E4571: Columbia University, Fall 2018
Data Science Institute
Industrial Engineering and Operations Research

Instructor: Brett Vintch PhD

with guest lectures by:
Sam Garrett | Senior Data Science Engineer at iHeartRadio
Brian Quinn | Lead Data Science at Comcast Cable

Course Description

Personalization is a key tool for enhancing customer experience across industries, thereby driving user loyalty and customer value. It is therefore no surprise that creating and enhancing personalization systems is also increasingly one of the core responsibilities of data science teams, and a key focus for many of the machine learning algorithms in the sector.

This course will focus on common personalization algorithms and theory, including behavior-based and content-based recommendation, commonly encountered issues in scaling and cold-starts, and state of the art research. It will also look at how businesses use, and misuse, these techniques in real world applications.

Prerequisites

Math: Linear algebra preferred, but not required
CS: A scripting language, preferably Python

Syllabus

Part I - Foundations of recommendation & personalization

Background and key questions

• History
• Evaluation - how do we know when we’re successful?
• Questions to keep in mind throughout course
  • The filter bubble
  • Rich get richer (popularity breeds popularity)
  • Cold start (users and/or items)
  • Accuracy
  • Interpretability
  • Scalability
  • Real-time vs batch
  • Serendipity
  • Diversity
  • How to incorporate new types of data
    • User or Item features
    • Real-time user context
  • Time-variance
    • User states or attribute drifts
    • Positive feedback loop
  • Aspirational vs actual (stated vs revealed)

Behavior-based personalization I

“Users with history like yours also like … “

• Nearest neighbors
• Collaborative filtering
• Matrix Factorization
  • Explicit
  • ALS

Part II - Advances in recommendation

Objectives, Evaluation, and Extensions

• Error metrics
• Ranking metrics
• Techniques for missing data as unobserved data
• Regularization

Behavior-based personalization II

• Approximate nearest neighbors
• Collective factorization & factorization machines
• Incorporating metadata and additional item or user dimensions
• Restricted Boltzmann Machines & Auto-encoders
• Bayesian approaches

Content based personalization

“Users that like content that [looks/sounds/reads] like this also might like…“

• Domains:
  • Semantic
  • Image
  • Audio

Hybrid approaches

• Ensembles
• Explicit combinations
  • Wide and deep

Part III - The personalization frontier, & personalization in the wild

• Active learning
• Reinforcement learning
• Context: time and location sensitivity
• Recommender rationale (e.g. tagging)
• Deep learning and neural network approaches

Evaluation

• Homework 1 (25%)
• Homework 2 (25%)
• Final project (50%)