Visualizing Matrix Factorization Using Self-Organizing Maps

My Take

This is a more detailed discussion of the matrix factorization approach that is usually used in recommender systems. For Zillow, discoverability is so important to the health of their business because housing is such a personal (and major, financially speaking) decision. This piece also discusses an added layer of complexity unique to their product – geography. Any approach for generating recommendations has to take into account the locality of where the user is searching. A great house that has everything the user wants but is located across the country is not really that great of a house, is it?

Their Take

One of the core methods used within Zillow’s home recommendation engine is collaborative filtering. There are several ways collaborative filtering can be implemented. Due to the lack of explicit user ratings here at Zillow, we use matrix factorization with implicit feedback. This method starts by constructing user-item interaction matrix where each entry expresses our confidence that user finds that particular item relevant. Note, that in our case items are individual listed homes. Next, we factorize this matrix into a low dimensional embedding of both users and items which together can be used to reconstruct the original interaction matrix.

https://www.zillow.com/data-science/visualizing-matrix-factorization/