Great work! Thank you for sharing some insights into the personalization of NYT content.

Did you try using image embeddings to find similar recipes? It makes sense that the NYT would have expertise in word embeddings, but I would have thought image embeddings of a recipe (e.g. the thumbnails) would be as informative as word embeddings and could also be implemented with pre-trained models. After all, an average recipe is probably a few hundred words and a picture is worth a thousand.

I am also curious why the reward function was changed in the multi-armed bandit approach to promote the diversity of recommendations. Wouldn't the "exploration" part of the multi-armed bandit algorithm accomplish this?

Thanks again for sharing this piece