Pricing Information Goods in an Agentbased Information Filtering System

1
Pricing Information Goods in an Agent-based
Information Filtering System

• Laura Maria Andreescu

Christos Tryfonopoulos MPII Saarbrücken
David Midgley INSEAD Fontainebleau
2
Outline

• Motivation
• Background
• ABIS
• Publisher selection and ranking
• Experiments
• Conclusions

3
Information Retrieval Scenario
simple query financial crisis
Agent
4
Information Filtering (IF) Scenario
continuous query financial crisis
Agent
5
Applications of IF

• News dissemination
• Sharing educational material (e.g., Edutella)
• Information alert for digital libraries
• Information alert for electronic marketplaces
• Stock market updates

6
Information Filtering System Architecture
Overview
Publisher1
Publisher2
meta-data
middleware
request for meta-data
meta-data
Publisher3
continuous query
Subscriber1
notify
meta-data e.g. docs available, publication
rate
7
Exact IF vs. Approximate IF

• Exact IF
• Subscribers are interested in notifications for
  all publications
• Approximate IF
• Subscribers are NOT interested in all matching
  publications
• Trade recall for scalability

8
ABIS Multi-agent Arhitecture
Agent Network
Directory Service
Subscription Service
Publication Service
9
ABIS Example
finance A1,A3,A4
crisis A1,A2,A5
financial crisis A1,A2,A3,A4, A5
finance A1,A3,A4
Agent 5
Agent 6
Agent 2
Agent 4
crisis A1,A2,A5
Agent 1
Agent 3
10
ABIS Example
Agent 5
Agent 6
finance crisis
Agent 2
no notification
finance crisis
Agent 4
Agent 1
Agent 3
11
ABIS

• information has a price
• 3 classes of agents
• choosing best top-k publishers that would monitor
  his query

12
Publisher selection

• Given a query q Which agents are most likely to
  publish documents matching q in the future?
• Subscriber uses the directory service (collecting
  per-term statistics of each query term) to
  compute publisher scores

Publisher Score
Information Quality
Price

• Publisher score used for ranking
• Scores are periodically recomputed, queries
  repositioned

13
Information Quality
Information Quality
Resource Selection
Agent Behaviour Prediction
Resource Selection

• identifies authorities
• based on IR techniques term/document
  frequencies,
• 
  collection sizes...

• how likely is a agent to publish documents
• of interest in the future
• based on time series analysis on IR metrics

Agent Behaviour Prediction
14
Experimental Evaluation

• Setup
• 100 agents containing
• 10 categories Music, Finance, Arts, Sports...
• each agent initial collection of 300 documents
• 15 random documents
• 10 not categorized
• 75 documents from a single category
• 10 agents specializing in each category
• 30 continuous queries
• comparison with MAPS (Minerva Approximate
  Publish/Subscribe System)

15
Prices
Random Prices
Prices Correlated Strictly with Quality
Spearman footrule metric adaptation
Prices Partly Correlated with Quality 1
Prices Partly Correlated with Quality 2
price
quality
16
Publishing behaviour
Recall
Consistent Publishing
Change in Publishing
price
quality
17
Conclusions

• Contributions
• Define an agent based arhitecture for approximate
  information filtering
• Proposal publisher ranking technique
• resource selection
• predicted behaviour
• cost of information
• Future Work
• Money Flow Publisher-Subscriber (in progress)
• Automatic Adjustment of Prices

18
Thank you!