The traditional collaborative filtering based approaches have certain lacunae like their inability to handle sparse data, cold-start and lack-of scalability when there are millions of items and/or users. The content based recommendation engines overcome cold start, but have issues in taking user feedback into account. Hybrid recommendation engines try to get the best of both worldds. We outline the embeddings based approach to build deep learning based hybrid recommendation systems in TensorFlow.
We outline how deep learning can be used to extract features of images, product meta-data (or domain ontology) and convert these into embeddings. The text+image embeddings plus the embedded latent features of both items and users (meta-data of users, including browsing and purchase history) is combined with a feed-forward deep learning network.
This is a short form of the three hour tutorial we gave at Strata Data conference in California in March 2018: https://conferences.oreilly.com/strata/strata-ca/public/schedule/detail/63818
Our code has been open sourced at: https://pypi.org/project/tfrecommender/
Note: This workshop requires cloud credits and cloud providers will be collecting participants data for generating coupons. These credits are exclusively for hands-on labs. We will be opening up RSVP for this workshop shortly.