Transcript Agent and Data Mining: Mutual Enhancement by Integration

Integration of Agent and Data
Mining
Longbing Cao
University of Technology, Sydney
Content
• Introduction
• Agents can enrich data mining
• Data mining can improve agents
• Ontology-based integration of agents and
data mining
• Demo
• Conclusions and directions
INTRODUCTION
Data mining & multiagent research
group at UTS
• Cross disciplinary researchers interacting at the group
• Integrated research of data mining and multi-agent
system
– http://datamining.it.uts.edu.au
• Real-world applications of the integration
– Capital markets
– F-Trade
Agents as a new computing
paradigm for complex problems
• Strengths
– Analyze and understand complex systems
– Deal with nonfunctional requirements
– Handle social complexity such as distribution,
dynamics, interaction, evolution, self-organization
– Build flexible infrastructure
• Weaknesses
– Lack machine learning capability
– Lack in-depth analytics
– Lack knowledge representation
Data mining and knowledge
discovery as an effective tool for indepth analysis
• Strengths
– Deep data analysis
– Deep knowledge discovery
• Weaknesses
– Nothing related to system infrastructure
– Deal with social complexity such as
distribution, dynamics
Bilateral enhancement of agents
and data mining by the integration
• Agents can enrich data mining
• Data mining can improve agents
• Mutual enhancement: integration between
data mining and multi-agent system
AGENTS can ENRICH DATA MINING
Building agent-based data mining
systems
• Agent-based data mining system
– F-Trade
• Agent-based distributed data mining
system
– Agent-based distributed data mining systems,
such as BODHI, PADMA, JAM, Papyrus
• Agents for multiple data source mining
• Agents for web mining
Data mining models as agents
• Intelligent data mining agents – modeling
data mining algorithms as agents
• Data mining model integrator – integrating
data mining algorithms
• Data mining model planner – smartly
managing data mining algorithms
• Data mining model recommender –
recommending appropriate algorithms
Agent-based mediation and
management of distributed and
large-scale data sources
• Data gateway agents for connecting data
sources
• Distributed data preprocessor agent
• Data integrator agents for data integration
• Agents for data clustering
• Agents for ensemble mining in distributed
data
• Agents for data sampling and assumption
User and interaction agents for
data mining
• Human agent interaction for data mining
• Agents for interactive mining
• Agents in human-guided mining
• Domain knowledge management using
agents
• User agents for preparing mining reports
• Agents for circulating mining results
Case study 1 -- F-Trade
Users/CMCRC/Instituations
(Anybody,anytime,anywhere,
from MAS & KDD & Finance)
Applications developers
KDD Researchers
(Frequent and abnormal patterns
discovery, optimization of
trading strategies, correlation
analysis)
Aims/Motivations:
Network
(Internet &
LAN)
AAMAS Researchers
(OCAS, AOSE, OADI, OSOAD)
(Services for system
components,algorithm and
multiple data sources)
F-Trade
(open automated
enterprise services,
and personalized
services)
Data Sources
(Diff. Providers: AC3, HK
market, CSFB, etc.
Diff. Formats: FAV, ODBC,
JDBC, OLEDB, etc.
)
• Research Service Provider for AAMAS and data mining
• Integrated Infrastructure for Financial Trading and Mining Support
Case study 1 -- F-Trade
System infrastructure
Case study 1 -- F-Trade
F-TRADE: Financial Trading Rules Automated Development & Evaluation
Case study 1 -- F-Trade
Algorithm
as an agent
Case study 1 -- F-Trade
AgentService
RegisterAlgorithm(algoname;inputlist;inputconstraint;outputlist;outputconstraint;)
Description:
This agent service involves accepting registration application submitted by role PluginPerson,
checking validity of attribute items, creating name and directory of the algorithm, and generating
universal agent identifier and unique algorithm id.
