​​Recommender Project Overview: 

​
​Hybrid ALS-DNN Model:

​​

​
​Overview

  • ​​The whole system consists of different models each suitable for a different situation and it decides automatically based on some parameters which to use the most
  • ​​This model
  • ​​Both ALS and DNN have their own pros and cons but in a complementary way. Using them together can be a way of solving popular recommender system challenges.
  • ​​Preprocessing
​​
  • ​​data loading and cleaning
  1. ​​drop duplicated records
  1. ​​dealing with missing values and outliers
  • ​​feature encoding
  • ​​
  • ​​matrix creation
  • ​​Candidate generation using WALS
​​
  • ​​demeanifying the matrix (takes care of sparsity)
  • ​​creating item and user embedding
  • ​​Uses user embedding to calculate item embedding and vice versa.
  • ​​candidate generation solves scalability issue
  • ​​WALS can be parallelized
  • ​​Automated Bayesian HP tuning
  • ​​Candidate ranking using DNN
​​
  • ​​adding item features
  • ​​semantic item embeddings are fed to the network as input.
  • ​​ranking using softmax model
  • ​​The trained model gives a prediction for every user-item pair to say if the item should be recommended or not.
  • ​​complex and non-linear feature extraction
  • ​​“deep and wide” implementation
  • ​​The theory states that “wide” linear models can effectively
  • ​​memorize sparse feature interactions using cross-product feature transformations, while “deep” neural networks can generalize to previously unseen feature interactions through low dimensional embeddings.
  • ​​In practice Google used this technique in Google Play store and saw improvement in users feedback.
  • ​​
  • ​​Bayesian HP tuning
​​
  • ​​Postproccessing
  • ​​removing repeated items
  • ​​re-ranking and choosing final recoms
​​

​
​Challenges and solutions

  • ​​cold start problem
  • ​​create average users
  • ​​using ALS
  • ​​matrix sparsity
  • ​​demeanify the matrix
  • ​​scalability
  • ​​candidate generation
  • ​​parallel computation
  • ​​using “jit” compiler
  • ​​ALS is an offline algorithm so the embeddings can be stored statically
  • ​​variable input size
  • ​​using embeddings from MF instead of one-hot encoded input
  • ​​diversity
  • ​​The result of the model is mixed with the results produced from other models such as pure CF or NCF to create final results
  • ​​including side features
  • ​​Using DNN
  • ​​exploration/exploitation trade-off
  • ​​Randomness is added so that the model can explore different niche items

​
​Questions to ask

  • ​​can we use the trained model for every data or should we retrain? 
​​
​​