hybrid web recommender systems robin burke presentation by jae-wook ahn 10/04/05

Download Hybrid Web Recommender Systems Robin Burke Presentation by Jae-wook Ahn 10/04/05

Post on 28-Dec-2015

215 views

Category:

Documents

1 download

Embed Size (px)

TRANSCRIPT

  • Hybrid Web Recommender SystemsRobin Burke

    Presentation by Jae-wook Ahn10/04/05

    Hybrid Web Recommender Systems

  • ReferencesEntre system & datasetBurke, R. (2002). Semantic ratings and heuristic similarity for collaborative filtering. AAAI Workshop on Knowledge-based Electronic Markets 2000. Feature augmentation, mixed hybrid exampleTorres, R., McNee, S., Abel, M., Konstan J., & Riedl J. (2004). Enhancing Digital Libraries with TechLens+. Proceedings of the 2004 Joint ACM/IEEE Conference on Digital Libraries.Hybrid recommender system UI issueSchafer, J. (2005). DynamicLens: A Dynamic User-Interface for a Meta-Recommendation System. Workshop: Beyond Personalization 2005, IUI05.Collaborative filtering algorithmSarwar, B., Karypis, G., Konstan, J., & Riedl, J. (2001). Item-based collaborative filtering recommendation algorithms. In Proceedings of the 10th international conference on World Wide Web.

    Hybrid Web Recommender Systems

  • Concepts and Techniques

    Hybrid Web Recommender Systems

  • Hybrid Recommender SystemsMix of recommender systems Recommender system classification knowledge sourceCollaborative (CF)Users ratings onlyContent-based (CN)Product features, users ratingsClassifications of users likes/dislikesDemographicUsers ratings, users demographicsKnowledge-based (KB)Domain knowledge, product features, users need/queryInferences about a uses needs and preferences

    Hybrid Web Recommender Systems

  • CF vs. CNUser-based CFSearches for similar users in user-item rating matrixItem-based CFSearches for similar items in user-item rating matrixCNSearches for similar items in item-feature matrixExample TF*IDF term weight vector for news recommendationItemsUsersRatings

    Hybrid Web Recommender Systems

  • Recommender System ProblemsCold-start problem Learning based techniquesCollaborative, content-based, demographic Hybrid techniques

    Stability vs. plasticity problemDifficulty to change established users profileTemporal discount older rating with less influence

    KB fewer cold start problem (no need of historical data)

    CF/Demographic cross-genre niches, jump outside of the familiar (novelty, serendipity)

    Hybrid Web Recommender Systems

  • Strategies for Hybrid RecommendationCombination of multiple recommendation techniques together for producing output Different techniques of different typesMost common implementationsMost promise to resolve cold-start problem

    Different techniques of the same typeEx) NewsDude nave Bayes + kNN

    Hybrid Web Recommender Systems

  • Seven Types of Recommender SystemsTaxonomy by Burke (2002)

    Weighted SwitchingMixedFeature combinationFeature augmentationCascadeMeta-level

    Hybrid Web Recommender Systems

  • Weighted HybridConcept

    Each component of the hybrid scores a given item and the scores are combined using a linear formula

    When recommenders have consistent relative accuracy across the product space

    Uniform performance among recommenders (otherwise other hybrids)

    Hybrid Web Recommender Systems

  • Weighted Hybrid ProcedureTrainingJoint ratingIntersection candidates shared between the candidatesUnion case with no possible rating neutral score (neither liked nor disliked)Linear combination

    Hybrid Web Recommender Systems

  • Mixed HybridConceptsPresentation of different components side-by-side in a combined listIf lists are to be combined, how are rankings to be integrated?Merging based on predicted rating or on recommender confidenceNot fit with retrospective dataCannot use actual ratings to test if right items ranked highlyExampleCF_rank(3) + CN_rank(2) Mixed_rank(5)

    Hybrid Web Recommender Systems

  • Mixed Hybrid Procedure

    Candidate generationMultiple ranked listsCombined display

    Hybrid Web Recommender Systems

  • Switching HybridConcepts

    Selects a single recommender among components based on recommendation situation

    Different profile different recommendation

    Components with different performance for some types of users

    Existence of criterion for switching decisionEx) confidence value, external criteria

    Hybrid Web Recommender Systems

  • Switching Hybrid Procedure

    Switching decisionCandidate generationScoring

    No role for unchosen recommender

    Hybrid Web Recommender Systems

  • Feature Combination HybridConcepts

    Inject features of one source into a different source for processing different data

    Features of contributing recommender are used as a part of the actual recommender

