Personalised Search on the World Wide Web
Presented by:Team Grape
About
Team Grape:Jin Wu
Kewei DuanLinh Duy ToMiaolai Han
Takazumi Matsumoto
The Paper:Personalized Search on
the World Wide Web
Alessandro MicarelliFabio GasparettiFilippo Sciarrone
Susan Gauch
The interactive stuff:MOT lesson
Grapple lessons: Text only, Depth first
Overview
1. Introduction2. A Short Overview on Personalised Search3. Contextualised Search4. Personalisation Based on Search Histories5. Personalisation Based on Rich Representations of User
Needs6. Collaborative Search Engines7. Adaptive Result Clustering8. Hyperlink-Based Personalisation9. Combined Approaches to Personalisation10. Conclusions
Introduction
Personalisation“adapting the results according to each user’s
information needs”(Micarelli et al., 2007, p. 195)
• Searching the WWW• Dealing with the information overload• Limitations of traditional search engines• Information access paradigms:– Searching by surfing (hyperlink directories)– Searching by query (Information Retrieval)– Recommendation (suggested items)
Content and Collaborative-based Personalisation
• Originally: information retrieval• Content-based: – Consider individuals - mostly used– Polysemy & synonymy leads to vocabulary problem
→ irrelevant information• Collaborative-based: – Consider models of different users– User similarity → similar information needs– Social navigation– Not employed in search engines
User Modelling in Personalised Systems
• User modelling/profiling techniques:– Track visited pages & search history → important feature
learned → more relevant information– Simplest cases: registration form or questionnaire– More complex cases: user model consists of a dynamic
information structure• Examples:
– Google Alert: explicit approach & routing query → limited– Google Personalized Search: deliver customised search based
on user profile• User modelling components affect search in 3 distinct
phases:– Part of retrieval process– Re-ranking– Query modification
Source of Personalisation
• Data mining & machine learning• Relevant feedback & query expansion– Explicit relevant feedback– Implicit relevant feedback
• Further sources: desktop search systems
An Overview on Personalisation Approaches
• Current context: based on implicit feedback using client-based software
• Search History: – Limited to web search history– Done during retrieval process → fast response
• Rich user models: explicit feedback → build rich representation of user needs
• Collaborative approach: relevant resources based on previous ratings by user with similar tastes & preferences
• Result clustering: results grouped into clusters, each related to same topic
• Hyper textual data: include additional factors in ranking algorithm
Contextual Search
• A new approach for search • The information system proactively suggests
information based on a person’s working context
• Just-in-Time IR (JITIR)• Rhodes
JITIR
• Monitors the user’s actions• Non-intrusive• Automatically identify relevant information• Retrieve resources automatically
Based on Agents
• Remembrance Agent • Margin Notes Agent• Jimminy Agent
Personalisation Based on Search Histories
Visa
Citizenship
Travel
Credit Card
Flight
Online Approaches
• Capture history information as soon as they are available, affecting user models and providing personalised results taking into consideration the last interactions of the user
• Two different types of information are collected: – submitted queries– snippets
Offline Approaches
• Exploit history information in a distinct pre-processing step, usually analysing relationships between queries and documents visited by users
• CubeSVD Algorithm based on the click-through algorithm
• Time-consuming
Personalisation Based on Rich Representations of User Needs
Three prototypesifWeb, Wifs, InfoWeb
• Based on complex representations of user needs (user models)
• Built using explicit user feedback on results• Based on frames and semantic networks (AI)
ifWeb
• User model-based intelligent agent• Weighted semantic network for user profile• Autonomous focused crawling to find related
documents based on previously identified documents
• Updates user profile using user feedback• Reduces the weight of unused concepts (rent)
Wifs
• Content-based approach• Filters HTML and text documents from AltaVista,
reordering links based on UM• Frame-based user model structure
A frame has slots which contains terms (topics), associated with other terms (co-keywords), forming a semantic network
• The terms are stored in a Terms DataBase that is created beforehand (by experts)
• Instead of traditional IR, the relevance of a document is calculated from the occurrence and relevance of terms in the document
Wifs
• Content-based approach• Frame-based user model structure
A frame has slots which contains terms (topics), associated with other terms (co-keywords), forming a semantic network
• The terms are stored in a Terms DataBase that is created beforehand (by experts)
• Filters HTML and text documents from AltaVista, reordering links based on UM
• Instead of traditional IR, the relevance of a document is calculated from the occurrence and relevance of terms in the documentRepresentations of the User model (a) and Document model (b)
(From Micarelli et al., 2007)
InfoWeb• Content-based approach• Adaptive retrieval of documents in digital
libraries, based on Vector Space (IR)• Stereotype knowledge base
Contains most significant documents for a specific category of user (domain), created beforehand (by an expert)
• k-means clustering on document collection beforehandEach cluster is seeded by a representative document for each class of user
• User model starts as a stereotype, evolves based on feedback
Collaborative Search Engine
• ‘SearchParty’ module– Social filtering– Stores user queries and the results users clicked
• Knowledge Sea– Social adaptive navigation system– Exploits both traditional IR and social navigation
approaches– Results represented by colour lightness
Collaborative Search Engine• Calculate similarity measures among user needs– Identified by queries, selected resources– Two queries might contain no common terms but
returns similar results– E.g. ‘PDA’ and ‘handheld computer’
• Statistical model– Based on the probability a page was selected for a
given query– Focus on relative frequency instead of content-
analysis techniques
Collaborative Search Engine
• Compass Filter– Based on web communities– Pre-processing the web structure– If user frequently visit a community, the results in
the same community are boosted
Adaptive Result Clustering
• Traditional Search Engines– Rank the list by similarity of query and page– Might take a long time– Important that users clearly describe what they
are looking for
• Organise the results– By grouping pages into folders and sub folders– On a graphical interactive map
Adaptive Result Clustering
• Clustering– Query process needs to be fast– Usually performed after retrieval of query results– Does not require pre-defined categories– Provides concise and accurate descriptions
• Further clustering systems– SnakeT– Scatter / Gather
Main algorithms:
•PageRank: PR value•HubFinder: hub value•HubRank: PR value & hub value
Hyperlink-Based Personalisation
Combined Approaches to Personalisation
Perform personalisation using multiple adaptive approaches•Outride:
Browsing history & current context•infoFACTORY:
Integrate web tools & services
Outride
Outride includes:• ContextualisaionInterrelated conditions that occur within an activity
• IndividualisationCharacteristics that distinguish an individual
infoFACTORY
• A large set of integrated web tools and services that are able to evaluate and classify documents retrieved following a user profile
• New• Has potential• Interesting
Conclusions
• Information is crucial to users• Need to filter and personalise resources to deal with
information overload successfully• Increases search engine accuracy and reduces time
wasted sorting through irrelevant results• Can be extended e.g. targeted advertising• Some systems already in use, others under
development (e.g. Semantic Web)• Future directions: – Predicting future user behaviour (plan-recognition)– Language semantic analysis (Natural Language Processing)
Thanks for listening
Any questions?