Foxtrot demo 1.10.2002 Foxtrot recommender system Demonstration Stuart E. Middleton David C. De Roure, Nigel R. Shadbolt Intelligence, Agents and Multimedia.

Foxtrot demo 1.10.2002

Foxtrot recommender systemDemonstration

Stuart E. MiddletonDavid C. De Roure, Nigel R. Shadbolt

Intelligence, Agents and Multimedia Research GroupDept of Electronics and Computer Science

University of SouthamptonUnited Kingdom

Email: [email protected]

• Problem domain• Ontological user profiles• Foxtrot approach• Demonstration• Future work

Foxtrot recommender systemDemonstration

Foxtrot demo 1.10.2002

• Problem domain

Recommender systemsNo need to specify explicit keywordsLearns the type of things you wantAutomatically looks for themRecommends relevant things when they are found

A real world problem domainOn-line research paper recommendation for researchersUniversity of Southampton staff and students

Information overload on the WWWToo many sites and pages to browseSearch engines need an explicit search query

Foxtrot recommender systemDemonstration

Foxtrot demo 1.10.2002

• Ontological user profiles

Training set for each class5-10 labelled examples per classExamples shared between users

Profile inferenceIs-a relationships used to infer interests not seen directlyInference used to improve profile accuracy

Profile feedbackProfile visualization allows profile feedbackFeedback used to improve profile accuracy

User profiles represented using an ontologyClasses represent research topicsOntology contains is-a relationships between classesUser profiles hold classes and current interest values

Foxtrot recommender systemDemonstration

Foxtrot demo 1.10.2002

Foxtrot recommender systemDemonstration

Foxtrot demo 1.10.2002

ResearchersResearchersResearchersResearchers

ClassifierClassifier

RecommenderRecommender

AgentAgentAgentAgentAIAI

AIAIAIAI

HCIHCI

Research paper databaseResearch paper databaseResearch paper databaseResearch paper database

WebWebWebWeb

WebWeb

World-WideWeb

World-WideWeb

Web ProxyWeb Proxy

ProfilerProfiler

• Foxtrot approach

• Foxtrot approach

Shared research paper databaseDatabase grows as users browse the internet

ClassifierBoosted k-Nearest Neighbour classifier

ProfilerBrowsed papers and explicit feedback indicate interestTime-decay and inference used to find interests

RecommenderPearson-r correlation finds similar peopleRecommended papers are those read by similar people

Unobtrusive monitoringWeb proxy used to monitor web browsing

Foxtrot recommender systemDemonstration

Foxtrot demo 1.10.2002

•D

emo

nst

rati

on

Foxtrot recommender systemDemonstration

Foxtrot demo 1.10.2002

Recommendation pageRecommendations appear as search results when you open the web page

Foxtrot recommender systemDemonstration

Foxtrot demo 1.10.2002

•D

emo

nst

rati

on

SearchingSearches are made by entering a search query into the edit boxes.

•D

emo

nst

rati

on

Foxtrot recommender systemDemonstration

Foxtrot demo 1.10.2002

SearchingSearches are made by entering a search query into the edit boxes. For instance, “speech” can be entered into the title search box.

Foxtrot recommender systemDemonstration

Foxtrot demo 1.10.2002

•D

emo

nst

rati

on

SearchingSearches are made by entering a search query into the edit boxes. For instance, “speech” can be entered into the title search box. A search class can also be entered by clicking on the … menu button.

Foxtrot recommender systemDemonstration

Foxtrot demo 1.10.2002

•D

emo

nst

rati

on

SearchingSearches are made by entering a search query into the edit boxes. For instance, “speech” can be entered into the title search box. A search class can also be entered by clicking on the … menu button.

Foxtrot recommender systemDemonstration

Foxtrot demo 1.10.2002

•D

emo

nst

rati

on

SearchingSearches are made by entering a search query into the edit boxes. For instance, “speech” can be entered into the title search box. A search class can also be entered by clicking on the … menu button.

Foxtrot recommender systemDemonstration

Foxtrot demo 1.10.2002

•D

emo

nst

rati

on

SearchingSearches are made by entering a search query into the edit boxes. For instance, “speech” can be entered into the title search box. A search class can also be entered by clicking on the … menu button.

Foxtrot recommender systemDemonstration

Foxtrot demo 1.10.2002

•D

emo

nst

rati

on

SearchingNow a search query has been entered, the search button can be clicked. This starts the search.

Foxtrot recommender systemDemonstration

Foxtrot demo 1.10.2002

•D

emo

nst

rati

on

SearchingOnce complete, the search results are displayed in the area below the search query. The recommendations are overwritten by the search result.

Foxtrot recommender systemDemonstration

Foxtrot demo 1.10.2002

•D

emo

nst

rati

on

FeedbackResearch papers can be opened by clicking on the green hyperlink.

Foxtrot recommender systemDemonstration

Foxtrot demo 1.10.2002

•D

emo

nst

rati

on

FeedbackResearch papers can be opened by clicking on the green hyperlink. PS, PDF, HTML and compressed versions are supported.

Foxtrot recommender systemDemonstration

Foxtrot demo 1.10.2002

•D

emo

nst

rati

on

FeedbackFeedback on individual papers can be provided via the radio buttons next to a paper.

Foxtrot recommender systemDemonstration

Foxtrot demo 1.10.2002

•D

emo

nst

rati

on

FeedbackFeedback on individual papers can be provided via the radio buttons next to a paper.

Foxtrot recommender systemDemonstration

Foxtrot demo 1.10.2002

•D

emo

nst

rati

on

Profile visualizationProfiles are visualized as a time/interest graph.

Foxtrot recommender systemDemonstration

Foxtrot demo 1.10.2002

•D

emo

nst

rati

on

Profile visualizationProfiles are visualized as a time/interest graph. Classes can be turned on and off.

Foxtrot recommender systemDemonstration

Foxtrot demo 1.10.2002

•D

emo

nst

rati

on

Profile visualizationUsers can draw interest bars onto the graph to indicate interest.

Foxtrot recommender systemDemonstration

Foxtrot demo 1.10.2002

•D

emo

nst

rati

on

Profile visualizationUsers can draw interest bars onto the graph to indicate interest.

Foxtrot recommender systemDemonstration

Foxtrot demo 1.10.2002

•D

emo

nst

rati

on

Profile visualizationUsers can draw interest bars onto the graph to indicate interest.

Foxtrot recommender systemDemonstration

Foxtrot demo 1.10.2002

•D

emo

nst

rati

on

Profile visualizationUsers can draw interest bars onto the graph to indicate interest.

Foxtrot recommender systemDemonstration

Foxtrot demo 1.10.2002

•D

emo

nst

rati

on

Profile visualizationUsers can draw interest bars onto the graph to indicate interest.

Foxtrot recommender systemDemonstration

Foxtrot demo 1.10.2002

•D

emo

nst

rati

on

Profile visualizationUsers can draw interest bars onto the graph to indicate interest.

Foxtrot recommender systemDemonstration

Foxtrot demo 1.10.2002

•D

emo

nst

rati

on

Profile visualizationUsers can draw interest bars onto the graph to indicate interest. Once drawn, bars can be moved and re-sized as required. Continued interest goes off the end of the graph.

• Future work

Task profileTask profiles would allow more than just a general profileTask analysis is a hard problem however

Agent metaphorA multi-agent-system could buy/sell ontological knowledgeAgents could trade personal information

Improved ontologyMore relationships than just is-a linksProjects, related areas, technology links…

Foxtrot recommender systemDemonstration

Foxtrot demo 1.10.2002