1、ExplainableRecommendation1Explainable AIAttention from Government Industry Academia?=2Explainable AIAttention from Government IndustryXAI AcademiaGDPR (General Data Protection Regulation)From Prof. Jitao Sang s CCDM Talk3Explainable AIInvests 1,600 engineers tosupport GDPR complianceAttention from G
2、overnment IndustryMoves more than 1.5billion users out of reachof European privacy law Academia4Explainable AIAttention from Government Industry Academia11 accepted papers mentionedinterpretation/explanation in the title5Traditional vs. Explainable Recommendation Traditional recommendation Explainab
3、le recommendation What, Who, When, Where WhyConnect the item with the user: persuasiveness, trust, satisfactionIt may help you betterunderstand some majordecisions of SatyaIt impacts how Satya thinksabout leadership6Explainable Recommendation for Ads7Explainable Recommendation for AdsBase AdCopyExpl
4、anation (Decoration)Long Ad Title (LAT)Consumer RatingTop AdsTwitterMerchant RatingReviewSite Link (ESL)8Application Scenarios In AdsSearch AdsBing Ads PlatformNative Ads on MSNNative Ads on 9Outline Definition and goals Forms of explanations Explainable recommendation pipelines10Outline Definition
5、and goals Forms of explanations Explainable recommendation pipelines11Goals of Explainable AIEffectivenessPersuasivenessReadabilityTransparencyTrustModelExplainabilityPresentationQuality12Relationships between the Goals: CorrelatedEvaluation results on 82 usersModelexplainabilityQuestionnaire: “I un
6、derstandQuestionnaire: “I candepend on the system.”Number of recommendations chosenwhat the system bases itsrecommendations on.”Not correlatedInterview: acceptance scenarioQuestionnaire: “I think that theartworks that the system recommendscorrespond to my art interests.”PresentationqualityUMUAI20013
7、8Relationships between the Goals: Trade-OffModelexplainabilityTrade-offTrade-offTrade-offPresentationqualityUMUAI200148Explainable Recommendation Systems1-800-FLOWERS.COM Elegant Flowers for LoversFog Harbor Fish HouseTheir tan tan noodles are made ofmagic. The chili oil is really appetizing.However
8、, prices are on the high side.EffectivenessPersuasivenessReadabilityTransparencyModelExplainabilityPresentationQualityTrustDefinition of Explainable Recommendation In general, give statements that support the recommendations WWW2013 Application specific Model-explainability help users understand the
9、 system behavior CHI2012 Presentation quality-Effectiveness help users make more accurate decisions IUI2015 Presentation quality-Persuasiveness convincing users to adopt recommendationsIUI2015, IUI200916Outline Definition and goals Forms of explanations Explainable recommendation pipelines17Forms of
10、 Explanations Three basic formsIUI200189Item-Based Explanations “You may like the item because it is similar to items you previously like”AmazonIUI200519User-Based Explanations “You may like the item because a user similar to you like this item”FacebookWWW201320Feature-Based Explanations “You may li
11、ke the item because this item contains features you like”IUI2005Arxiv2018SIGIR201421Dialog-Based ExplanationsMicrosoft XiaoIce (小冰)KDD201622Structured Overview ExplanationsNewsMapCHI200323Outline Definition and goals Forms of explanations Explainable recommendation pipelines24Existing Methods Post H
12、ocUsers ndation Recommended Explanation(, )items MethodExplanationItems Can we enhance persuasiveness (presentationquality) in a data-driven way?Existing Methods - EmbeddedUsers ExplanationExplanation RecommendedMethoditems Items Can we build an explainable deep model(enhance model explainability)?E
13、xisting Methods - EmbeddedUsers ExplanationExplanation RecommendedMethoditems Items Is there a pipeline that better balancepresentation quality and model explainability?The Wrapper PipelineUsers Explanation RecommendedRecommendationExplanationMethoditems model (, )Items Our ContributionsA feed-back
14、aware generativemodel that enhancepersuasiveness (presentationquality) in a data-driven wayPotentialusersModelexplainabilityPresentationqualityEasy todeliverAdvertisers,E-commercewebsitesPost-hocEmbeddedWrapperAlgorithmdevelopers,researchersAn explainable deeprecommendation model(enhances model expl
15、ainability)Futureadvertisers,algorithmdevelopersA new reinforcement learning pipelinethat better balances presentation qualityand model explainabilityOur ContributionsA feed-back aware generativemodel that enhancepersuasiveness (presentationquality) in a data-driven wayPotentialusersModelexplainabil
16、ityPresentationqualityEasy todeliverAdvertisers,E-commercewebsitesPost-hocEmbeddedWrapperAlgorithmdevelopers,researchersFutureadvertisers,algorithmdevelopersFeed-back Aware Generative Model Application scenario: Ads Mainstreamed, revenue increasedAdvertiser PlatformSearch AdsNative Ads / ONative Ads
17、 / MSNExample ResultsThe model has the ability togenerate persuasive phrasesInput AdTitlejob applications onlineNew: job application online. Apply today & find your perfect job!Now hiring - submit an application. Browse full & part time positions.3 open positions left - apply now! Jobs in your areaO
