ฝึกพูดภาษาอังกฤษด้วยเทคนิค Shadowing จากวิดีโอ: Recommendation System Infra Basics 1

Hey everyone, we're trying something new with shorter form content to teach you important concepts.

This is a bit different than our longer form problem breakdowns or deep dives, but I'm hoping you learned something new and it helps you to grow as an engineer.

Let us know what you think in the comments if we should be doing more of this.

Today, I want to make a quick video to walk you through some of the architecture behind recommendation systems.

This isn't going to be an ML focus video, we'll get to those later, but I wanted to share a few of the techniques used in modern systems

because they end up being broadly useful in designing systems that both operate at scale and don't necessarily have correct answers.

Just to be clear, I'm focusing on the software engineering side here.

If you think about recommendation systems, they usually have a pretty simple objective.

You have a user and you want to find all the things that we might think the user is going to want.

This might be YouTube videos or a buyable item on an e-commerce platform or social media posts or whatever.

There's really two big challenges here.

First, how can we know what the user wants?

That seems ill-defined and might require a bunch of different input.

What do we know about the items?

What do we know about the user?

What about their relationships?

Fundamentally, this is an ML problem and I'm going to skip it for now

because it's not the thrust of what we want to focus on for this video.

But the second item is pressing.

How do we make this system scale?

If we have billions of videos or trillions of posts, how can we possibly find the one the user actually wants in a reasonable amount of time?

This seems like an especially hard needle in a haystack problem, and it is.

But this isn't about making machine learning models run fast.

There's more to it.

Let me illustrate the problem quickly for video recommendations and then we'll move on to some solutions.

If you were to approach this problem, you might start with a blackbox ml function, which takes a user and a video and tells me how likely the user is to watch the video.

In this case, my black box outputs a score after several milliseconds.

This is an important building block which will tell me that a given user will hate this particular cat video, but love Evan's Ticketmaster breakdown.

Great!

With this function available, I can now try to make my recommendation system.

Let's apply this function to every video in the catalog.

This will give us a score for every possible video.

We can then sort the scores and return the video which is most likely to be watched.

But wait, billions of videos means I need to run through my ML black box billions of times.

This means I need a ton of time, maybe two weeks to figure out the best video for them to watch.

Crap, that's not going to work.

Since this is really slow, maybe we take it offline.

Let's pre-calculate our recommendations for each user and then when they come online we can feed them out of a cache.

Easy, right?

No, if we have billions of items and billions of users, suddenly we need a data center the size of Texas in order to make recommendations.

This is untenable and it's slow.

I don't want to wait two weeks after I register for my TikTok videos to start streaming in.

Okay, okay, you get the gist.

So we need to do a little trickery here and the naive approach won't work.

Modern recommendation systems really have three stages.

Candidate generation, ranking and re-ranking.

What we've been talking about so far is ranking.

Taking a bunch of videos and applying an ML model to score them, which results in an ordering for us to return to our clients.

But ranking can't really scale to billions of possibilities, as we've just seen.

It's just too time consuming and resource intensive.

We need a better way.

Enter candidate generation.

Candidate generation is a simple idea.

We know some classes of videos are much more likely to be watched than others.

As an example, if I'm subscribed to MrBeast, it's likely that I'm going to watch other videos he creates.

Similarly, the most viewed videos on the platform have very broad appeal.

If you wanted a dead simple recommendation system, you'd just take the top videos in the platform and feed those to users on a loop.

That's still how radio works today.

But we can do a bit better now that we have our ranker.

We'll take each of these heuristics that we just talked about and call them candidate generators.

A candidate generator will take in some information about the user to generate a list of candidates

or videos we think the user will want to watch.

We can union together all of the candidates from each of the candidate generators.

Then we'll take our ML function we had earlier, which we now call a Ranker, to score each of these candidates.

Since we now have order a thousand rather than a billion items, this is much more tractable.

The videos that emerge with the highest scores are the ones that we want to recommend.

And as long as our candidate generators are reasonably exhaustive, this approximates the much more expensive system.

I'll say it again for emphasis here.

we're taking some heuristics to cut the problem down and make it more feasible.

We don't have to rank every video ever to get good results.

That's the essence of candidate generation.

Before we move on, there's one other type of candidate generator which is worth discussing.

The top videos on the platform or videos from channels I've subscribed to are pretty easy to get.

