Skip to main content

User-Item (Content Similarity)

Description​

The User-Item Content Similarity policy computes scores based on the direct similarity between a user's content embedding and an item's content embedding. It requires pre-computed embeddings based on user attributes and separate pre-computed embeddings based on item attributes. It does not pool embeddings based on interactions but relies on inherent alignment between user and item features (e.g., user interests vs. item tags).

Policy Type: user-item-content-similarity Supports: scoring_policy

Configuration Example​

scoring_policy_user_item_content_similarity.yaml
policy_configs:
  scoring_policy:
    policy_type: user-item-content-similarity
    user_feature_names: ["user_interests"]    # Name of the user features to compute similarity with
    item_feature_names: ["item_tags"]         # Name of the item features to compute similarity with
    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).