18、pen positions left - apply now! Job application online.7 open positions left - apply now! Jobs in your areaSales positions open. Hiring now - apply today!The model candifferentiate similarinputsOutputAdDescriptionsDiversified resultsInput AdTitleUS passport applicationFind US passport application an
19、d related articles. Search now!Quick & easy application. Apply for your passport online today!Quick & easy application. Find government passport application and related articles.Government passport application. Quick and easy to search results!Start your passport online today. Apply now & find the b
20、est results!Open your passport online today. 100% free tool!OutputAdDescriptionsOur ContributionsA feed-back aware generativemodel that enhancepersuasiveness (presentationquality) in a data-driven wayPotentialusersModelexplainabilityPresentationqualityEasy todeliverAdvertisers,E-commercewebsitesPost
21、-hocEmbeddedWrapperAlgorithmdevelopers,researchersFutureadvertisers,algorithmdevelopersOur ContributionsPotentialusersModelexplainabilityPresentationqualityEasy todeliverAdvertisers,E-commercewebsitesPost-hocEmbeddedWrapperAlgorithmdevelopers,researchersAn explainable deeprecommendation model(enhanc
22、es model explainability)Futureadvertisers,algorithmdevelopersThe 33rd AAAI Conference on Artificial Intelligence (AAAI 2019)Honolulu, Hawaii, USAExplainable Recommendation ThroughAttentive Multi-View LearningJingyue Gao1,2, Xiting Wang2,*, Yasha Wang1, Xing Xie21 Peking University, 2 Microsoft Resea
23、rch AsiaMotivation: An Explainable Deep Model Build a network based an explainable deep hierarchy to improveaccuracy and explainability simultaneouslyIsA RelationshipIsAMicrosoft Concept GraphFeature hierarchyChallenges Model multi-level explicit features from noisy and sparse data Shrimp Seafood Me
24、at Users may be interested in Seafood even if they mainly mention Shrimp andMeat in reviews Generate explanations from the multi-level structure Features could be semantically overlapping Simultaneously puttin Shrimp and Seafood in explanations will degrade userexperience.Deep Explicit Attentive Mul
25、ti-View Learning ModelHierarchical Propagation Propagate user-feature interest over the hierarchical structureFeature embedding is trained beforehand using GloVeto capture both semantic and hierarchical informationAttentive Multi-View Learning Features at different hierarchical levels are regarded a
26、s different viewson user interest and item quality In each single view, we extend EFM to EFM+ by adding rating biases Loss in each viewAttentive Multi-View Learning A common paradigm of Multi-View Learning Enforce agreement on predictions from multiple views Co-regularization lossAttentive Multi-Vie
27、w Learning Predictions from all views ar attentively combined for final prediction Joint learningPersonalized Explanation Generation Select features from the feature hierarchy for explanation Whether a feature is important in recommendation Whether user is interested in How well item performs on The
28、 weight of the view that belongs to Define a Utility Function for each featurePersonalized Explanation Generation Objective Maximize the sum of utilities of selected features Constraint Features cannot be simultaneously selected with their ancestors in the hierarchy Constrained Tree Node Selection P
29、roblemPersonalized Explanation Generation Dynamic Programming (, ): maximum sum of utilities if selecting nodes in the subtreerooted at ; : total number of children of (, , ): maximum sum of utilities if selecting nodes from the first children of ; : id of -th child of ;Evaluation Study on model acc
30、uracy Parameter sensitivity analysis Study on explainabilityDatasetsAmazonTuples: (user, item, rating, review, time)AmazonYelpStudy on Model Accuracy G1: only use the observed rating matrix G2: knowledge-based G3: leverage textual reviews EFM: state-of-the-art method for mining feature-level explana
31、tions DeepCoNN, NARRE: deep-learning based DEAML-V: a variant of our DEAML without attention mechanismParameter Sensitivity Analysis Number of latent factors Weight of co-regularization Weight of errors of each reviewStudy on Explainability Quantitative analysisScores (1-5) annotated by users Qualit
32、ative analysisVisualization of user interest overthe feature hierarchya 30-year-old male Yelp usera 26-year-old female Yelp userOur ContributionsPotentialusersModelexplainabilityPresentationqualityEasy todeliverAdvertisers,E-commercewebsitesPost-hocEmbeddedWrapperAlgorithmdevelopers,researchersAn ex
33、plainable deeprecommendation model(enhances model explainability)Futureadvertisers,algorithmdevelopersOur ContributionsPotentialusersModelexplainabilityPresentationqualityEasy todeliverAdvertisers,E-commercewebsitesPost-hocEmbeddedWrapperAlgorithmdevelopers,researchersFutureadvertisers,algorithmdeve
