the Web with Agents having Mobility and Intelligence MCS Defense - PowerPoint PPT Presentation

Loading...

PPT – SWAMI: Searching the Web with Agents having Mobility and Intelligence MCS Defense PowerPoint presentation | free to view - id: 13be3b-OWNlM



Loading


The Adobe Flash plugin is needed to view this content

Get the plugin now

View by Category
About This Presentation
Title:

SWAMI: Searching the Web with Agents having Mobility and Intelligence MCS Defense

Description:

The Web grows beyond the ability to navigate it. ... Work as the user browses (not offline); Trigger search mechanisms to find more pages; ... – PowerPoint PPT presentation

Number of Views:74
Avg rating:3.0/5.0
Slides: 52
Provided by: glass
Category:

less

Write a Comment
User Comments (0)
Transcript and Presenter's Notes

Title: SWAMI: Searching the Web with Agents having Mobility and Intelligence MCS Defense


1
SWAMISearching the Web with Agents having
Mobility and IntelligenceMCS Defense
  • presented by
  • Mark Kilfoil, BCS

Supervisor A. Ghorbani
2
Schedule
  • Introduction
  • Background Motivation Overview
  • Implementation
  • Design Construction
  • Evaluation
  • Goals Methodology Interpretation
  • Conclusion
  • Summary Contribution Future Work

3
Introduction Background A Confused User
4
Introduction Background An Assisted User
5
Introduction Background
  • The Web grows beyond the ability to navigate it.
  • The Web is a highly irregular space, so some
    intelligent tool is required to assist users.
  • Users are different, so the tool should reflect
    the particular user personally.
  • Users are the same, so the tool should use
    similarities between users to assist a particular
    user.

6
Introduction Background Adaptive Web Systems
  • Adaptive Web systems (both client, server and
    proxy) are one approach change the browsing
    experience based on evidence of user interest or
    common user behaviour.
  • In particular, the changes important here are in
    the adaptation of navigation.
  • Navigation suggestions given should be based on
    user interests, which are multiple and changing.

7
Introduction Background User Interest
A user's level of interest in a particular topic
changes over time
8
Introduction Motivation
  • To build a framework for an intelligent web
    browsing assistant that can
  • Learn a user's multiple, distinct interests on
    its own
  • Reflect the user's changing interests
  • Work as the user browses (not offline)
  • Trigger search mechanisms to find more pages
  • Allow community-like interaction with other
    users browsing.
  • To synthesize multiple current technologies into
    one framework
  • Hierarchical topic relationships
  • Dynamic user profiles
  • Agent technologies
  • Page representation and similarity.

9
Introduction Overview
  • SWAMI
  • Agents organize themselves into a representation
    of their user by watching user browsing
    behaviour
  • These agents can create specialized search agents
    to perform specific searches for peak interests,
    rather than all interests.

10
Introduction Overview
11
Implementation Design Topic Structure
12
Implementation Design Page Cluster Structure
13
Implementation Design Agent Structure
14
Implementation Design SWAMI Structure
15
Implementation Design Page Representation
  • Turn page into a bag of words
  • remove special characters and turn into lowercase
  • stem words (Porter stemming)
  • remove words on stoplist
  • count frequency of word occurrence
  • For each context the page is to be considered in,
    calculate the weight of each term

16
Implementation Design Agent Wealth
  • In order to determine if an agent is doing well,
    a score was created which takes into account
    arbitrary measures of success wealth.
  • The wealth of an agent combines four factors
  • Size of agent this measures not just as the raw
    number of pages that the agent represents, but
    considers how many pages the agent could have,
    and how recently it acquired pages.
  • Search success this measure combines both the
    quality of the pages found and how recently they
    were found.
  • Search acceptance this measure compares the
    number of recommendations made to the number of
    recommendations followed.
  • Previous wealth this measure simply allows an
    agent which has performed well to rest on its
    laurels and diminish more slowly.