We can do this with basic database queries.

But how can I build candidate generators for things like a video similar to the one I just liked

or videos watched by users like me?

The answer to this is to use what are called embeddings or vector of numbers.

I'm going to simplify a bit and invoke the black box of ML here.

But let's pretend we have a function which takes in an input, in this case a video, and outputs an embedding or a series of numbers.

These numbers are special in that videos with smaller distance, you could think of this like distance in 3D space here, have some property.

It could be that we've trained our model, our black box, so that videos which are liked by similar users are clustered together.

Or we could try to group them semantically based on their content, like I've done here.

Or we could try to find videos which irritated the same people, the options are literally What we can see is that the cat video and dog video have smaller distance.

If you calculate the Euclidean distance in 3D space, then the cat video and a chess match,

they aren't as semantically similar as the cat video and the dog video.

Just like before, if we want to find the smallest distance, our most naive approach is to calculate the distance between our input video and every other video in the catalog.

This is way too slow, but it's going to find the video that is most semantically similar or liked by similar users or whatever.

A more efficient way to do this is we can generate embeddings for our entire catalog of videos, including the new ones that are created.

Once we have embeddings, we store them in a vector database.

Popular options here include commercial offerings like Pinecone or open source solutions like Face or In this case,

I want all my videos together with their embeddings in this database.

Vector databases are optimized with algorithms like hierarchical navigable small worlds, which enable very fast approximate nearest neighbor search.

Basically, they're much, much faster than trying to calculate distances on each item.

With my database constructed, I can query it.

So for videos similar to the one I just liked,

I'll grab the embedding for that video I just liked and query the database for the nearest neighbors of that embedding.

This will return to me the IDs of all the videos which have embeddings with small distance to my input.

In effect, my database now answers which videos are similar to this one.

Those are my candidates, which I can feed to my ranker just like I fed the top 100 most viewed platform videos.

100

I can have separate tables or databases for each style of embedding.

101

So I can have a separate database of users to answer questions about videos liked by users similar to me.

102

It's common inside ranking and recommendation systems to have dozens of these indexes

103

or candidate generators and managing them is an important problem for the software engineers working on these systems.

104

That's candidate generation.

105

Now that we've completed candidate generation and hand-waved past ranking, we'll get to that later, let's talk briefly about re-ranking.

106

There are a lot of behaviors of recommendation systems that we want to tweak.

107

We probably want to give new videos a boost so

108

that they get a chance to get exposure and for us to learn about them.

109

This is fundamental to the trade-off between exploration and exploitation.

110

We not only want to show you the video you're most likely to like, but we also want to expand your horizons a bit.

111

We also want to make sure we're not recommending videos that a user has blocked or they might be sensitive to.

112

Adjusting the full recommendation stack for each of these use cases would be really expensive.

113

If we needed special indexes that don't contain videos from creators the user is blocked, that candidate generator is much less efficient.

114

Instead, in the re-ranking phase, we're going to apply tweaks post hoc.

115

If we need to provide the user with 10 ordered recommendations,

116

our ranking stage might produce 25 and our re-ranking stage can reorder or remove based on business rules,

117

privacy settings, etc. This is actually an implementation of a very common pattern you'll see in real-world system design

118

and it applies to search engines, privacy solutions, ranking systems and more.

119

Sometimes it's easier to design for the 99% case

120

and have a layer to fudge the final results

121

and make them how you want rather than build a system which solves 100% of the problem.

122

I'm going to stop there from going too long.

123

There's obviously a lot more we can get into.

124

Let me know if you want to see more in the comments.

125

But just to recap some key points here, we've covered how a multi-stage architecture, in this case using candidate generation,

126

enables scalability for massive systems and makes seemingly impossible problems possible.

127

Next, we showed how you can use embeddings and vector search to solve fuzzy problems cheaply.

128

And finally, we showed how production systems are often layered with fine-grained adjustments applied at the last level.

129

This makes problems more tractable.

130

I hope you found this video valuable and if you'd like more please subscribe.

131

Otherwise until next time talk to you later.

