Facebook paper 2019

Deep Learning Recommendation Model for Personalization and Recommendation Systems

Brief Introduction

Two primary perspective

• Recommendation systems
  • Content filtering → collaborative filtering → Neighborhood method
• Predictive analytics
  • Statistical models to classify or predict the probability of events
    • From simple regression to models with deep networks
  • With embedding,which transform one-hot vectors into dense representations

Propose a personalization model with such two perspectives

• Uses embeddings to process sparse features into dense features
• Interacts these features using the statistical techinques (Factorizaion Machine)
• Finds the event probability with MLP

Background Concept

Embedding

• An embedding is a mapping of a discrete variable to a vector of continuous numbers
  • Embeddings map each category to a dense representation in an abstract space
  • One-hot or multi-hot vector to dense representation
• Dot-product of two shows similiarity

• Verification of embeddings

Matrix Factorization

MF is a simple embedding model — google developer page