PhD Defense - A Context Management Framework based on Wisdom of Crowds for Social Awareness Applications

A Context Management Framework based on Wisdom of Crowds for Social Awareness applications Adrien JOLY PhD Candidate, supervisor: Prof. Pierre MARET, LaHC CIFRE: Alcatel-Lucent Bell Labs France + INSA-Lyon, LIRIS, UMR5205

Un cadre de Gestion de Contextes fondé sur l’Intelligence Collective pour améliorer l’efficacité des applications de Communication Sociale Adrien JOLY CIFRE: Alcatel-Lucent Bell Labs France + INSA-Lyon, LIRIS, UMR5205 Encadré par: Prof. Pierre MARET (LaHC), Johann Daigremont (ALBLF)

Context — Problem and state-of-the-art Approach — Context-based filtering of social streams Framework — Context tag clouds Evaluation — Perceived relevance Conclusion & future work Agenda of this presentation

Context Approach Framework Evaluation Conclusion A PhD thesis in a moving industrial research lab 2003-2006: INSA Lyon + QUT Brisbane 2007-2008: Ambient Services Group, ALU Research & Innovation “ Study of semantic models in context-aware software” Projects: Context Services, EASY Interactions, IMS-based applications 2008-2010: Social Communications Dept., ALU Bell Labs France “ Context-aware Social Networking” Team projects: Hyperlocal Social Networks, Multimodal Support… Strategic projects: Environment as a Service , One Million Conversations With support from colleagues: Service Infrastructure Research Domain (SIRD), ALU Bell Labs France LIRIS (DRIM), INSA Lyon

Context Approach Framework Evaluation Conclusion Introduction to social networks and information overload “ Aware” user Social updates Activities / Status Updates / Contacts

Context Approach Framework Evaluation Conclusion Introduction to social networks and information overload “ Aware” user Activities / Status Updates / Contacts Social updates Information overload* * aka Cognitive Overflow Syndrome [Lahlou 1997]

Context Approach Framework Evaluation Conclusion Introduction to social networks and information overload Filter “ Aware” user Activities / Status Updates / Contacts Needed Social updates and productive

Context Approach Framework Evaluation Conclusion Problem statement Goal Maintain awareness, reduce information overload Proposition Filter social feeds, recommend most relevant social updates Constraints & requirements Privacy: control in the hands of users Poor context: social updates are short by nature No training is possible: new meaning can emerge anytime

Context Approach Framework Evaluation Conclusion State of the Art: Computer-Supported Collaborative Work Fundamental CSCW concepts [Dourish 1992] Awareness: understanding of the activities of others, which provides a context for your own activity Context : object of collaboration, and the way in which the object is produced used to ensure that individual contributions are relevant to the group’s activity Shared feedback presenting feedback on individual users’ activities within the shared [work]space Similar concept on today’s social networks: newsfeed Link between interaction context and relevance Peripheral awareness / Notification systems CANS [Amelung 2005]: activity-based notifications in Sakai CMS  a broader context must be taken in account Social translucence: making activities visible [Erickson 2000] From Knowledge Management to Knowledge Communities Importance of conversations: e.g. give credit

Context Approach Framework Evaluation Conclusion State of the Art: content-based filtering Content-based filtering… Recommend items which content match users’ preferences/profile … for social matching: Recommend web pages accordingly to evolving profile [Balabanovic 1997] … or similarity of content with manipulated documents’ [Budzik 2000] Extract context features from software-based activities [Dragunov 2005] Multidimensional activity/context model for notes [van Kleek 2009] Evolving user profile, learnt from implicit content rating [My6sense.com]

Context Approach Framework Evaluation Conclusion State of the Art: collaborative filtering Collaborative filtering… Find items that were selected by same users Find people who selected similar items … for social matching “ SoMeONe” [Agosto 2005]: items = topics extracted from web bookmarks Using categories from DMOZ/ODP (static) “ groop.us” [Bielenberg & Zacher 2005] items = tags clusters from Delicious bookmarks Insufficient features from Delicious for recommending relevant items ContactRank [Delalonde 2007] Closed vocabulary is too strict Tags are good features to consider [Hotho 2006, Niwa 2006]

Context Approach Framework Evaluation Conclusion Proposal: context-aware filtering C ollaborative filtering via content [Pazzani 1999] Proposal: Hybrid filtering based on context features Publication : Context-Awareness, the Missing Block of Social Networking , International Journal of Computer Science and Applications 2009. Java Dev I/O PHP Similar contexts Physical, virtual and social sensors Open set of context features User’s activity Recommend social updates

Context Approach Framework Evaluation Conclusion Research issues Previous work Context knowledge useful to determine relevance of shared interactions Context can be extracted from content of users’ tasks Problems: closed vocabularies, learning phases, lack of metadata… Proposed contribution Apply collaborative filtering via content , with crowd-sourced metadata Research questions How to model context knowledge as features for filtering social updates? How does contextual similarity perform as a criteria for recommending relevant social updates, from users’ point of view? How do crowd-sourced metadata compare with content-based features?

Agenda of this presentation Context Approach Framework Evaluation Conclusion

Context Approach Framework Evaluation Conclusion Similarity of context, our hypothesis C A is the context of a user U A sharing a piece of information I A . C X is the context of a user U X that is a potential recipient of this information. Hypothesis: I A is relevant to U X if C A is similar to C X A A = Travel in Asia U A = Alice I A = « Check out my amazing picture ! » A B = Working Java U B = Bob I B = « What database should I use ? » A C = Browsing map U C = Christine I C = « Looking for holiday locations… »

Context Approach Framework Evaluation Conclusion Similarity of context, our hypothesis C A is the context of a user U A sharing a piece of information I A . C X is the context of a user U X that is a potential recipient of this information. Hypothesis: I A is relevant to U X if C A is similar to C X C A = Travel, Asia C C = Travel C B = Java Dev. A A = Travel in Asia U A = Alice A B = Working Java U B = Bob I B = « What database should I use ? » A C = Browsing map U C = Christine I C = « Looking for holiday locations… » Similar context: travel No relevant match for this context I A = « Check out my amazing picture ! »

Context Approach Framework Evaluation Conclusion What is context ? Context [Dey, 2001] : « any information that can be used to characterize the situation of an entity » From physical sensors: From computer-based actions: Location Surrounding people Other sensors Communication history Web browsing history Document history

Context Approach Framework Evaluation Conclusion From physical context sensors to applications – usual approach Context sensors Applications Interpretation Acquisition db Usual representation scheme for context information: Ontology-based / semantic Interoperability Requires ont. modeling Lack of semantic data Scalability issues Publication : Context-Aware Mobile and Ubiquitous Computing for Enhanced Usability: Adaptive Technologies and Applications (IGI Global , 2009) Context Management Framework

Context Approach Framework Evaluation Conclusion From physical, virtual and social context sensors to applications – our approach Context Management Framework Context sensors Social Applications Interpretation Acquisition db Proposed representation scheme for context information: Contextual tag clouds Machine-interpretable Crowd-sourced model Easy to edit Emergent meaning Updates Paris Notre-Dame Café Cloudy Crowded Sitting with:Pierre

Context Approach Framework Evaluation Conclusion Our approach, the “Big Picture” Christine’s contextual cloud: McDonalds Chatelet DriveIn Radio Alice France Californication Paris RedHotChiliPeppers Wifi_SSID_5874 Mario is shopping near Chatelet Mario: « let’s grab a coffee at SB ! » Alice has just shared a photo [ view ] Lucie is listening to Californication [ i like this ] Kevin: « new McChicken is great! » Car incident 32 meters away Alice GPS Wifi McDonald’s restaurant Radio currently playing a song Christine

Context Approach Framework Evaluation Conclusion Context Aggregation and Filtering process Social updates Aggregator Sniffers Notifier Filter User Actions and tags Contextual clouds Notifications Context Interfaces Abstraction and weighting Services

Context Approach Framework Evaluation Conclusion Considering a first context dimension: browsing activity Social updates Aggregator Sniffers Notifier Filter User Actions and tags Contextual clouds Notifications Context Interfaces Abstraction and weighting Services Publication : “ Workspace Awareness without Overload” Smart Offices and Other Workspaces, Intelligent Environments 2009

Context Approach Framework Evaluation Conclusion How to synthesize the contextual tag cloud from web browsing ? The user opens a web page…

Context Approach Framework Evaluation Conclusion How to synthesize the contextual tag cloud from web browsing ? Low level and static author description Automatic content analysis Mining semantic concepts from content People-entered tags (wisdom of crowds) 1) URL is sent to the Context Aggregator 2) Content is analyzed by enhancers (including web services)

Context Approach Framework Evaluation Conclusion Contextual Tag Clouds, vector space model and algebra Sample tag cloud R : (normalized) Aggregation of a set V of normalized Tag Clouds  normalized sum: Relevance of Tag Cloud R with S  cosine similarity: 0.1 0.1 0.3 0.5 « Discount » « Flight » « Asia » « Travel »

Context Approach Framework Evaluation Conclusion Contextual Tag Clouds, extraction and enhancement functions 1. Extracting weighted terms from: Resource Metadata Title Keywords Description Parameters = 50 = 10 = 1

Context Approach Framework Evaluation Conclusion Contextual Tag Clouds, extraction and enhancement functions 2+3. Extracting weighted terms from: 2. Search Query ambient, awareness 3. Resource Location video, all, alcatel-Lucent

Context Approach Framework Evaluation Conclusion Contextual Tag Clouds, extraction and enhancement functions Extracting weighted terms from: Social Annotations w poster = 11, w work = 11, w gtd = 10, w done = 10, w inspiration = 7, …

Context Approach Framework Evaluation Conclusion Contextual Tag Clouds, extraction and enhancement functions Extracting weighted terms from: Semantic Analysis of content MIT, Tim Berners-Lee, …