📺 ช่องเดียวกัน

Elasticsearch Deep Dive w/ a Ex-Meta Senior Manager for System Design Interviews

University StudentsBusiness English+2

API Design in System Design Interviews w/ Meta Staff Engineer

University StudentsBusiness English+2

Redis Deep Dive w/ a Ex-Meta Senior Manager

University StudentsBusiness English+1

Consistent Hashing: Easy Explanation for System Design Interviews

University StudentsWorking Adults+1

✨ วิดีโอแนะนำ

126 ประโยค

How I use AI as a senior software engineer and 5 principles of AI explained in 10 minutes

University StudentsBusiness English+1

34 ประโยค

Marketing Research Essentials Intro, Models & Intelligence

University StudentsBusiness English+1

405 ประโยค

Netflix product manager mock interview and tips (with ex-Netflix Lead PM)

University StudentsBusiness English+2

125 ประโยค

Introduction to statistical prediction models

University StudentsBusiness English+1

164 ประโยค

System Design Mock Interview: Design Facebook Messenger

University StudentsBusiness English+1

162 ประโยค

If I started a YouTube channel from scratch, I’d do this

University StudentsBusiness English+1

133 ประโยค

Data Ingestion Best Practices! How to Ingest Data Efficiently at Enterprise Scale!

University StudentsBusiness English+1

443 ประโยค

Complete Roadmap for Backend Software Engineers (START HERE!)

University StudentsBusiness English+1

237 ประโยค

How Social Media Algorithms Actually Work (And How to Beat Them)

University StudentsBusiness English+1

182 ประโยค

We Let $200 GPT-Pro Think for 7 Days on Unsolved Math

University StudentsBusiness English+1

47 ประโยค

Essence of linear algebra preview

en-USUniversity Students+2

18 ประโยค

SD91: Reliability vs Scalability

University StudentsBusiness English+1

เทคนิค Shadowing คืออะไร?

Shadowing เป็นเทคนิคการเรียนรู้ภาษาที่ได้รับการรับรองทางวิทยาศาสตร์ พัฒนาขึ้นสำหรับการฝึกนักแปลมืออาชีพ วิธีการนี้เรียบง่ายแต่ทรงพลัง: คุณฟังเสียงภาษาอังกฤษจากเจ้าของภาษาและพูดตามทันที — เหมือนเงาที่ตามผู้พูดด้วยช่วงเวลาห่าง 1-2 วินาที การวิจัยแสดงว่าเทคนิคนี้ปรับปรุงความแม่นยำในการออกเสียง ทำนองเสียง จังหวะ การเชื่อมเสียง การฟังเข้าใจ และความคล่องแคล่วในการพูดได้อย่างมีนัยสำคัญ

เทคนิค shadowing: อ่านคู่มือฉบับเต็มทีละขั้นตอน →

ฝึกพูดภาษาอังกฤษด้วยเทคนิค Shadowing จากวิดีโอ: Recommendation System Infra Basics 1

Proximity Search & Geospatial Indexes Explained

Distributed Transactions Explained: 2 Phase Commit vs Saga Pattern

Why Shopify Moved Inventory Reservations from Redis to MySQL

Concurrency in Low-level Design Interviews w/ Staff Engineer

CAP Theorem in System Design Interviews

API Gateways in System Design Interviews w/ Ex-Meta Staff Engineer

Message Queues in System Design Interviews w/ Meta Staff Engineer

Elasticsearch Deep Dive w/ a Ex-Meta Senior Manager for System Design Interviews

API Design in System Design Interviews w/ Meta Staff Engineer

Redis Deep Dive w/ a Ex-Meta Senior Manager

Consistent Hashing: Easy Explanation for System Design Interviews

How I use AI as a senior software engineer and 5 principles of AI explained in 10 minutes

Marketing Research Essentials Intro, Models & Intelligence

Netflix product manager mock interview and tips (with ex-Netflix Lead PM)

Introduction to statistical prediction models

System Design Mock Interview: Design Facebook Messenger

If I started a YouTube channel from scratch, I’d do this

Data Ingestion Best Practices! How to Ingest Data Efficiently at Enterprise Scale!

Complete Roadmap for Backend Software Engineers (START HERE!)

How Social Media Algorithms Actually Work (And How to Beat Them)

We Let $200 GPT-Pro Think for 7 Days on Unsolved Math

Essence of linear algebra preview

SD91: Reliability vs Scalability

สิ่งที่คุณจะได้เรียนรู้

ฟังเสียงเหล่านี้

พูดเหมือนเจ้าของภาษา

เทคนิค Shadowing คืออะไร?