It doesn't, unfortunately. One way to circumvent the issue would be to compute content-based similarities for the new item against old items, then pick the most similar old item as a proxy for the new item. Now you can use the interaction history of the proxy item to to make recommendations based on the new item.
How would it work on new items? Embedding does not have information about them.
It doesn't, unfortunately. One way to circumvent the issue would be to compute content-based similarities for the new item against old items, then pick the most similar old item as a proxy for the new item. Now you can use the interaction history of the proxy item to to make recommendations based on the new item.