Centralization vs. Decentralization Issues in Internet-based KMS: Experiences from Expertise...
description
Transcript of Centralization vs. Decentralization Issues in Internet-based KMS: Experiences from Expertise...
July 13, 2000 TWIST 2000 Yimam & Kobsa
Centralization vs. Decentralization Issues in Internet-based KMS: Experiences from
Expertise Recommender Systems
Dawit Yimam, GMD-FIT.MMK &
Alfred Kobsa, UCI, ICS
July 13, 2000 TWIST 2000 Yimam & Kobsa
Outline
• Background• First centralized approach• Alternatives - to centralize or decentralize ?• DEMOIR• Summary
July 13, 2000 TWIST 2000 Yimam & Kobsa
Expert Recommenders/Finders
• Systems to help users in tracing human information and/or expertise sources in organizations
• part of knowledge management and knowledge sharing services.
• Traditionally done by manual construction and search of expertise descriptions of people, e.g.,
+ Expert Databases (“knowledge directories”)+ Personal web pages on the Web
• Automatically mining implicit sources of expertise evidence from electronic resources of an organization and its people.
Background
Alternatives
First appr.
DEMOIR
Summary
July 13, 2000 TWIST 2000 Yimam & Kobsa
Characterizing Expert Finders
1. Expertise evidence/indicator source recognition and gathering
2. Expertise modeling- Expertise indicator extraction- expertise model representation
3. Expertise model deployment- query mechanisms- matching operation- output delivery/presentations- adaptation and learning operations
Background
Alternatives
DEMOIR
Summary
First appr.
July 13, 2000 TWIST 2000 Yimam & Kobsa
Query-time expertise modeling
Web Site Indexing
Web DocumentsIndex
Background
Alternatives
DEMOIR
Summary
GlimpseFIT Peoples’and otherWeb Pages WebGlimpse
First appr.
July 13, 2000 TWIST 2000 Yimam & Kobsa
Query-time expertise modeling
Query(Boolean)
Background
Alternatives
DEMOIR
Summary
Web Site Indexing
Web DocumentsIndex
GlimpseFIT Peoples’and otherWeb Pages WebGlimpse
ExpertQuery
Interface
First appr.
July 13, 2000 TWIST 2000 Yimam & Kobsa
Query-time expertise modeling
ExpertDatabase
(Name, URL)
SearchRanked List of
ExpertsBackground
Alternatives
DEMOIR
Query(Boolean)
Web Site Indexing
Web DocumentsIndex
GlimpseFIT Peoples’and otherWeb Pages WebGlimpse
ExpertQuery
Interface
Search Result(passages containing
Keywords)
ExpertiseModeler& Tracer
First appr.
Summary
July 13, 2000 TWIST 2000 Yimam & Kobsa
Query-time expertise modeling
• Shortcomings:
+ high latency in query processing+ personal sources hard to include+ non-document sources (e.g. recommendation from people,
social relations, etc.) hard to include+ full reliance on availability of some search engine+ limited exploitation of info due to lack of persistent expertise
models
Background
Alternatives
DEMOIR
First appr.
Summary
July 13, 2000 TWIST 2000 Yimam & Kobsa
Building apps on text Indexes
• Existing Web indexing systems use centralized indexes of distributed resources/collections.
• Distributed Indexing needed to cope with ever growing information on Internet.
• But, currently centralized global indexes (though may be distributed in a tightly coupled manner) consistently outperform decentralized indexing and query approaches.
• This favors centralizing the applications to be built on them.
Background
Alternatives
DEMOIR
First appr.
Summary
July 13, 2000 TWIST 2000 Yimam & Kobsa
Pre-generation of Expertise Models
• Alternative 1: Personal expert finding agents + Decentralized multi-agent system.+ Expertise modeling as well as searching done by self-managing
personal agents residing in experts’ computers (e.g. Vivacqua, 1999; Foner, 1997).
• Alternative 2: Aggregated expertise modeling+ Based on centralized expertise models (that are either
dynamically aggregated or linked to a pre-constructed ontology) (e.g. simple versions in Kautz & Selman, 1998; Krulwich & Burkey, 1996).
+ Can be distributed among tightly coupled cluster of machines.
Background
Alternatives
DEMOIR
First appr.
Summary
July 13, 2000 TWIST 2000 Yimam & Kobsa
Personal expert finding agents
Agent communicatio
n
Agent1
-----------------ModelExpertise
FindExpert
PersonalExpertise
Model
Agent2
-----------------ModelExpertise
FindExpert
PersonalExpertise
Model
Agentn
-----------------ModelExpertise
FindExpert
PersonalExpertise
Model
Agent3
-----------------ModelExpertise
FindExpert
PersonalExpertise
Model
Background
Alternatives
DEMOIR
First appr.
Summary
July 13, 2000 TWIST 2000 Yimam & Kobsa
Aggregated Expertise Modeling
ExpertFindingServer
AggregatedExpertise
ModelBackground
Alternatives
DEMOIR
First appr.
Summary
July 13, 2000 TWIST 2000 Yimam & Kobsa
Aggregated Expertise Modeling
ExpertFindingServer
AggregatedExpertise
Model
Gateway(broker)
localExpertise
Model
Server1
localExpertise
Model
Server1
localExpertise
Model
Servern
LAN
Background
Alternatives
DEMOIR
First appr.
Summary
July 13, 2000 TWIST 2000 Yimam & Kobsa
Aggregated Expertise Modeling
ExpertFindingServer
AggregatedExpertise
Model
Gateway(broker)
localExpertise
Model
Server1
localExpertise
Model
Server1
localExpertise
Model
Servern
Gateway(broker)
Server1
Server1
Servern
CentralExpertiseModel
LAN
LAN
Background
Alternatives
DEMOIR
First appr.
Summary
July 13, 2000 TWIST 2000 Yimam & Kobsa
Analysis
Factor Agents AggregatedLocality (processing/load distribution) Facilitated limited
Mining personal resources easier difficult
Personal Privacy preservation At the hand of the expert At the hand of maintainer
No single point of failure Single point of failure(backup mechanisms needed)
ScalabilityRobustness (in face offailure)
Extendability Easy by adding new agents Depends on design
Central administration No Yes
Experts feeling in control Yes Mostly not
Background
Alternatives
DEMOIR
First appr.
Summary
July 13, 2000 TWIST 2000 Yimam & Kobsa
Analysis (contd.)
Factor Agents AggregatedOrganization-wide access to expertiseinfo.
Limited (e.g. to expertnetwork, etc.)
facilitated
Multi-purpose/optimal utilization ofexpertise info (analysis, visualization,browsing, etc.)
Limited facilitated
Sources of expertise evidence mined Mostly limited to personalresources
Organizational resources(repositories, databases,Web/Internet, etc.)
Query Performance (scalability) Low (due to the need toconsult many agents)
High due to single location ofinformation
Knowledge-based/statisticaltechniques support
poor Good
Coordination overhead High (e.g. getting agentsfind and interact with oneanother)
low
Background
Alternatives
DEMOIR
First appr.
Summary
July 13, 2000 TWIST 2000 Yimam & Kobsa
Hybrid Approach
• Combine distributed agents with centralized expertise model server - “local-central” approach
• How ?1. Decentralized + centralized Expertise modeling
• Lightweight personal agents for personal sources• Configurable gatherers for organizational resources
2. Centralized (but “distributable”) expertise information server3. Decentralized Exploitation of expertise information (through
clients)
Background
Alternatives
DEMOIR
First appr.
Summary
July 13, 2000 TWIST 2000 Yimam & Kobsa
DEMOIR - A Hybrid Architecture
Organizational InformationResources
Expertise-indicator
Source Gatherers
Source TypeIdentifier
SourceWrapper2
SourceWrapper1
SourceWrappern
...
EISM
Ontology,Organizationalstructure, etc.
AggregatedExpertise
Model
ExpertModels
RemoteExpert Details
API ClientsFusers
Expertise Information Space
Background
Alternatives
DEMOIR
First appr.
Summary
July 13, 2000 TWIST 2000 Yimam & Kobsa
DEMOIR - A Hybrid Architecture
Gathering(decentralize
dCentralized)
Modeling(decentralized/centralized)
Exploitation(decentralized
)
Background
Alternatives
DEMOIR
First appr.
Organizational InformationResources
Expertise-indicator
Source Gatherers
Source TypeIdentifier
SourceWrapper2
SourceWrapper1
SourceWrappern
...
EISM
Ontology,Organizationalstructure, etc.
AggregatedExpertise
Model
ExpertModels
RemoteExpert Details
API ClientsFusers
Expertise Information Space
Summary
July 13, 2000 TWIST 2000 Yimam & Kobsa
Summary/Observation
• Centralized and decentralized options have their advantages and disadvantages.
• Many problem domains involve both “centralizable” and “decentralizable” tasks
Challenges:+ isolating such tasks and identifying the tradeoffs b/n
centralizing and decentralizing their operations+ If both approaches are used, how to get them work together
Background
Alternatives
DEMOIR
Summary
First appr.
July 13, 2000 TWIST 2000 Yimam & Kobsa
Summary/Observation
• Centralization/decentralization is only one dimension of a system’s architecture. Relate to:
+ size/complexity of system (e.g. number of different parts, dynamism of their interaction, etc.)
+ heterogeneity of data and their sources+ accessibility (e.g. permissions/privacy constraints, manner of
use)+ communication patterns among components
keep these in mind and analyze how they affect centralization/decentralization decision.
Background
Alternatives
DEMOIR
Summary
First appr.
July 13, 2000 TWIST 2000 Yimam & Kobsa
Summary/ObservationWhat we did (in retrospect):
1. Identify system requirements/tasks
2. Identify and analyze centralized and decentralized alternatives of performing identified tasks
thereby identify and evaluate general centralization and decentralization factors in the problem domain.
3. Specify optimum system components as well as architecture (i.e. trying to achieve advantages and avoid disadvantages of alternatives)
aim at flexibility to allow varying degrees of centralization and/or decentralization to suit different deployment environments.
Background
Alternatives
DEMOIR
Summary
First appr.