17
Implementation Design Agent Wealth Formula
18
Implementation Design Agent Lifecycle
  • Birth when a parent agent discovers a subcluster
    of pages.
  • Aging as each page visit passes, the agent will
    diminish in wealth if nothing positive happens
    for it.
  • Retirement if an agent falls below a threshold
    of wealth, it will be placed in retirement.
    While in that area, it cannot execute searches or
    cannot have children, but it continues to be able
    to bid on incoming pages and continues to age.
    Thus, their wealth continues to diminish over
    time.
  • Reinstatement an agent can come back from
    retirement if a) it wins a bid for an incoming
    page, or b) an agent wants to create a new agent,
    but the retired agent is close enough to what the
    new agent would be.
  • Removal the agent is removed entirely from the
    system if its wealth drops below a lower
    threshold.

19
Implementation Design Lifecycle States
20
Implementation Design Similarity Clustering
A cluster of 6 pages A, B, C, D, E, F
Page A is closest to B page B is closest to C
and page C is closest to B.
Thus, A, B, C is a potential subcluster.
Similarly, D, E, F is another subcluster.
21
Implementation Design Searching
  • Proximity-Priority Search starting with a given
    set of pages, follow the most promising links
    that are close the starting point repeat.
  • Rendezvous Server send a representative to a
    community to interact with other representatives.
  • Local Expert send a representative to a remote
    host to interact with local experts with inside
    knowledge.
  • Search Engine-Based (unimplemented) submit
    feature terms to a search engine and
    filter/measure results.

22
Implementation The Integrated Browser
23
Implementation Browser Agent Details
24
Evaluation Goals
  • Can this system model a users interests? Are
    short-, long- and recurring-term interests
    identified and handled appropriately?
  • Can the system create appropriately specialized
    searches?
  • Can knowledge be shared?
  • Note efficiency not a goal!

25
Evaluation Methodology
  • Closed, artificial set of web pages representing
    reasonable assumptions
  • Pages are created to reflect membership in
    groups if the system places pages from the same
    group together (without knowing about the group
    explicitly), it has grouped them successfully.
  • Links between pages are created to reflect the
    assumption that they are there because the pages
    most likely are similar.
  • Given assumptions about user behaviour, simulate
    users, watching for the appropriate reactions
    from the system.
  • A user following a set of pages within the same
    group at the same time indicates an interest is
    rising if that interest is sufficient, a new
    agent representing it should be created.
  • Repeatedly browsing a set of pages from the same
    group should reinforce a representation of those
    pages.
  • Ignoring a group of pages should devalue those
    agents which represent it.

26
Evaluation Interpretation
27
Evaluation Searching
First instance
Second instance
28
Conclusion
  • SWAMI is a framework which captures a users
    multiple interests with varying level of details
    by observing what they browse.
  • SWAMI can reflect short, long and recurring
    interests.
  • SWAMI can trigger specialized searches for
    important interests.
  • SWAMI allows information to be shared with a
    community of users, to allow a word of mouth
    collaborative search mechanism.

29
Conclusion Contribution
  • Synthesized existing technologies and ideas.
    Innovations include
  • applying agents to the incremental, hierarchical
    clustering
  • simple partitioning mechanism
  • rarified features in a hierarchical clustered
    environment
  • allow multiple types of search mechanisms in the
    same architecture
  • proximity-priority search
  • rendezvous server.

30
Conclusion Contribution
  • Created an incremental, hierarchical, agent-based
    user interest representation with some search
    capabilities.
  • Demonstrated that such a system can capture the
    multiple interests and the different time-scales
    of interest, and can trigger specialized searches.

31
Conclusion Future Work
  • More dynamic weighting of important weight
    factors. Should other factors be considered?
  • Real user testing on real pages.
  • Efficiency!
  • Multiple, simultaneous search mechanisms and a
    way to decide between them!
  • Alternative page representation and similarity
    functions.

32
Conclusion Summary
  • A dynamic, incremental, hierarchical, active
    profile can capture a user's interests and the
    changes in those interests over time.
  • While it has been demonstrated to be feasible, it
    is calculation intense (due to the large number
    of comparisons).
  • SWAMI provides a basis for experimenting further
    with page representation and comparison methods,
    search and collaboration methods, and interest
    identification and representation.

