Building Neural Collaborative Filtering recommendation model

Key Takeaways

hi there welcome back to our video series of building recommendation systems with sensor flow my name is way and I'm a developer Advocate at Google in this video we'll be discussing a recommendation model neuro collaborative filing and showing you how to build such a model using tensorflow to core API we have covered a lot of ground in this video series but mostly the tools we discussed are high level apis which are made easy to use on purpose but there are some occasions when you really want to go below the surface and use a centerflow core apis to build a recommendation engine from scratch this sometimes makes sense and we're going to show you how today we're going to build a classical recommendation model called neurometrix factorization new MF new MF comes out of the 2017 research paper by H and his collaborators and it is an instantiation of neur collaborative FY it is a general framework for collaborative furthering of recommendations in which a neuron Network architecture is used to model user item interactions unlike traditional models umf does not resort to Matrix factorization with an inner product on latent features of users and items it replaces the inner product with a multi-layer pectrum that can learn an operatory function from data architecture wise new umf is an ensemble of generalized Matrix factorization gmf and multi-layer perception MLP gmf applies a linear kernel to model the lat feature interaction and MLP applies a nonlinear kernel to learn the interaction function from data new MF is a fused model of gmf and MLP to better model the complex user item interactions and unifies the strength of of linearity of MF and nonlinearity of MLP for modeling the user item latent structures new MF allows gmf and MLP to learn separate embeddings and combines the two models by concatenating their last hidden layers specifically starting at the bottom we have using embedding and moving embeddings from this embeddings we generate user and movie latent Vector for Matrix factorization and MLP layer separately after passing through the generalized Matrix factorization layer and MLP layer we concatenate the outputs of this layers and get the predicted Logics from a dense layer for computing the loss that's it that's the whole architecture we actually provide an excellent implementation of new MF in the tensor FL motor Garden for context tensorflow model Garden is a repository for reference model implementations for 4 2 and it includes more than 50 state-of-the-art models across Vision NP and recommendation engines torfl Motel Garden in a was taking full advantage of torl 2 for research and production it is a community-driven repository with growing contributions and it's a top 10 python package with most unique contributors in 2020 teslow Mel Garden has an implementation of this architecture let's walk through the code to understand how to build it with tlow to first we start with user and item input layers this is quite straightforward next we call Constructor model method to generate the logic by passing the user and item inputs we're going to Define Constructor model function next in Constructor model function we Define the user and item embeddings using the user and item inputs note the embedding Dimension we set as the second parameter this embeddings will be shared by both the Matrix factorization layer and MLP layers and we will extract user and item latent vectors from this embeddings next we're going to use several Lambda layers to slice the embeddings into user and item latent vectors to do the slicing we Define two utility functions MF slice Funk and MLP slice Funk functions as you can see the Matrix factorization lat vectors takes the first two mfdm elements from the embedding while the MLP lat vectors takes the remaining embeddings next for generalized Matrix factorization we're going to use the multiply layer to do element wise multiplication for MLP we're going to first concatenate the user and item latest vectors into a MLP vector and then pass it through several MLP layers lastly we concatenate The Matrix factorization vector and the MLP Vector into a single Vector predictive Vector here and generates the logic FR it using a des layer then we return the model finally we can compute the cross entropy loss from the predicted soft Max loges converted from the predicted loges so that's it that's all the code to build the newer collaborative feding model of course this is just the modeling part you still need to write code to handle data preprocessing model training performance evaluation and so on but hopefully the code walk through here helps you understand how to build a good neur collaborative feding model so to summarize today we walked us through how to build a neuro collaborative fed model with tensor 2 API you can check out the original research paper and the code in tensor modor Garden for additional details with that thank you for watching this video please stay tuned for more updates from us [Music] [Music]