What Are Embeddings?
Embeddings are dense vector representations (typically 64-512 dimensions) that capture semantic meaning.
**Key insight:** Similar entities have similar embeddings. In a well-trained embedding space:
- Users with similar preferences are close together
- Items that appeal to similar users are close together
- User-item proximity indicates relevance
**Learning embeddings:**
- Train on interaction data (clicks, purchases, ratings)
- Optimize so positive interactions have high similarity
- Negative sampling ensures non-interactions have low similarity
The magic: embeddings capture patterns that explicit features miss. A user embedding might encode "prefers indie films with strong female leads" without anyone defining that category.