Implicit

1
Implicit

• An Agent-Based Recommendation System for Web
  Search
• Presented by
• Shaun McQuaker
• shaun_mcquaker_at_hotmail.com
• Presentation based on paper
• Implicit An Agent-Based Recommendation System
• Alexander Birukov, Enrico Blanzieri, and Paolo
  Giorgini
• Department of Information and
• Communication Technology
• University of Trento Italy

2
Overview

• Problem Definition
• Implicit Culture and SICS
• Implicit System Structure
• Experimental Results
• Related Work
• Conclusions

3
Problem Definition

• Increasing amount of web content
• On July 2004 there were 285,139,107 hosts on the
  Internet
• Finding relevant information is a hard task
• Approximately 56.3 of the Internet users search
  the web at least once per day
• 33 rarely look at second page of results

4
Problem Solutions

• Authority-based search engines
• Recommendation systems
• Systems that deal with the content of the web
  pages
• Systems that use a collaborative approach
• Agents and multi-agent systems
• Software agent that assists its user

5
Solution Implicit

• Agent-based recommendation system
• Intended to improve web search of a community of
  people with similar interests
• Based on the concept of Implicit Culture

6
Implicit Culture Motivation

• An agent interacting in a new environment
• Humans experience culture shock
• New user of a system, where is the printer?
• Solutions
• Just ask someone
• Represent relevant knowledge and give it to the
  agent
• Agent with observational and learning capabilities

7
Implicit Culture basic definitions (1)

• Let P be a set of agents, O a set of objects, A
  a set of actions. We define
• Environment eÍ P ÈO
• Scene as the pair , where B Í e, and A Í A
• Situation as , where a Î P and s is a
  scene
• Executed situated action as the action executed
  in given situation.

8
Implicit Culture basic definitions (2)

• Random variable ha,t that describes the action
  that the agent a executes at the time t
• Expected action as the expected value of ha,t ,
  E(ha,t )
• Situated expected action as the expected value of
  ha,t given a situation E(ha,t )
• Cultural constraint theory for a group GÍ P, as
  a theory on the situated expected actions of the
  agents of G
• Cultural action w.r.t. G, as an executed action
  that satisfies a cultural constraint theory for G

9
Implicit Culture Solution

• Provides a method where new agents can behave
  similarly to existing agents.
• Control the environment
• Change environment to express implicit knowledge
  of the agent.
• Directory Finder for services
• Existing agents may have optimized behaviour thus
  a new agent entering performs in an optimal manner

10
Implicit Culture System

• Has goal of achieving implicit culture
• Achieves it by
• Building validated cultural constraints from
  observations of situated actions
• Presenting scenes to agent such that their
  actions satisfy this constraint
• Directory recommends service that best fits
  request

11
SICS

• Systems for Implicit Culture Support
• Goal produce Implicit Culture phenomenon
• Architecture
• Observer, stores executed situated actions done
  by agents in the group
• Inductive module, uses actions to produce a
  cultural constraint theory
• Composer, using theory and actions to manipulate
  scenes faced by the agents

12
SICS Overview
S0
Observer stores in a data base the situated
executed actions of the agents of G.
Inductive Module
S
Inductive Module using the data from the DB
induces a cultural constraint theory S. Can use
clustering techniques, a priori learning.
DB
Observer
Composer

• Composer proposes to a group G a set of scenes
  such that the expected situated actions satisfies
  S.
• Two sub-components
• Cultural Actions Finder
• Scene Producer

13
SICS Composer

• Cultural Actions Finder
• Takes as input the theory S and executed situated
  actions of G and produces cultural actions that
  satisfy S.
• Scenes Producer
• Takes one of the cultural actions produced by CAF
  and executed situated actions of G, and produces
  scenes such that the expected situated action is
  the cultural action.
• Directory Finder Example
• Cultural theory request(x,DF,s) inform(DF,x,y)
  - request(x,y,s)
• Agent in G makes request(x,DF,s)
• CAF produces request(x,y,s)
• SP proposes y to provide service s, thus
  inform(DF,x,y)
• It is now expected that the agent (x) will chose
  y to provide service s

14
Implicit

• Implemented in JADE
• SICS module incorporated in agent to produce
  recommendations
• Agents communicate with outside search source,
  Google.
• Agents are collaborative
• Send messages between each other

15
Implicit Messages

• Query Message
• Information about user query or agent query
• Reply Message
• Contains recommended link or ID of another agent
• Feedback Message
• Contains accepted/reject links or agent Ids.

16
Implicit Usage (1)
17
Implicit Usage (2)
18
Experimental Purpose

• Understand how the insertion of a new member into
  the community affects the relevance, in terms of
  precision and recall, of the links that are
  produced by SICS.
• Also after a certain number of interactions, will
  personal agents be able to propose links accepted
  in previous searches?

19
Experimental Measurements

• Link is relevant to a particular keyword if
  probability of acceptance is above a certain
  threshold (0.1)
• Precision is the number of suggested relevant
  links to total number of suggested links.
• Recall is the ratio of proposed relevant links to
  the total number of relevant links

20
User Interaction

• User profiles replace user interaction.
• 10x10 matrix of keywords vs. rank
• Values denote probability that link is relevant
• Assume all users are similar, thus personal
  profile is derived from a base profile.
• User accepts only one link, other suggested links
  are rejected.
• Datasets replace queries to Google.

21
Sample User Profile
22
Experiment Details

• SICS module suggests links for keywords after
  observing user acceptance.
• Suggestions are given by other agents based on
  their user profiles
• User will accept or reject suggest links.
• Feedback is sent
• Relevant/Irrelevant links are enumerated
• Precision and Recall are calculated

23
Experimental Results

• More agents more relevant link suggestions
• Agents with same profile in community of 4 or 5
  agents performed on average better across all
  tests
• Agents have determined which link is the most
  relevant given a group of agents with the same
  profile (interests).
• An Implicit Culture has been established

24
Related Work

• InfoSpiders, analyze hyperlinks on current page
  to propose new documents
• Goal-oriented web search
• What to do if my pet is sick?
• Take it to a veterinarian, return closest
  veterinarian office
• Referral Network
• Agents have interest, expertise, neighbours
• Can query, provide answers or referrals
• Ontology to facilitate knowledge sharing

25
Future Work

• Improve composer module by using association
  rules
• Analyze social relations between agents
• Hybrid Referral Network and Implicit Culture
• Using ontologies agents could connect to related
  communities
• Search each community for relevant links.

26
Conclusions

• Agents interacting in Implicit Culture allow
  better recommendations to be made
• Prevents new agents from searching from scratch
• Uses power of other agents as well as a search
  engine
• Process is transparent to user

27
References

• Birukov Alexander, Blanzieri Enrico, Giorgini
  Paolo (2005), Implicit An Agent-Based
  Recommendation System, Department of Information
  and Communication Technology, University of
  Trento, Italy.
• Blanzieri Enrico, Giorgini Paolo (2000), From
  Collaborative Filtering to Implicit Culture a
  general agent-based framework, ITC-IRST Trento,
  Italy, University of Trento, Italy.
• Lin Weiyang, Alvarez A. Sergio, Ruiz Carolina
  (2001), Efficient Adaptive-Support Association
  Rule Mining for Recommender Systems, Microsoft
  Corporation, Department of Computer Science,
  Boston College, Department of Computer Science,
  Worcester Polytechnic Institute.