User Embeddings Part 3: Applications and Real-World Impact
User Embeddings Part 3: Applications and Real-World Impact
In the previous parts of this series, we understood what user embeddings are, how they’re created, and how models learn to represent users. Now, in Part 3, we explore how companies like TikTok, YouTube, Netflix, and even Amazon put user embeddings to work. We also look into advanced use-cases like user intent modeling, personalization, and ethical concerns.
How Companies Use User Embeddings
1. TikTok: The For You Feed
TikTok doesn’t just recommend content based on likes or hashtags. It builds a dynamic, ever-evolving embedding for each user. Every scroll, rewatch, pause, like, or skip becomes a signal.
The embedding is continuously updated through real-time user behavior, and a nearest-neighbor search finds videos with similar content embeddings. That’s how TikTok learns your humor, preferences, and even mood.
2. YouTube and Netflix: Multi-Tower Models
These platforms use multi-tower architectures where:
- One tower processes the user embedding
- One tower processes the content embedding
A dot product or similarity score helps rank videos/shows tailored to the user’s interest profile.
1
2
3
# Simplified PyTorch example
similarity = torch.matmul(user_embedding, item_embedding.T)
ranking = torch.argsort(similarity, descending=True)
This helps in surfacing “just right” content, even if the user hasn’t watched it before.
3. Amazon: Personalized Search and Ranking
Amazon learns not just what you search, but how long you browse, how you compare items, and what you buy (or abandon). These behaviors generate multiple embeddings:
- Shopping intent embedding
- Brand affinity embedding
- Price sensitivity embedding
These are used to personalize product ranking in real time.
Application: Zero-Click Personalization
Zero-click personalization means showing a user content without them having to search or click. Embeddings make this possible:
- Cold Start: New user embeddings are initialized via location, device, or demographic heuristics.
- Warm Start: A few sessions are enough to model behavior.
- Mature Embeddings: Long-term embeddings + short-term context embeddings provide optimal recommendations.
Beyond Recommendations: What Else Are They Used For?
- Ad Targeting: Embeddings help advertisers match users with high-conversion potential audiences.
- Fraud Detection: Sudden deviations in embeddings from normal behavior can indicate account compromise.
- Dynamic Pricing: Pricing models use embeddings to estimate willingness-to-pay.
Real-Time Adaptation
Some systems (e.g., Instagram Reels) update embeddings in real time. This requires fast approximate nearest neighbor (ANN) search, using tools like:
- Facebook FAISS
- ScaNN by Google
- Pinecone (vector DB as a service)
Ethical Considerations
User embeddings raise concerns about:
- Privacy: Are embeddings too detailed? Can they be deanonymized?
- Echo Chambers: Constant personalization can limit diversity of thought.
- Manipulation: Over-targeting can exploit vulnerabilities.
It’s important for ML engineers to log usage, monitor drift, and provide transparency to users.
Wrapping Up
User embeddings are how machines try to understand us, our likes, habits, needs, and emotions. They’re a double-edged sword: powerful for convenience and personalization, but also capable of subtle manipulation. As builders, we must handle them with care.
In Part 4, we’ll go deeper into building a recommendation engine using embeddings from scratch in PyTorch, covering contrastive learning, sampling strategies, evaluation metrics like NDCG, and more.
Suggested Reading: