1 / 12

Wolfgang Woerndl , Henrik Muehe, Stefan Rothlehner, Korbinian Moegele

Explore decentralized, item-based collaborative filtering for Windows Mobile PDAs, enabling context-aware recommendations in mobile scenarios like tourist guides. Our system optimizes storage, supports group recommendations, and integrates context scores for personalized experiences.

mgoff
Download Presentation

Wolfgang Woerndl , Henrik Muehe, Stefan Rothlehner, Korbinian Moegele

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. Context-Aware Recommendations in Decentralized, Item-Based Collaborative Filtering on Mobile Devices Wolfgang Woerndl, Henrik Muehe,Stefan Rothlehner, Korbinian Moegele Technische Universitaet Muenchen Munich, Germany

  2. Motivation • Recommender systems • Successful application for example in online shops • As yet mostly centralized systems • Mobile scenario promising, more focused information access • Decentralized, mobile recommender systems • Recommendation on client, connection to server not required • Potential advantages with regard to privacy • Goal of this work • Design, implementation and test of a decentralized recommender system for Windows Mobile PDAs + context-aware application • Recommender systems basics, design of our approach • Contextualization, example scenario: mobile tourist guide

  3. Recommender Systems • Recommender systems • Individual approaches  Consider only active user • Collaborative filtering  Consider also ratings of other users • Collaborative Filtering (CF) • User-based CF • Determine similar users  Drawbacks include cold start, performance • Item-based (or model-based) collaborative filtering • Calculate model of pairwise item similarities based on ratings  Advantage: can be pre-computed in advance • Recommend items that are similar to items that have been positively rated by the active user in the past

  4. Decentralized, Mobile Recommender System • Decentralized approach • Peers in system exchange rating vectors of their users • Each peer computes local matrix of item-item similarities • Recommendation based on rating vector of user • PocketLens • Decentralized approach from research literature • Stores intermediate results when calculating item similarity • Disadvantages • Very big data model (item-item similarity) • Limited extensibility, only new rating vectors • Our system implements decentral, item-based collaborative filtering for PDAs

  5. Our approach • Optimization of storage requirements • Extensibility of model by introducing versioned rating vectors • Store history of ratings • Allow for changing and deleting ratings • Integration of group recommendations • Users in front of shared public display • Implemented scenario: display, rateand recommend images on PDA • Windows Mobile (.NET Compact Framework) • Tested in small user study (13 users)

  6. Scenario with Public Shared Display

  7. Contexualization • So far, mobilerecommender on PDA, but not context-aware • Goal: Adapttothecurrentusersituation (time, position, …) • Proposedmethodis a combination score, e.g. linear: • score = a * cf-score + b * ctx-score • scores: +1 best value; -1 worstvalue • cf-score: ratingpredictionaccordingtotheexplained item-basedcollaborativefiltering • ctx-score: score accordingtothecurrentusercontext • Forexample, currentdistancetopoint-of-interest (POI) • ctx-score = -1 meaning, forexample • POI istoo far away • Restaurantiscurrentlyclosed

  8. Scenarios • Mobile exhibitionguide • Search for products, exhibitors or places of interest etc., additional functions such as appointment schedule, virtual business cards • Context: Locationofexhibitionboothsandhalls • Indoorpositioningwithbluetooth-basedinfrastructure • Mobile cityguidefortourists • So far: application on PDA displaysinformation (image/text) andplayaudiofilefornearest POI (sight) • Extended withexplaineddecentral CF method • Devices are not networked, model isupdatedwhentouristsreturndevice • Context: POIs in vicinity, rankedaccordingto CF score • Rating acquisition • Explicit: User entersrating (good/mediocre/bad, or 5 starscale) • Implicit: System determineratingbyusageofaudiofile • Implementedandtested

  9. User Interface Mobile City Guide

  10. Evaluation • User study with real users (tourists) in Prague • 2 weeks in late september, 30 volunteering participants, aged between 17 and 76, various nationalities • Users could use city guide with recommender system for free and fill out questionnaire, instead of paying for rental • Questionnaire with 10 questions • Positive feedback, users liked application and recommender system • Example question: „I felt that the mobile guide selected sights according to my interests/ratings“ • 17% totally agree, 36% agree, 23% tend to agree7% tend to disagree, 10% disagree, 7% totally disagree • User are happy to give away some information about themselves in return for personalized recommendations • But users gave mixed feedback regarding explicit rating dialogue • Rely more on implicit ratings in future

  11. Conclusion • Summary • Implementing a decentralizedrecommendersystemfor PDAs • Innovationsincludestorageoptimization, improvedextensibilityandgrouprecommendations • Contextualizedapplication in a mobile touristguide • User studywithearly promising results • Current & futurework • Mobile touristguide, ongoingstudyw.r.t. recommendationquality • Are recommended POIs reallymoreinterestingforusersthan just nearest POIs? • Are thananydifferencesbetweenrecommendationsbased onimplicit versus explicit ratings? • Improvecontext-awareness, refinemethodtocombine CF andcontextscores • Test in otherscenario(s)

  12. Context-Aware Recommendations in Decentralized, Item-Based Collaborative Filtering on Mobile Devices Wolfgang Woerndl woerndl@in.tum.de Questions?

More Related