You are a true gem, already been following lot of your contents in youtube. You articulate the system design cases and solutions so well with in-depth explanations of the technologies used.
Hi! what are the ranking models exactly taking in and putting out? I thought that from the retrieval, nodes containing the lists of candidates would have scores associated to each candidate based on the neighbors being close. And then filtering happens with bloom filters. thank u!
Hey Clara! So in the retrieval phase, we hit our embedding caches to find all of the possible suggestions that we may give to a user. Then we filter, to ensure that we don't show the user any content that they've seen, or shouldn't be shown to them for whatever reason. In the ranking phase, we have room to rank a more limited search space of potential items (because we limited them before) according to user specific criteria. If we retrieved items that already had a score, it would mean that we would either need a copy of the embeddings per user (probably impossible) or that we couldn't account for any user specific information when making our rankings (not ideal). Hope this makes sense!
This video randomly popped up in my recommendations (pun intended) and it was a very good recommendation. Really like the detailed explanation. Follower += 1
Have You thought about how you will work with a recommendation system if we have addition data on users such as age/ location/handset/Time based recommendation and their activity as well as adding different types of tags such as Content/ category/ genre/ actors/ director/ content creator etc.
How would you say this would deviate from our traditional recommendation system pipeline? At least to me, these just seem like more features with which to make decisions!
If the embedding model is trained offline, how will it recommend most up to date data? (ie a new video got created 10 mins ago) In that case, I'm a bit confused on how it's real-time? Also do we not need to use Spark Streaming anywhere in this design? Thanks!
It probably won't. That being said, you can combine those suggestions with things like subscriptions (see Twitter design) to deliver videos of creators that you subscribe to to your news feed
Why would the recommendation cache not grab data from our items db? isn't the hadoop file storage not preprocessed from our spark job yet, so that data isn't valuable for recommendations?
Well I would imagine we'll have to search the items DB when we look at our most recently interacted-with items, get their neighbors, and then fetch the items corresponding to those IDs, unless the items themselves are denormalized and stored in the neighbor index.
One more question: Does the embedding model + index only look at the item embeddings? (videos recently watched) or does it take user information as well (I set my preferences as liking action movies) Also how would it worked for a cold start user? (user does not have any "'last 5 videos watched" to get 1000 nearest neighbors?
Embedding model doesn't have any user info or else we'd have to duplicate all the embeddings, for the first open we can just show popular videos or something
Embedding model doesn't have any user info or else we'd have to duplicate all the embeddings, for the first open we can just show popular videos or something
The real time part is just asking for more videos. That being said, I suppose you could use stream processing to create embeddings of newly created videos on the fly and then include those as candidates for new suggestions.
@@Shufjskkskf Ah ok - just describes any changes made to the set of items that we could possibly be recommending (adds, removes, edits of an existing item) as resulting from user action
@@jordanhasnolife5163 could that be a good usecase to store it in a feature store? could the user activity also be stored in a feature store? in case of a music recommendation system are those infos regarding: song titles, text, images, artists, music genre, audio analytics?
You are a true gem, already been following lot of your contents in youtube. You articulate the system design cases and solutions so well with in-depth explanations of the technologies used.
Thanks Surendra ♥️♥️
I have been asked recommendation systems in system design interview because I worked in a team that does it. Really helpful
Hi! what are the ranking models exactly taking in and putting out? I thought that from the retrieval, nodes containing the lists of candidates would have scores associated to each candidate based on the neighbors being close. And then filtering happens with bloom filters. thank u!
Hey Clara!
So in the retrieval phase, we hit our embedding caches to find all of the possible suggestions that we may give to a user. Then we filter, to ensure that we don't show the user any content that they've seen, or shouldn't be shown to them for whatever reason.
In the ranking phase, we have room to rank a more limited search space of potential items (because we limited them before) according to user specific criteria. If we retrieved items that already had a score, it would mean that we would either need a copy of the embeddings per user (probably impossible) or that we couldn't account for any user specific information when making our rankings (not ideal).
Hope this makes sense!
@@jordanhasnolife5163Thanks for the answer + thanks for the help! love your videos.
This video randomly popped up in my recommendations (pun intended) and it was a very good recommendation. Really like the detailed explanation. Follower += 1
Thanks Rahul! Funny to see the system in action haha
Have You thought about how you will work with a recommendation system if we have addition data on users such as age/ location/handset/Time based recommendation and their activity as well as adding different types of tags such as Content/ category/ genre/ actors/ director/ content creator etc.
How would you say this would deviate from our traditional recommendation system pipeline? At least to me, these just seem like more features with which to make decisions!
If the embedding model is trained offline, how will it recommend most up to date data? (ie a new video got created 10 mins ago) In that case, I'm a bit confused on how it's real-time? Also do we not need to use Spark Streaming anywhere in this design? Thanks!
It probably won't. That being said, you can combine those suggestions with things like subscriptions (see Twitter design) to deliver videos of creators that you subscribe to to your news feed
Why would the recommendation cache not grab data from our items db? isn't the hadoop file storage not preprocessed from our spark job yet, so that data isn't valuable for recommendations?
Well I would imagine we'll have to search the items DB when we look at our most recently interacted-with items, get their neighbors, and then fetch the items corresponding to those IDs, unless the items themselves are denormalized and stored in the neighbor index.
One more question: Does the embedding model + index only look at the item embeddings? (videos recently watched) or does it take user information as well (I set my preferences as liking action movies)
Also how would it worked for a cold start user? (user does not have any "'last 5 videos watched" to get 1000 nearest neighbors?
Embedding model doesn't have any user info or else we'd have to duplicate all the embeddings, for the first open we can just show popular videos or something
Embedding model doesn't have any user info or else we'd have to duplicate all the embeddings, for the first open we can just show popular videos or something
Great video Jordan! Nicely explained.
hello.. all model training seems offline with the 2nd approach..so how is it real time?
The real time part is just asking for more videos. That being said, I suppose you could use stream processing to create embeddings of newly created videos on the fly and then include those as candidates for new suggestions.
Can you a video to talk more about KAFKA? why should we use it and when in SD?
I recommend watching my stream processing video - it's a log based message broker.
Hey just a silly question are you same guy in your dp , if yes you are a gem and a meme legend too
No I'm not haha, I just also enjoy the meme - it may be time for me to change it tbh
I saw comments and thought it's a good video but It was a huge waste of time. Maybe interesting for people who have Zero idea about tech.
Kek
You're the Yin to my Neetcode :)
Can you please elaborate more on the neighbouring index?? and how the vector model linked to this index? Anyways Great content (y)
It's a map from (vector) to [closest neighboring vectors]. Feel free to elaborate on your question and I'll do my best to help!
what are item change events?
Mind giving me a timestamp to reference?
@@jordanhasnolife5163 in the architecture diagram here at timestamp 306 sec
@@jordanhasnolife5163 top right in diagram at 5:13
@@Shufjskkskf Ah ok - just describes any changes made to the set of items that we could possibly be recommending (adds, removes, edits of an existing item) as resulting from user action
@@jordanhasnolife5163 could that be a good usecase to store it in a feature store? could the user activity also be stored in a feature store? in case of a music recommendation system are those infos regarding: song titles, text, images, artists, music genre, audio analytics?
I personally liked your ipad notes and prev method more than this.
My iPad notes is my current method - that's why I'm remaking them
great content!