User Profile Construction
What you'll learn: How to create a user preference profile by summarizing features from items a user has liked or interacted with.
The Core Idea
Once you've represented items as feature vectors (as in Item Feature Representation), the next step is to represent users in the same feature space. User profile construction aggregates the features of items a user has positively interacted with—watched, clicked, purchased, rated highly—to create a single vector that represents that user's preferences.
Think of it like creating a "composite portrait" of a user's tastes. If you've liked five sci-fi movies with futuristic settings and strong female leads, your profile should highlight those features. This allows the system to recommend new items whose features align with your aggregated preferences.
Common Aggregation Methods
The simplest approach is averaging: take all items the user interacted with positively and compute the mean of their feature vectors. This creates a centroid in feature space representing typical preferences.
More sophisticated methods include:
- Weighted averaging: weight recent interactions more heavily, or weight by rating strength
- TF-IDF weighting: emphasize rare features the user likes (similar to text retrieval)
- Learned embeddings: train a neural network to map interaction history directly to user embeddings
Example Without Code
Imagine a user has rated three movies positively. Each movie has features [action_score, romance_score, comedy_score]:
- Movie A:
[0.9, 0.1, 0.2] - Movie B:
[0.8, 0.0, 0.3] - Movie C:
[0.7, 0.2, 0.1]
The user profile (simple average) would be: [0.8, 0.1, 0.2] — indicating a strong preference for action with mild comedy.
Key Takeaway: User profile construction transforms interaction history into a feature vector that summarizes preferences, enabling similarity-based recommendations in the same space as item features.