Companion article
The Embedding Trap in Signed-Affinity Recommendation Systems
Interactive tool

Signed Affinity Recommendation Explorer

Compare naive text-embedding cosine similarity against signed weighted-sum matching. Adjust the user profile sliders to see where the two models diverge — and why the sign of an affinity score must live on the vector, not in the text.

Preset scenarios

User affinity profile

Product catalogue

Model comparison
Naive text embedding + cosine
Weighted signed vector sum + cosine