Skip to main content

User (Content Similarity)

Description​

The User-Content Similarity policy implements content-based filtering focused on user attributes. It requires user embeddings based on their profile features (e.g., text embeddings from bios, categorical embeddings from demographics). An item's representation is dynamically computed by pooling the embeddings of users who have positively interacted with that item. A target user is recommended items whose aggregated user pool embedding is most similar to the target user's own content embedding.

Policy Type: user-content-similarity Supports: embedding_policy, scoring_policy

Configuration Example​

scoring_policy_user_content_similarity.yaml
policy_configs:
  scoring_policy: # Can also be used under embedding_policy
    policy_type: user-content-similarity
    pool_fn: "mean"         # Strategy for pooling interacting user embeddings ('mean', 'max', etc.)
    distance_fn: "cosine"   # Similarity metric ('cosine', 'dot')

References​

  • Lops, P., de Gemmis, M., & Semeraro, G. (2011). Content-based recommender systems: State of the art and trends. Recommender systems handbook, 73-105. (General overview).