2006 just called and wants its #RecSys models based solely on user and item IDs back! 🙃

Semantic IDs might be as revolutionary as I keep hearing (or they might not), but…I’m already tired of comparisons to a straw man of ID-based recommendation modeling back in “the bad old days.”

If you’ve spent the past decade or two having cold-start, sparsity, and generalization issues while wrangling enormous embedding tables learned from lowest-common denominator features like lists of interacted items, you missed a lot of practical industry knowledge about how to make recommenders that live up to people’s natural expectations by combining general purpose deep model architectures (e.g. DLRM, DCN) with specific domain knowledge (e.g. ontologies and taxonomies.)

For example: how to reflect features across the user/item boundary, how to enrich user and item representations with distributional features, and how to transfer learning between implicit and explicit preferences. We’ve been out here doing this stuff for a while now, and I don’t expect it to go away any time soon.

via Carl H. on LinkedIn