Pragmatic Deep Learning for image labelling - An application to a travel recommendation engine主讲人：Dataiku数据科学家 Alexandre Hubert Introduction and Context Iterative building of a recommender systemLabeling ImagesPragmatic deep learning for dummiesPost Processing AKA: Image for BI on steroidsOutlineResultsMore images !Dataiku• Founded in 2013 • 90 + employees, 100 + clients • Paris, New-York, London, San Francisco, SingaporeData Science Software Editor of Dataiku DSSDESIGNLoad and prepareyour dataPREPAREBuild yourmodelsMODELVisualize and shareyour work ANALYSERe-execute your workflow at easeAUTOMATEFollow your production environmentMONITORGet predictionsin real timeSCOREPRODUCTIONE-business vacation retailerNegotiate the best prize for their clientsDiscount luxury• Key FiguresSale Image is paramountPurchase is impulsive18 Millions of clients. Hundreds of sales opened everydaySpecificitiesHighly temporary sales-> Classical recommender system fail-> Time event linked (Christmas, ski, summer)Expensive Product-> Few recurrent buyers-> Appearance counts a lotIterative Building of a Recommender SystemBasic Recommendation EnginesOther FactorsOne Meta Model to Rule Them AllRecommenders as featuresMachine learning to optimize purchasing probabilityCombineRecommendDescribeCleaning, combining and enrichment of dataRecommendation EnginesOptimization of home displaythe application automatically runs and compiles heterogeneous dataGeneration of recommendations based on user behaviourEvery customer is shown the 10 sales he is the most likely to buyCustomer visitsPurchases Sales ImagesMetal model combine recommendations to directly optimize purchasing probabilityMeta ModelRecommender system for Home Page Ordering+7% revenueSales information (A/B testing)Batch Scoring every nightWhy use Image ? We want do distinguish« Sun and Beach »« Ski »A picture is worth a thousand wordsSales ImagesIntegrating Image InformationLabeling ModelPool + Palm Trees Hotel+ MountainsPool + Forest + Hotel + SeaSea + Beach +Forest + HotelSales descriptions vectorCONTENT BASEDRecommender SystemImage Labelling For Recommendation EnginePragmatic Deep learning for “Dummies”Using Deep Learning modelsCommon Issues “I don’t have GPUs server”“I don’t have a deep leaning expert”“I don’t have labelled data” (or too few)“I don’t have the time to wait for model training ”I don’t want to pay to pay for private apis” / “I’m afraid their labelling will change over time”“I don’t have (or few) labelled data” -> Is there similar data ? Solution 1 : Pre trained models PLACES DATABASEUSSUN DATABASE205 categories2.5 M images307 categories110 K imagestower: 0.53skyscraper: 0.26swimming_pool/outdoor: 0.65inn/outdoor: 0.06Solution 1 : Pre trained models If there is open data, there is an open pre trained model !• Kudos to the community• Check the licensingExample with Places (Caffe Model Zoo) :Solution 2 : Transfer Learning Credit : Fei-Fei Li & Andrej Karpathy & Justin Johnson http://cs231n.stanford.edu/slides/winter1516_lecture11.pdfPLACES DATABASEOUR DATASUN DATABASETraining (optional)Pre-trained modelVGG16tower: 0.53skyscraper: 0.26Re-TrainingTransferred Data :Last convolutional layer featuresRe-trained modelTensorFlow2 fully connected layersCaffe Model ZooGPUCPUGPULeverage existing knowledge !Solution 2 : Transfer Learning Accuracy: 72%, Top-5 Acc: 90 % > state of the art on dataset alonePost Treatment & Results(Or how we transfer the labelling information)Using Images information for BI on steroids Labels post-processingComplementary informationRedondant informationIssue with our approach:Solution : NMF Matrix FactorizationDimension ReductionExplicabilitySparsityBalancednessImage content detectionTopic scores determine the importance of topics in an imageTOPICTOPIC SCORE (%)Golf course – Fairway – Putting green31Hotel – Inn – Apartment building outdoor30Swimming pool – Lido Deck – Hot tub outdoor22Beach – Coast - Harbor17TOPICTOPIC SCORE (%)Tower – Skyscraper – Office building62Bridge – River – Viaduct38Results ? 1) Visits : • France and Morocco• Pool displayed2) First Recommendation • Mostly France & Mediterranean • Fails to display pools3) Only Images recommendation • Pool all around the world• Does not respect budget 4) Third column = Right Mix 1)2)3)4)ConclusionDo iterative data science !Start simple and grow Evaluate at each stepsImage labelling = BI on steroidsTransfer LearningKick-start your projectGain time and moneyAny Data Scientist can do itDeep Learning Don’t start from scratch !Is there existing data ? Is there a pre-trained model ?Learned along the way What’s next ? Attractiveness = % visits with tag / % sales with tag For ski sales, indoor pictures performs betterWhat’s Next ? KenyaPragueBerlinCambodiaWhat’s Next ? Customize the Image !KenyaPragueBerlinCambodiaThank you for your attention !
发表评论取消回复