Scalable Machine Learning Pipeline For Meta Data Discovery From eBay Listings
- 1. Spark for Metadata Discovery Who We Are Metadata Discovery and Challenges Spark Solution Summary Scalable Machine Learning Pipeline for Metadata Discovery from eBay Listings Qing Zhang, Rui Li eBay Spark Summit 2016, June 6-8 San Francisco
- 2. Spark for Metadata Discovery Who We Are Metadata Discovery and Challenges Spark Solution Summary Table of Contents 1 Who We Are 2 Metadata Discovery and Challenges 3 Spark Solution 4 Summary
- 3. Spark for Metadata Discovery Who We Are Metadata Discovery and Challenges Spark Solution Summary eBay Structured Data Metadata discovery and management Listing classification Catalog and mapping listing to product Inventory insights
- 4. Spark for Metadata Discovery Who We Are Metadata Discovery and Challenges Spark Solution Summary eBay Structured Data Metadata discovery and management Listing classification Catalog and mapping listing to product Inventory insights
- 5. Spark for Metadata Discovery Who We Are Metadata Discovery and Challenges Spark Solution Summary eBay Structured Data Metadata discovery and management Listing classification Catalog and mapping listing to product Inventory insights
- 6. Spark for Metadata Discovery Who We Are Metadata Discovery and Challenges Spark Solution Summary Metadata Zoom In Important name-value pairs: brand - Dell Selling flow item specifics Search navigation Powers internal applications
- 7. Spark for Metadata Discovery Who We Are Metadata Discovery and Challenges Spark Solution Summary Metadata Discovery memory box weiss yamaha fisher-price generic modway duluth trading mek usa dnm other sahara club gokey longhorn outdoor gear trax wolverine sk spiderman vintage mixed orchard corset Highly rely on manual review Unfamiliar candidates The same candidate appears in multiple categories
- 8. Spark for Metadata Discovery Who We Are Metadata Discovery and Challenges Spark Solution Summary Challenges in Metadata Discovery Term Site Categories scott james US Men’s Clothing: Blazers & Sport Coats Men’s Clothing: Pants Men’s Clothing: Casual Shirts Men’s Clothing: Dress Shirts tiella US Chandeliers & Ceiling Fixtures Lighting Parts & Accessories turf US Sports Mem, Cards & Fan Shop: Cards: Football turf UK Collectables: Cigarette/Tea/Gum Cards: Cigarette Cards: Other Cigarette Cards
- 9. Spark for Metadata Discovery Who We Are Metadata Discovery and Challenges Spark Solution Summary Data Driven Approach for Brand Discovery Utilize seller input item specifics Utilize supply demand signals from sellers and buyers Training data available from previously reviewed candidates
- 10. Spark for Metadata Discovery Who We Are Metadata Discovery and Challenges Spark Solution Summary Supervised Machine Learning Approach Data : 35,000 previously human reviewed metadata candidates Feature : supply and demand signals Prototypes with Python Scikit Logistic regression, gradient boosting trees, random forest etc Random forest F1 0.878
- 11. Spark for Metadata Discovery Who We Are Metadata Discovery and Challenges Spark Solution Summary Production Development In the past, train offline and implement prediction component on production File transferring and configurations are time-consuming Dev Local Production
- 12. Spark for Metadata Discovery Who We Are Metadata Discovery and Challenges Spark Solution Summary Spark and Spark MLlib MLlib GraphX Spark provides powerful data processing APIs MLlib is a comprehensive machine learning package powered by Spark Regression, classification, clustering, dimensionality reduction etc Efficient development with local model, and flexible file access
- 13. Spark for Metadata Discovery Who We Are Metadata Discovery and Challenges Spark Solution Summary The Machine Learning SystemSpark Spark MLlib Feature Training Prediction
- 14. Spark for Metadata Discovery Who We Are Metadata Discovery and Challenges Spark Solution Summary Model Training with MLlib val pipeline = new Pipeline () .setStages(Array(labelIndexer , featureIndexer , rf , labelConverter )) val evaluator = new MulticlassClassificationEvaluator () .setLabelCol("indexedLabel") . setPredictionCol ("prediction") val cv = new CrossValidator () .setEstimator(pipeline) .setEvaluator(evaluator) . setEstimatorParamMaps (paramGrid) .setNumFolds (5)
- 15. Spark for Metadata Discovery Who We Are Metadata Discovery and Challenges Spark Solution Summary Evaluations Model F1 Python Scikit prototype 0.878 MLlib local 0.865 MLlib Hadoop (200 executors, production) 0.862 MLlib Hadoop (400 executors) 0.861 MLlib Hadoop (50 executors) 0.857 MLlib Hadoop (2 executors) 0.862 The performance variations among implementations are acceptable
- 16. Spark for Metadata Discovery Who We Are Metadata Discovery and Challenges Spark Solution Summary Speed Stage Data Size Time Feature Generation 1.73 Billion 6 min Train 33,000 8 min Prediction Input 650,000 4 min Capable of running the job daily Speed up the metadata discovery process, from months to days
- 17. Spark for Metadata Discovery Who We Are Metadata Discovery and Challenges Spark Solution Summary Newly Discovered Brand Brand Probability Milkies 0.84 BEABA 0.85 OXO 0.83 Lorex 0.87 Plan Toys 0.85 Safety 1st 0.82 Blabla 0.81 Combi 0.88 Graco 0.88 TotsBots 0.85 Realtree 0.85
- 18. Spark for Metadata Discovery Who We Are Metadata Discovery and Challenges Spark Solution Summary Summary Spark enables fast iterations of ML application development MLlib is comprehensive, and well integrated with Spark framework Dev and test locally, straightforward production deployment Compact code : 600 lines Need better understanding of the ML algorithm implementations in MLlib
- 19. Spark for Metadata Discovery Who We Are Metadata Discovery and Challenges Spark Solution Summary Acknowledgement Thejas Durgam Anu Mandalam Meital Tahar Zahav & eBay SDO Team Jean-David Ruvini
- 20. Spark for Metadata Discovery Who We Are Metadata Discovery and Challenges Spark Solution Summary Thank You! Qing Zhang, qzhang12@ebay.com Rui Li, ruili1@ebay.com