310 likes | 823 Vues
PATCH Workshop, UM’07, Corfu, Greece June 25, 2007 A Framework for Guiding the Museum Tour Personalization Mykola Pechenizkiy, Toon Calders Information Systems Group Department of Computer Science Eindhoven University of Technology the Netherlands Outline Introduction
E N D
PATCH Workshop, UM’07, Corfu, Greece June 25, 2007 A Framework for Guiding the Museum Tour Personalization Mykola Pechenizkiy, Toon Calders Information Systems Group Department of Computer Science Eindhoven University of Technologythe Netherlands “A Framework for Guiding the Museum Tour Personalization” by M. Pechenizkiy & T. Calders
Outline • Introduction • Motivation and goals • Personalization and adaptation of the cultural heritage content • Personalization process • The basic approaches for personalization • Nonintrusiveness: efficient learning of user preferences • What is special in personalization of access to the museum artworks? • The generic framework: Optimally Personalized Museum Tour • Formal description of the museum tour personalization • Evaluation methodologies for personalization • Challenge of Scientific Evaluation of Personalization • The methodological framework for evaluating and guiding personalization process • Discussions and further research “A Framework for Guiding the Museum Tour Personalization” by M. Pechenizkiy & T. Calders
Motivation • Information overload - too much content! • Too many artworks to see them all at one visit • Diversity of content and diversity of visitors’ needs • Web-access to museums collections • Introduce the existing galleries, collections, artworks • Educate virtual visitors • Recommend virtual visitors what they may want to do in the museum • Suitable galleries, collections, or personalized tours • CHIP • “I know what you’ll see in the museum next <Sunday, month, summer, …>” “A Framework for Guiding the Museum Tour Personalization” by M. Pechenizkiy & T. Calders
Goals • to tailor personalized access to a visitor’s (potentially changing) interests and preferences without demanding to express them explicitly and without increasing visitor’s intrusiveness. • Interest vs. interests: coverage • Recommending a tour, not an individual artwork. • to start offering the most relevant information (recommendations) to the (possibly first-time) visitors as soon as possible while trying to minimize the users’ intrusion. “A Framework for Guiding the Museum Tour Personalization” by M. Pechenizkiy & T. Calders
Sources for the introductory slides • “AI Techniques for Personalized Recommendations” IJCAI’03 Tutorial by Konstan et al. • “Comparing Human Recommenders to Online Systems” by Rashmi Sinha & Kirsten Swearingen • “Toward the Next Generation of Recommender Systems: A Survey of the State-of-the-Art and Possible Extensions” and • “Personalization Technologies: A Process-oriented Perspective” by G. Adomavicius and A. Tuzhilin “A Framework for Guiding the Museum Tour Personalization” by M. Pechenizkiy & T. Calders
Personalization process • Customization (adaptable) vs. personalization (adaptive) • customization (or adaptability) assumes active user participation (a visitor has a possibility to configure the adjustable properties of the application) and explicit input (manually creating and/or editing an own profile). • In personalized and adaptive applications not a visitor, but the system is responsible for automatic personalization of structure, content and its outlook according to visitor’s preferences, which can be either also learnt by the system automatically, or, alternatively, the necessary information can be explicitly provided by the visitor. “A Framework for Guiding the Museum Tour Personalization” by M. Pechenizkiy & T. Calders
Personalization process • understanding, who is the user and what kind of content is of his or her interest, through user modelling process that often consists of some relevant data collection, its analysis and transformation to actionable knowledge; • delivering the personalized content, • measuring and evaluating the impact of personalization on the visitor’s satisfaction in particular and on achieving goals defined by the resources provider in general “A Framework for Guiding the Museum Tour Personalization” by M. Pechenizkiy & T. Calders
Recommendation Process “A Framework for Guiding the Museum Tour Personalization” by M. Pechenizkiy & T. Calders
Recommendation Process “A Framework for Guiding the Museum Tour Personalization” by M. Pechenizkiy & T. Calders