    Adding new features into the mix

    Not combining components, just combining knowledge source

    Hybrid Web Recommender Systems

  • Feature Combination Hybrid Procedure

    Feature combination In training stageCandidate generationScoring

    Hybrid Web Recommender Systems

  • Feature Augmentation HybridConceptsSimilar to Feature CombinationGenerates new features for each item by contributing domainAugmentation/combination done offlineComparison with Feature CombinationNot raw features (FC), but the result of computation from contribution (FA)More flexible to applyAdds smaller dimension

    Hybrid Web Recommender Systems

  • Feature Augmentation Hybrid Procedure

    Hybrid Web Recommender Systems

  • Cascade HybridConceptsTie breakerSecondary recommenderJust tie breakerDo refinementsPrimary recommenderInteger-valued scores higher probability for tiesReal-valued scores low probability for tiesPrecision reduction Score: 0.8348694 0.83

    Hybrid Web Recommender Systems

  • Cascade Hybrid ProcedureProcedurePrimary recommenderRanksBreak ties by secondary recommender

    Hybrid Web Recommender Systems

  • Meta-level HybridConcepts

    A model learned by contributing recommender input for actual recommender

    Contributing recommender completely replaces the original knowledge source with a learned model

    Not all recommenders can produce the intermediary model

    Hybrid Web Recommender Systems

  • Meta-level Hybrid ProcedureProcedureContributing recommender Learned modelKnowledge Source ReplacementActual Recommender

    Hybrid Web Recommender Systems

  • Experiments

    Hybrid Web Recommender Systems

  • Testbed Entre Restaurant RecommenderEntre SystemCase-based reasoningInteractive critiquing dialogEx) Entry Candidates Cheaper Candidates Nicer Candidates ExitNot narrowing the search by adding constrains, but changing the focus in the feature space

    Hybrid Web Recommender Systems

  • Testbed Entre Restaurant Recommender (contd)Entre DatasetRatingEntry, ending point positive ratingCritiques negative ratingMostly negative ratingsValidity test for positive ending point assumption strong correlation between original vs. modified (entry points with positive ratings) Small in size

    Hybrid Web Recommender Systems

  • Evaluation MethodologyMeasures ARC (Average Rank of the Correct recommendations)Accuracy of retrieval At different size retrieval setFraction of the candidate set (0 ~ 1.0)Training & Test set5 fold cross validation random partition of training/test setLeave one out methodology randomly remove one item and check whether the system can recommend itSessions SizesSingle visit profiles 5S, 10S, 15S Multiple visit profiles 10M, 20M, 30M

    Hybrid Web Recommender Systems

  • Baseline AlgorithmsCollaborative Pearson (CFP)Pearsons correlation coefficient for similarityCollaborative Heuristic (CFH)Heuristics for calculating distances between critiquesnicer and cheaper dissimilarnicer & quieter similarContent-based (CN)Nave Bayes algorithm compute probability that a item is liked / dislikedToo few liked items modified candidate generationRetrieve items with common features with the liked vector of the nave Bayes profileKnowledge-based (KB)Knowledge-based comparison metrics of EntreNationality, price, atmosphere, etc.

    Hybrid Web Recommender Systems

  • Baseline EvaluationsTechniques vary in performance on the Entre dataContent-based (CN) weakKnowledge-based (KB) better on single-session than multi-sessionHeuristic collaborative (CFH) better than correlation-based (CFP) for short profilesRoom for improvementMulti-session profiles

    Hybrid Web Recommender Systems

  • Baseline Evaluations

    Hybrid Web Recommender Systems

  • Hybrid Comparative StudyMissing components

    Mixed hybridNot possible with retrospective data

    Demographic recommenderNo demographic data

    Hybrid Web Recommender Systems

  • Results Weighted Hybrid performance better in only 10 of 30CN/CFP consistent synergy (5 of 6)Lacks uniform performance KB, CFHLinear weighting scheme assumption fault

    Hybrid Web Recommender Systems

  • Results Switching KB hybrids best switching hybrids

    Hybrid Web Recommender Systems

  • Results Feature CombinationCN/CFH, CN/CFPContributing CNIdentical to CFH, CFPCFH maintains accuracy with reduced datasetCF/CN Winnow modest improvement

    Hybrid Web Recommender Systems

  • Results Feature AugmentationBest performance so farParticularly CN*/CF*Good for multi-session profiles

    Hybrid Web Recommender Systems

  • Results CascadeCFP/KB, CFP/CNGreat improvementAlso good for multi-profile sessions

    Hybrid Web Recommender Systems

  • Results Meta-level HybridsCN/CF, CN/KB, CF/KB, CF/CNNot effectiveNo synergyWeakness of KB/CN in Entre datasetBoth components should be strong

    Hybrid Web Recommender Systems

  • DiscussionDominance of the hybrids over basic recommendersSynergy was found und