Home

Implementation details

embedding layer

• input -> embedding is a fully connected layer
• we build 4 embeddings, user-MLP, user-GMF, movie-MLP, movie-GMF
• if MLP is 3 layers, then embedding length is 2x predictive factors (see [evaluation](### evaluation))

MLP

• ReLu activation
• each higher layer has 1/2 the number of units
• 3 hidden layers
• last layer is lenth of predictive factors
  • each earlier layer is 2x the previous

GMF

total system

• pretrain GMF and MLP
• ADAM during pretraining, vanilla SGD after

evaluation

• leave one out evaluation
• randomly sample 100 items not interacted by the user, and rank the test item among 100 items
• Hit rate and NDCG
• random sample 1 interaction for each user as validation data, and tune hyper-parameters on it
• sample four negative instance per positive instance
• last layer of NCF = predictive factors, evaluated at [8,16,32,64]