Consumers vs. providers • User-centric • Find what I want • Know I will like it • Trust system to help me • Team up with my friends to defeat evil marketers • Provider-centric • Show people what they will buy • Learn what people want so you have it • Learn how much they want it so you charge as much as possible “A Framework for Guiding the Museum Tour Personalization” by M. Pechenizkiy & T. Calders
Key Problems • Gathering “known” ratings for matrix • Explicit • Ask people to rate items (what items?) • Implicit • Learn ratings from user actions • Extrapolate unknown ratings from known ratings • Mainly interested in high unknown ratings • Key problem: matrix of ratings is sparse • most people have not rated most items, unless it is a controlled experiment or aka pre-test for evaluation of users tastes • Three groups of approaches • Content-based; Collaborative; Hybrid • Evaluating extrapolation methods “A Framework for Guiding the Museum Tour Personalization” by M. Pechenizkiy & T. Calders
Design of Personalization Process • develop good metrics to determine personalization impact; • study the feedback-integration problem and develop novel methods to address it; • investigate the goal-driven design process in order to achieve better personalization solutions. “A Framework for Guiding the Museum Tour Personalization” by M. Pechenizkiy & T. Calders
The basic approaches for personalization • Content-based methods • analyze the common features among the items I a visitor rated highly and recommend those items that are similar to I • Collaborative-based methods • search for peers of a visitor that have similar preferences and then recommend those items that were most liked by the peers • User-to-user or Item-to-item collaborative filtering • Hybrid approaches • combine collaborative and content-based methods • Cascade, parallel, meta • Memory-based algorithms (lazy-learners) • heuristics that can predict ratings based on memorizing and searching the entire collection of previously rated artworks by the visitors • Model-based algorithms • use the collection of ratings to learn a model, which is then used to make rating predictions “A Framework for Guiding the Museum Tour Personalization” by M. Pechenizkiy & T. Calders
Content-based recommendations • Main idea: recommend items to customer U similar to previous items rated highly by U • Artwork recommendations • recommend artworks with same painter, style, year, etc. • or with “similar” content … “A Framework for Guiding the Museum Tour Personalization” by M. Pechenizkiy & T. Calders
Limitations of content-based approach • Finding the appropriate features • i.e., features on paintings themselves as images (not their annotations) • Overspecialization • Never recommends items outside user’s content profile • People might have multiple interests • Recommendations for new users • How to build a profile? “A Framework for Guiding the Museum Tour Personalization” by M. Pechenizkiy & T. Calders
User-user/item-item Collaborative Filtering Submit/store ratings, compute correlations, request recommendations, identify neighbors, select items, predict rating “A Framework for Guiding the Museum Tour Personalization” by M. Pechenizkiy & T. Calders
Limitations of Collaborative Filtering • Collaborative filtering cannot recommend new items: no one has rated them • Random • Content analysis • Collaborative filtering cannot match new users: they have rated nothing • Provide average ratings • User agents collect implicit ratings • Put users in categories • Carefully select items for users to rate “A Framework for Guiding the Museum Tour Personalization” by M. Pechenizkiy & T. Calders
Five basic types of approaches “A Framework for Guiding the Museum Tour Personalization” by M. Pechenizkiy & T. Calders
5 approaches to recommendation and their typical positive (above) and negative (below) aspects, according to Burke (2002) “A Framework for Guiding the Museum Tour Personalization” by M. Pechenizkiy & T. Calders
Hybridization methods Motivation: the various techniques have partly complementary strengths and weaknesses • Weighted • The scores (or votes) of several recommendation techniques are combined together to produce a single recommendation • Switching • The system switches between recommendation techniques depending on the current situation (short−term and long−term models) • Mixed • Recommendations from several different recommenders are presented at the same time (e.g. Amazon’s web pages) • Feature combination • Features from different recommendation data sources are thrown together into a single recommendation algorithm (CBR) • Cascade • One recommender refines the recommendations given by another • Feature augmentation • Output from one technique is used as an input feature to another • Meta-level • The model learned by one recommender is used as input to another “A Framework for Guiding the Museum Tour Personalization” by M. Pechenizkiy & T. Calders