Context Approach Framework Evaluation Conclusion Context Aggregation and Filtering process –- in the enterprise Social updates Aggregator Sniffers Notifier Filter User Actions and tags Contextual clouds Notifications Context Interfaces Abstraction and weighting Services

Context Approach Framework Evaluation Conclusion Implementation Firefox extension (Javascript) to track web browsing Windows daemon (C++) to track opened PDF documents Lightweight HTTP Server (Java) + 5 tag extractors (Java) incl. 2 web service wrappers Jetty-based HTTP Server (Java) DWR for server-push (Java) Off-line scripts (Java+shell) Firefox sidebar (HTML+Javascript) Mobile application (Java for android) Aggregator Sniffers Notifier Filter Social updates User Actions and tags Contextual clouds Notifications Context Interfaces Abstraction and weighting Services

Context Approach Framework Evaluation Conclusion Demonstration of the ECN prototype Demonstration : Enterprise Contextual Notifier, Contextual Tag Clouds towards more Relevant Awareness . CSCW 2010, the ACM Conference on Computer Supported Cooperative Work (ACM, 2010)

Context Approach Framework Evaluation Conclusion Requirements and plan Hypothesis: Recommended social updates are relevant when users’ contexts are similar To evaluate: Relevance of social updates based on contextual similarity Relevance of social updates to the context of their posting Differences between context from virtual and social sensors Experimentation plan: (1 week) 2 personalized surveys per user Indexing as tags clouds & Matching contexts

Context Approach Framework Evaluation Conclusion From browsing activity to social matching Temporal indexing period = 10 mn. Common tags: JAVA, DEV Common tags: TRAVEL  Recommend u5’s social update to u1  Recommend u3’s social update to u7

Context Approach Framework Evaluation Conclusion 1. Relevance of social updates based on contextual similarity Matching

Context Approach Framework Evaluation Conclusion Survey #1 … and 3 social updates with various relevance scores, for each context 1 2 3 4 1 2 3 4 Survey #1 : For each user, 5 personal contextual clouds are proposed…

Context Approach Framework Evaluation Conclusion Survey #1 results 1/2  rarity of good matches (few participants  few common tags)

Context Approach Framework Evaluation Conclusion Survey #1 results 2/2  Accuracy = 72% (based on MAE between relevance scores and ratings) Accuracy

Context Approach Framework Evaluation Conclusion 2. Relevance of social updates to the context of their posting

Context Approach Framework Evaluation Conclusion Survey #2 Survey #2 : For each user’s social update, Evaluation of relevance between social updates and context of posting Results Average relevance rating: 50.3% (over 59 social updates), including: - 71% for social bookmark notifications - 38% for tweets ( ≈ 41% of “me now” statuses on twitter [Naaman’2010]) Publication : Contextual Recommendation of Social Updates, a Tag-based Framework . International Conference on Active Media Technology (Springer, 2010)

Context Approach Framework Evaluation Conclusion 3. Differences between context from virtual and social sensors Combining virtual and social sensors: good compromise between quantity and quality of matches 280k Number of matches 40k 170k 130k 70k 10k Low precision matches High precision matches

Context Approach Framework Evaluation Conclusion Conclusion Goal : Maintain awareness, reduce information overload Previous work : Recommendation of documents and people, based on content of browsed documents Problems: closed vocabularies, learning phases, lack of metadata… Hypothesis : A social update is relevant to a person, if this person’s context is similar to the context of the sender at the time of posting Proposition : Aggregate context as tags, from virtual and social sensors, for ranking social updates

Context Approach Framework Evaluation Conclusion Contributions and findings Contributions A tag-based context management framework, and software implementation A working social awareness application that recommends relevant updates A methodology, and tools to evaluate the performance of such applications Findings (using web browsing activity as context) : Encouraging results: 72% accuracy Half social updates are relevant to web browsing context, depending on nature Ready for integration in Bell Labs research projects As a working Social Radar implementation for “ One Million Conversations” As a context management system for “ Environment as a Service”

Context Approach Framework Evaluation Conclusion Future work Quality of contextual tag clouds, to be improved Semantic analysis, clustering, and filtering of tags Dynamic weights (based on time) Additional context sensors, to be supported E.g. leverage physical sensors and geo-localized social streams Context tag cloud manipulation interface, to be studied Add graphical representations of tags Multidimensional/hierarchical tag clouds Edit tags and their weights

www.alcatel-lucent.com Thank you for your attention! Merci pour votre attention !

PhD Defense - A Context Management Framework based on Wisdom of Crowds for Social Awareness Applications

More Related Content

What's hot (20)

Viewers also liked (19)

Similar to PhD Defense - A Context Management Framework based on Wisdom of Crowds for Social Awareness Applications (20)

More from Adrien Joly (6)

PhD Defense - A Context Management Framework based on Wisdom of Crowds for Social Awareness Applications

Editor's Notes