33
Selected Bibliography
  • Balabanovic, M. (1997). An adaptive web page
    recommendation service. In Proceedings of the 1st
    International Conference on Autonomous Agents,
    pp378-385, Marina del Rey, CA, USA.
  • Brusilovsky, P. (1997). Efficient techniques for
    adaptive hypermedia. Intelligent Hypertext
    Advanced Techniques for the World Wide Web, 1326
    pp12-30.
  • Chen, C. C. and Chen, M. C. (2002). PVA A
    self-adaptive personal view agent. Journal of
    Intelligent Information Systems,
    18(2/3)pp173-194.
  • Freitag, D., Joachims, T., and Mitchell, T.
    (1995). WebWatcher A learning apprentice for the
    World Wide Web. Working Notes of the AAAI Spring
    Symposium Information Gathering form
    Heterogeneous, Distributed Environments, pp6-12.
  • Godoy, D. and Amandi, A. (2000).
    PersonalSearcher An intelligent agent for
    searching web pages. In Advances in Artificial
    Intelligence, IBERAMIA-SBIA 2000, volume 1952 of
    Lecture Notes in Artificial Intelligence,
    pp43-52. Springer.
  • Lieberman, H. (1995). Letizia An agent that
    assists web browsing. In Proceedings of the
    Fourteenth International Joint Conference on
    Artificial Intelligence, pp924-929, San Mateo,
    CA, USA. Morgan Kaufmann Publishers Inc.
  • Schwab, I., Kobsa, A., and Koychev, I. (2000).
    Learning about users from observation. In
    Adaptive User Interfaces Papers from the 2000
    AAAI Spring Symposium, Menlo Park, CA, USA. AAAI
    Press.

34
Comments / Remarks
  • Any questions?

35
The End!
36
Implementation Design
  • Diagram of system. Diagram should show Browser,
    Interface Agent (with Retirement Area),
    Representation Agents, Search Agents, Rendezvous
    Server and Local Smart Server.

37
Implementation Design
  • Diagram of clustering. Shows the mandala of
    the clustering system, which contains several
    pages and the similarities between them. It
    highlights the best mutual similarities,
    demonstrating how you put together a cluster.

38
Implementation Design Page Representation
  • Page features and feature weights. Note that
    features that exist in a supercluster are
    diminished when determining features in a
    subcluster (because all feature in the
    supercluster are already likely to be in the
    subcluster, and thus, they dont tell you
    anything special about the subcluster).

39
Implementation Construction
  • Simple, custom agent solution for specific needs.
  • Standard page representation, similarity
    measurements.
  • Simple best-fit clustering.
  • A wealth calculation for each agent which
    contains several arbitrarily-constructed factors.

40
Implementation Design Rarifying Features
41
Implementation Design Page Representation
A page is represented as a bag of words that is,
a set of words without any respect to the order
in which they appear.
Each word is de-capitalized and common
prefixes and suffixes are removed (via Porter
stemming).
The words are filtered through a stoplist, which
removes common but generally unimportant words
like and, but, or the.
42
Implementation Design Proximity-Priority Search
  • Proximity-Priority Search starting with a given
    set of pages, follow the most promising links
    that are close the starting point repeat.

43
Implementation Design Rendezvous Server
  • Rendezvous Server send a representative to a
    community to interact with other representatives.

44
Implementation Design Local Expert
  • Local Expert send a representative to a remote
    host to interact with local experts with inside
    knowledge.

45
Implementation Design Search Engine-Based
  • Search Engine-Based (unimplemented) submit
    feature terms to a search engine and
    filter/measure results.

46
Evaluation Interpretation
  • Show short, long, recurring and diminishing
    interests.
  • Show Rendezvous Server.

47
Implementation Design
  • Discussion of how short, long and recurring
    interest can be judged using weight function.
  • Show factors involved in agent performance,
    leading up to the weight function. Dont explain
    the whole function here, but have the slides on
    hand to show the formulas.

48
Implementation Design
  • Hierarchy of agents, organized based on groupings
    of pages.
  • A master agent, the Interface Agent, interacts
    with the Browsing Window.
  • Representation agents form the reflection of the
    users interests.
  • Agents grow when they appear to represent well,
    diminish when they are less important, and
    eventually fade away.
  • Search agents incorporating specific search
    capabilities can be spawned by representation
    agents when they are wealthy enough.

49
Implementation Design Agent Wealth Formula
50
Implementation Design Agent Wealth Formula
51
Implementation Design Agent Wealth Formula
About PowerShow.com