Affinity Scoring

Affinity scores are computed from a weighted composite of explicit and implicit behavioral signals. Explicit signals include ratings, likes, and wishlist additions. Implicit signals—often more voluminous—include:

• Click-through rate (CTR) on product pages
• Dwell time and scroll depth
• Add-to-cart and checkout initiation events
• Search query reformulation patterns
• Cross-category browsing sequences Each signal type is assigned a decay function to prioritize recency, ensuring that a purchase three years ago does not outweigh a browse event from three minutes ago.

User interest is non-stationary; an affinity for winter coats in January should not influence recommendations in July. Temporal decay functions apply a half-life to every interaction. Common implementations include:

• Exponential decay: score = base_weight * e^(-λt) where λ controls the decay rate
• Linear decay over a fixed window (e.g., 30 days)
• Event-based decay triggered by lifecycle changes (e.g., a purchase resets the browse affinity for that category) The half-life parameter is typically tuned per vertical—fast fashion requires shorter windows than automotive or real estate.

Modern affinity scoring moves beyond a single numeric score to a multi-dimensional vector capturing preference across orthogonal axes:

• Category affinity: Preference for product categories (electronics, apparel)
• Brand affinity: Loyalty to specific manufacturers or labels
• Price-band affinity: Sensitivity and preference for discount, mid-tier, or luxury price points
• Attribute affinity: Preference for specific features like color, material, or size
• Channel affinity: Propensity to engage via email, push notification, or in-app message These vectors are stored in a feature store and served to downstream ranking models at inference time.

Affinity scores are enriched through collaborative filtering techniques that identify latent preference patterns across user cohorts. Key approaches include:

• User-based CF: Users with similar affinity vectors are clustered; items popular within a cluster are surfaced to all members
• Item-based CF: Items frequently co-engaged with by the same users receive cross-affinity boosts
• Matrix factorization: Techniques like Alternating Least Squares (ALS) decompose the user-item interaction matrix into dense user and item embeddings, which serve as the foundation for affinity computation This hybrid approach mitigates the cold start problem by leveraging cohort behavior when individual history is sparse.

Affinity scores are not batch-computed overnight; they update in near real-time as new events stream in. This requires:

• Event-driven architecture using platforms like Apache Kafka or Amazon Kinesis
• Streaming aggregation with windowed computations (tumbling, sliding, or session windows)
• Incremental model updates that adjust scores without full retraining
• Write-through caching to Redis or Aerospike to serve updated scores with sub-millisecond latency This ensures that a user who just abandoned a cart for a specific SKU immediately sees related recommendations on their next page load.

To prevent engagement fatigue and filter bubbles, affinity scoring systems implement guardrails:

• Saturation capping: After N exposures to a category without conversion, the affinity score plateaus or decays to prevent over-recommendation
• Exploration bonuses: A small epsilon value is added to under-represented categories to ensure the system continues to explore new interests
• Negative affinity signals: Explicit dismissals, returns, and rapid bounce-backs decrement scores, creating a true bidirectional preference model These mechanisms ensure the system respects evolving user intent rather than trapping users in a static preference echo chamber.