Our current focuses • Noninrusiveness • As accurate model as possible in as few rating requests as possible • Coverage of user interests • If someone is interested in landscaped and also in portraits, but lesser than in landscapes, what happens? • Recommending tour not an individual artwork • Implies new challenges and constrains • By now – our focus is coverage • In general – many other things are interesting (e.g. physical placement of artworks) “A Framework for Guiding the Museum Tour Personalization” by M. Pechenizkiy & T. Calders
Noninrusiveness Efficient Learning of User Preferences (Active Learning) • ActiveCP approach • utilizes information about items controversy and popularity • VC-WMP algorithm • clusters items by categories in order to reduce the dimensionality and sparseness of the score matrix and applies a majority vote learner with selection of votes based on the correlation of user profiles • Entropy-driven active learning algorithm • allows to better balance learning efficiency and user satisfaction • Transductive experimental design • explores available unrated items and selects such items that are on the one side hard-to-predict and on the other side representative for the rest of the items “A Framework for Guiding the Museum Tour Personalization” by M. Pechenizkiy & T. Calders
Optimally Personalized Museum Tour A Generic Framework of the Optimally Personalized Museum Tour Problem • with every object oO , a set of characteristics c(o) is associated: • the nightwatch: {rembrandt, 17th century, oil paint, militias} • a user u which has a preference u(o) for every object oO • coverage of the Tour • quality of the Tour • benefit Function • offline vs. online settings “A Framework for Guiding the Museum Tour Personalization” by M. Pechenizkiy & T. Calders
Evaluating Predictions • Compare predictions with known ratings • Root-mean-square error (RMSE) • Another approach: 0/1 model • Recall/coverage • Number of items/users for which system can make predictions • Precision • Accuracy of predictions • Receiver operating characteristic (ROC) • Tradeoff curve between false positives and false negatives • Narrow focus on accuracy sometimes misses the point • Cautions in data interpretation • Users may like/”buy” items regardless of recommendations • Users may also avoid seeing certain artworks they might have seen based on recommendations “A Framework for Guiding the Museum Tour Personalization” by M. Pechenizkiy & T. Calders
Relevance Feedback “A Framework for Guiding the Museum Tour Personalization” by M. Pechenizkiy & T. Calders
A/B test-based guided personalization (RE – recommendation engine, RI – recommended items, URF – user relevance feedback “A Framework for Guiding the Museum Tour Personalization” by M. Pechenizkiy & T. Calders
Research directions Enabling • Similarity function • Confidence in visitors tastes • Evaluation • Guiding personalization process Advanced • profiling techniques based on data mining • finding actionable rules, sequential patterns, and signatures • adjust recommendations to the context in which it is offered • take into consideration the when, where, and with whom, etc contexts into consideration • Track and handle concept drift: • changes due to changes in hidden contexts • Changing user habits • Previous history may not accurately predict present tastes in arts “A Framework for Guiding the Museum Tour Personalization” by M. Pechenizkiy & T. Calders
Discussion and Further Research • developing methods that utilize some of the more advanced profiling techniques based on data mining • finding actionable rules, sequential patterns, and signatures • adjust recommendations to the context in which it is offered • take into consideration the when, where, and with whom, etc contexts into consideration • hidden contexts: concept drift • Changing user habits • Previous history may not accurately predict present tastes in arts • scientific evaluation of personalization • high-quality controlled experiments • fair estimating the benefits and limitations of certain personalization technique “A Framework for Guiding the Museum Tour Personalization” by M. Pechenizkiy & T. Calders
Contact Info Mykola Pechenizkiy Information Systems Group Department of Computer Science Eindhoven University of Technologythe Netherlands E-mail: m.pechenizkiy@tue.nl http://www.win.tue.nl/~mpechen THANK YOU! MS Power Point slides of other recent talks and full texts of selected publications are available online at: http://www.win.tue.nl/~mpechen/talks/talks.html “A Framework for Guiding the Museum Tour Personalization” by M. Pechenizkiy & T. Calders