Currently i am working on a recommendation problem , I dont want to recommend same article to user that i have already recommended before in past / recently . (Here i dont have any table which keeps track of what i have recommended to user already).Is there any specific way to manage this , atleast my recommender should not recommend recent artiles which he/she has read in last 1-2 months .
2 Likes
You could use something like a bloom filter to filter out results in a fairly space efficient manner. It would usually be implemented as part of your sampler. Or as a separate step after scoring. You can just filter out results if you’ve already seen them.
Here are some notes I have on Bloom filters:
- Bloom Filters for System Design Interviews | by SystemDesign | Tech Wrench | Medium
- mmh3 · PyPI and & and | bitwise operations.
- Would storing a fairly large ~kilobyte AttributeValue - Amazon DynamoDB bloom filter per user be accurate enough?
or see the example in this paywalled site instead:
Also, I’ve read about (but never implemented) fancier alternatives to Bloom filters. These can be set up to work with a timeout.
An aside:
I also recall a discussion in the Fast AI course about something similar to this. It was part of a discussion of the amazon store’s recommendations. It’s not exactly what you are talking about, but the point is relevant. For some recommenders, you don’t want to show recommendations that the user is probably already aware of. If I recall correctly, the example went something like this. A book by the same author that you just read is probably not the correct recommendation. A better recommendation might be a book in a similar genre. So in addition to filtering
1 Like