Building Recommender System with the Two-Tower Architecture
The repo provides boiler plate code for building recommender systems utilizing the Two-Tower-Architecture. The module that enables this is the freshly-backed tensorflow-recommenders, which curates all the necessary methods based on recent research conducted by Google researches. The approach takes with focus on "candidate retrieval", which revolves against (in-batch) softmax and adapts a framework of a categorical problem. Additionally, an extension is possible to "candidate rating" would rather require numerical expression of query-candidate affinity like star ratings or other means of numeric preferance expression.
The model was build using proprietary data on beauty products and is not sharable. The bc can be adopted when working on similar projects. The presentation of the project can be found under the link. https://slides.com/kirillkasjanov/recommender-systems