34、lopersA new reinforcement learning pipelinethat better balances presentation qualityand model explainabilityA Reinforcement Learning Frameworkfor Explainable RecommendationXiting Wang, Yiru Chen, Jie Yang, Le Wu, Zhengtao Wu, Xing XieMicrosoft Research Asia, Peking University, Tsinghua University, H
35、efei University of Technology, USTCContribution Design a reinforcement learning framework for explainablerecommendation Model-agnostic Has good model explainability Can flexibly control the presentation quality Instantiate the agents with personalized-attention-based neuralnetworks Evaluate the effe
36、ctiveness of our method by using both offlineexperiments and evaluation with human subjects54Problem Definition Input: user ID and/or some side informationitem ID : interpretable component User set , is a user Item set , is an item A recommendation model to be:explained (, )Attributes like “price -
37、low”;user review; product image; Output Explanation =The th interpretable component is selectedThe th interpretable component is not selectedReinforcement Learning Framework Advantages: model-agnostic, model-explainability, presentation quality56Reinforcement Learning Framework Advantages: model-agn
38、ostic, model-explainability, presentation qualityIndependent57Reinforcement Learning Framework Advantages: model-agnostic, model-explainability, presentation qualityAgent 2 can predict (, )Model-explainability reward:58Reinforcement Learning Framework Advantages: model-agnostic, model-explainability
39、, presentation quality will increase if is good given 59Couple AgentsSentence-level Explanation60Optimization Goal Maximizing expected rewardReward Model-explainabilityPresentation quality(Sentiment) ConsistencyReadabilityConciseness:desirable sentence lengthCoherencereturns the sentiment score of61
40、Optimization Method Doubly Stochastic Policy GradientAgent 1Agent 262Offline EvaluationEach training sample: user, item, rating, review text63Different recommendationmodels to be explainedOffline EvaluationPresentation qualityModel explainabilityPresentation qualityModel explainability64Offline Eval
41、uationDifferent parameter settings: desirable explanation lengthPresentation qualityModel explainabilityModel explainabilityPresentation quality65Evaluation with Human Subjects Ask the participants to choose the explanations that are most usefulin helping them decide whether they will go to the rest
42、aurants66Evaluation with Human SubjectsP3Frequent words in reviews:P4Words related to servicesWords related to food67Summary Enhance persuasiveness in a data-driven way Build an explainable deep model to improve model explainability Better balance presentation quality and model explainability Future
43、 work Integrate heterogeneous symbolic knowledge for improving explainability Combine models from computational psychology to enhance presentationquality Conversational interpretability based on generative modelsVenueYearTitleAuthorsDatasetMethodPresentationExplainMajor ConclusionInsightmechanism12I
44、UI2005Explaining recommendations: Mustafa Books (40,000)LIBRA: (weighted) combiningrankings of content andcollaborative filtering resultsSorted list ofFeature/Item/UserYesSatisfaction: KSE (feature) ISE (item) NSE (user)Dividing into KSE, ISE, andNSE. Conclusion (34 users).satisfaction vs. promotion
45、BilgicICDE(Workshop)2007A survey of explanations inrecommender systemsNavaTintarevOne of them:structuredoverviewSummary of userinteractions, presentationtypes, commercial systems,major aspects.34UMUAI20082009The effects of trust in andacceptance of a content-basedart recommenderHenrietteCramerArtUse
46、r study of a content-basedmethod. Three groups: why grouppresents related items, uncertainSorted list ofitems (clickwhy?)YesYesTransparent - Acceptance. Evaluation results on 82 users.Transparency !- trust.Uncertainty does not help.Significant correlationsbetween metrics.Questionnaire. Survey.Whyjus
47、tificationtransparency.556 users. Item, feature, user,tag-based. Descriptive tags.12 usersIUITagsplanations: explainingrecommendations using tagsJesseVigMovieLensMusicJointly consider tag relevance andtag preference. Handle tag quality& redundancy with filtering.Sorted list oftagsBest: RelSort, wors
48、t: RelOnly(justification, effectiveness,mood). Subjective Factual56CHI(ExtendedAbstract)20122013The role of transparency inrecommender systemsRashmiSinhaAsk users to evaluate 5 existing(commercial) music recommendersystemsUsers like and feel moreconfident about systems theyperceive as transparentWWW
49、Do social explanations work?Studying and modeling theeffects of social explanations inrecommender systemsAmitSharmaFriendPop,RandFriend,OverallPop,GoodFriend,GoodFrCountGood friends help increaselikelihood of clicking. Therating is not influenced.The artists name and photomatter too. Familiaritymatt
50、ers.78KDDSIGIR2014201420152015Jointly modeling aspects,ratings, and sentiments formovie recommendation(JMARS)Explicit factor models forexplainable recommendation g Zhangbased on phrase-levelsentiment analysisFLAME: a probabilistic model Yao Wucombining aspect basedopinion mining andIMDbJMARS: factor
