Day #181 - Neural-network field aware factorisation machines for online-behaviour Prediction - Gunjan Sharma - Very Good Session

Key Lessons

• CTR - Click through rate (click or Not)
• CVR - Click View Rate (Install or not) 

Existing Approaches

• Linear Models, Feature Engineering
• LR good at generalization
• Decision Trees rigid about combinations
• Difficult for hand crafted features, Moved to Neural Network as they can find better features
• Categorical features handling, Neural Network for Regression problem

Factorization Machine

• One hot Vector representation
• Vector dot product and addition of the values
• PublisherVector*AdvertiserVector + AdvertiserVector*GenderVector + GenderVector*PublisherVector
• Weighted RMSE loss function
• Learn all the features using NN

Example cases

• Movie, Features
• Cross product of features
• Works better than LR

Field Aware Factorization Machine

• K dimension representation for cross features
• Capture second order interaction, Given more degrees of freedom
• Could not beat tree based model

DeepFM

• Neural Networks + Factorization machines
• Dense Embedding (Latent Features)
• Create the vector
• Pass through network
• Factorization - Second Order, NN - Higher order interactions

Somewhere I felt this like the cosine similarity based feature engineering :)

Neural Factorization Machine

• K Vector representation
• Change formula for vector dot product and vector summations
• Coming up with newer way of representing relationships matters
• Deep FFM has production latency issues, training time
• More Hidden layers - Better predictions

Lesson - Try Multiple techniques of feature engineering, Work on trying out new concepts in research papers, collaborate and keep learning, nn is the way to go forward..

Happy Mastering DL!!!