Scalable Machine Learning Pipeline For Meta Data Discovery From eBay Listings
•
4 likes•1,381 views
The document describes eBay's use of Spark and MLlib to build a scalable machine learning pipeline for metadata discovery from product listings. eBay developed supervised machine learning models using Spark MLlib to classify product listings and discover new metadata brands with high accuracy. The models were trained on historical metadata and product data, and can process large amounts of data quickly to discover new metadata on a daily basis, speeding up metadata discovery from months to days. Spark MLlib allows for efficient model development locally and simple deployment to production at scale.
1 of 20
Download to read offline
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
More Related Content
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