# Production Challenges
# Evolution of Architectures
Before 2006: Content-based filter, Collaborative filter
2006: Factorization-based Methods
2009: Learning to Rank; Gradient Boosting Decistion Tree (Pairwise Ranking)
2015: Wide & Deep , DeepFM/xDeepFM , DIN/DIEN , DLRM ,DCN
# Pros vs Cons
Unique challenges
- Large Embedding Tables
- Model Heterogenity : MLP, Transformers
- Performance Variance: recommendation context, - model colocation
Pros:
- Storage Capacity much larger
- Compute Intensity is lower than CNN
- Sparse Irregular memory accesses
200 different recommender models running concurrently in FB
Three Types:
- Embedding Dominated
- MLP Dominated
- Attention Dominated
