Item2Vec
Word2Vec applied to item recommendations
Just as Word2Vec learns meaning from word co-occurrence in sentences, Item2Vec learns relationships from item co-occurrence in user behavior sequences.
If the pattern "people who viewed A also viewed B" repeats, vectors for A and B converge. This is the core collaborative filtering signal — and it works without an explicit user-item matrix.
Simple to implement
Feed item sequences into gensim's Word2Vec module. A prototype takes about 10 lines of code.
However, it doesn't fully leverage sequence order. For order-sensitive recommendations ("what to watch next?"), GRU4Rec is a better fit.
How It Works
Collect per-user behavior sequences (purchase, click, view)
Feed sequences as "sentences" into Word2Vec
Train with Skip-gram + Negative Sampling
Recommend via cosine similarity on learned vectors
Pros
- ✓ Extremely simple to implement (10 lines with gensim)
- ✓ Partially handles cold start items
Cons
- ✗ Does not fully utilize sequence order
- ✗ Focuses on item similarity rather than user personalization