Role: PluginPerson
Pre-conditions:
-A request of registering an algorithm has been activated by protocol
SubmitAlgoPluginRequest
-A knowledge base storing rules for agent and service naming and directory
Type: algorithm.[datamining/tradingsignal]
Location: algo.[algorithmname]
Inputs: inputlist
InputConstraints: inputconstraint[;]
Outputs: outputlist
OutputConstraints: outputconstraint[;]
Activities: Register the algorithm
Permissions:
-Read supplied knowledge base storing algorithm agent ontologies
-Read supplied algorithm base storing algorithm information
Post-conditions:
-Generate unique agent identifier, naming, and locator for the algorithm agent
-Generate unique algorithm id
Exceptions:
-Cannot find target algorithm
-There are invalid format existing in the input attributes
Agent plugand-play
Case study 1 -- F-Trade
Agent for
multiple data
sources
management
Case study 1 -- F-Trade
Agent for
reporting
Visual Reports
Point Reports
Transactions Reports
Summary Reports
Input Reports
Case study 2 – agent-based WEKA
Case study 3 – ensemble
DATA MINING can IMPROVE AGENTS
Data mining-driven multiagent
learning
• DM-driven learning in MAS
– Coordination learning
– Individual learning
– Group/collective learning
– Distributed learning
– Online/offline learning
Data mining-driven evolution and
adaptation in MAS
• Evolution of MAS based on hidden rules,
so mine these rules and fill into the agent
knowledge base for designing evolutionary
agent systems
• Adaptive capability mining for enhancing
agent’s adaptation
• Self-organization rule mining
Data mining for agent
communication, planning and
dispatching
• Cluster and classification
• Class/segment-based communication
• Class-based planning and dispatching
DM-based User modeling
• Modeling user behavior from DM
– Game player modeling
– Trader’s behavior modeling
– Trader’s role modeling
• User-agent interaction based on user
modeling
– Trader agents’ interface design
– Trader-agent interaction rule design
DM-based User servicing
• DM-based agents for serving users
– Visualization mining for reporting
– Customer-relationship management for
customer care
• DM-based recommender agents
– Stock recommender
– In-depth rule recommender
– Trading rule-stock recommender
Case study - learning
• Agent learning via machine learning
– Reinforcement learning
– Evolutionary multiobjective methods
– Evolutionary algorithm
– Markov decision process
– Temporal difference method
Case study – user modeling
• Trader’s behavior modeling
• Trader’s role modeling
– Market order
– Limit order
MarketOrder
LargeMarketOrder
January
February
Large market orders analysis
Case study - servicing
• Pairs trading
– Mining correlated stock pairs
– Correlated stock miner agent
– Stock pairs recommender
– Pairs trading strategy solution
Case study - servicing
• Optimized rules
– Mining in-depth rules
– In-depth rule miner agent
– User interface agent
– Optimized rules recommender
– Optimized trading strategy solution
Case study - servicing
• Rule-stock pairs
– Mining rule-stock pairs
– Rule-stock pair mining agent
– User interface agent
– Rule-stock pair recommender
– Trading strategy solution
Return on investment
ONTOLOGY-BASED INTEGRATION OF
AGENTS AND DATA MINING
Ontology for domain
understanding and interaction
• Domain ontology for understanding the
domain problems
• Problem-solving ontology
• Task ontology
• Method ontology
Ontology for knowledge
management
• Ontology for organizing agent systems
• Ontology for organizing mining algorithms
• Ontology for user interaction
• Managing domain ontology/task
ontology/problem-solving
ontology/method ontology
Ontology-based system
architecture
• Multi-domain ontological space
– Related problem domains
– Agent ontology domain
– Data mining ontology domain
• Hybrid ontology structure for
organizing ontologies crossing
multiple domains
Ontological engineering for the
integration
• Ontology namespace
• Ontology mapping structure
• Semantic rules for ontology mapping
• Ontology transformation
• Ontology query
• Ontology search and discovery
Business Profiles
Task View
Domain Ontology
Task Ontology
Businesslogic View
BL Ontology
Method View
Problemsolving System
Resource View
Method Ontology
Resource Ontology
PS Ontology
DO-to-PSO linkage
Internal PSO linkage
Stock
FinancialOrder
f
Limit Order
Market Order
...
Stop Order
in
s
pa
ta
rtnc
of
eof
su
bc
la
ss
-o
OrderOperation
Amend
Price
Enter
Dealer
Trade
Delete
Date
Time
Cantr
Volume
Algorithm Agent
Input
ontologies
...
...
Output
ontologies
Resource
ontologies
...
...
Knowledge
ontologies
...
….
M1 ? M2
N1 = N2 || N1  N2
N1  N2
M1  M2
Equivalence,
similarity
Synonyms, encoding,
conventions,
paradigms,
scaling
M1  M2
Scope, coverage,
granularity
Generalization,
specialization
 =M1M2
Naming conflict,
homonymy
Disjointness,
antonyms
M1M2
<min(M1,
M2)
Scope, coverage,
granularity
Overlapping
M1  M2
Naming, encoding
Instantiation
-  (part_of (A, B)  part_of (B, C))  part_of (A, C)
-  (substitute_to (A, B)  substitute_to (B, C)) 
substitute_to (A, C)
Ontology 1
Root Concept
- fixed
- resident
- id
- local_fee
- remote_fee
- IP
- business
Ontology 2
Root Concept
- home
- local_call
- intraprovince
- interprovince
- international
- Hongkong
- Taiwan
- Macau
- IP
- business
conceptto-concept
attributeto-concept
attributeto-attribute
1 attribute-tom*attribute
Rule 4.
-  (A AND B),  B ::= substitute_to(A, B)
 A OR B, the resulting output is A or B
Rule 5.
-  (A AND B),  B ::= disjoint_to(A, B)
 A AND B, the resulting output is A and B
DEMO
CONCLUSIONS and DIRECTIONS