WRadoslaw
commented
9 months ago State of Recommendations on 30/01/2024 - 3pm
Orion
Already done
GraphQL server:
- homepage resolver ✅
- similar videos resolver ✅
- channel follow handler ✅
- new query to get the next video (after the current one ends - this will require a new scenario in Recombee) ✅
Auth API:
- syncing new users ✅
Interactions API:
- limits requests per IP ✅
- limits requests per single item interactions ✅
- supports: click, video portion, video consumption, video rate ✅
- channel click handler ✅
Processor:
- adds new and updated videos to the provider ✅
- adds new reactions to the provider (only if the user exists in DB) ✅
- has a flag to sync from 0 or only after the export block ✅
- adding channels to the provider ✅
To be implemented
GraphQL server:
- new query to support personalized search
Processor:
- creating a service for language recognition based on the title and description
TBD:
- low prio - cron jobs that will periodically check Recombee for empty items (synced by
createCascade) fill them up or delete them. The same should be done for users but in reverse order, but only if Orion schema allows the user ID to be empty. Currently, in Orion, we don't clean up anonymous users if they are inactive for any time, it would be good to introduce such a thing and remove such users from Recombee as well.
Atlas
Already done
- homepage feed ✅
- video view - similar feed ✅
Not yet implemented
- watch next - at the end of the video
- personalized search
- channel interactions
- low prio - make all of the above optional
An important part of all apps will be to power recommendations. Orion, as the primary data layer and API for apps will need to, in some way, accommodate recommendation engine integrations. I have limited experience on this topic, but doing a bit of research I found that Tensorflow Recommenders seems like a really good fit, as an end-to-end ML powered recommendation framework, for use-cases very similar to Atlas. So keeping in mind how to accommodate recommendations, ideally through some flexible abstraction that allows people to use their own approach, but where this frame is taken as a sort of benchmark framework. Here is a really exciting playlist that takes you through how it works: https://www.youtube.com/playlist?list=PLQY2H8rRoyvwLbzbnKJ59NkZvQAW9wLbx