Item-Based KNN
Recommend items similar to those a user has previously interacted with
Item-Based KNN computes pairwise item similarity from the training interaction matrix. At inference, it recommends items that are most similar to the items a user has already engaged with.
When to use:
- Recommendation where items have stable similarity (product catalogs, media libraries)
- "Users who bought this also bought…" style recommendations
- When user preferences are better modeled through item relationships
Input: User-item interaction history Output: Ranked list of similar items per user based on past interactions
Model Settings (set during training, used at inference)
K (N Neighbors) (default: 40) Number of similar items to consider per recommendation. More neighbors produce more diverse but potentially less relevant recommendations.
Similarity Metric (default: cosine)
Metric for item-item similarity. cosine is standard for sparse interaction data.
Min Support (default: 1) Minimum number of shared users between items to compute similarity.
Inference Settings
No dedicated inference-time settings. Items are ranked by aggregated similarity to the user's interaction history.