A user-item interaction matrix maps users to items, with each cell recording interaction strength. Positive signals such as clicks, purchases, and ratings can populate the relationship values. section 4.2.3 section 9.3.1
The matrix supports recommendation and personalized search in two progressively richer ways:
- Use overlapping interactions to find users who engaged with the same item, then boost other items those users also engaged with. section 4.2.3
- When the matrix is sparse, factor it into a user-feature matrix and an item-feature matrix, capturing latent preferences and item similarities. section 4.2.3 section 9.3.1
Weighted interactions
A purchase receives more weight than an add-to-cart action, which receives more weight than a click. Empty cells indicate no observed interaction, while factorization can fill them with predicted preferences. section 9.3.1
