Explore a comprehensive conference talk on machine learning applications at Amazon, delivered by Rajeev Rastogi. Dive into various ML use cases including address quality improvement, product packaging optimization, substitute recommendations, demand forecasting, and product classification. Learn about key challenges in developing question-answering bots for product pages and techniques for learning semantically rich representations. Examine approaches to detect and correct product catalog defects, and understand the size recommendation problem using Bayesian modeling. Gain insights into leveraging customer and product features, incorporating customer personas, and the benefits of Bayesian inference in e-commerce applications.
Machine Learning at Amazon by Rajeev Rastogi
International Centre for Theoretical Sciences via YouTube
Overview
Syllabus
Start
Machine Learning @ Amazon
Numerous ML Applications
Address Quality
Product Packaging
Product Substitutes
Product Recommendations
Product Demand Forecasting
Product Classification
Product Matching
Insights Extraction from Reviews
Outline
Amazon Product Pages
Question & Answering Bot
Product Feature Questions
Product Comparison/Compatibility Questions
Key Challenges
Learning Semantically Rich Representations
Results for Different Loss Functions
Qualitative Results
Learning Representations with Attention
Amazon's Product Catalog
Title Defects
Image Defects
Product Attribute Mismatches
Text Attribute Extraction
Image Classification/Attribute Extraction
Mismatch Detection
Size Recommendation Problem
Motivation
Our Approach
Our Approach Contd
Bayesian Modeling Benefits
Intuition
Data Likelihood
Generative Model
Bayesian Inference
Polya-Gamma Augmentation [Polson et al. 2013]
Polya-Gamma Augmentation Contd
Gibbs Sampling Algorithm
Predictive Distribution
Experimental Results
Leveraging Customer and Product Features
Incorporating Customer Persona
Summary
Q&A
Taught by
International Centre for Theoretical Sciences
