Product Affinity Graph

The graph is a weighted, directed network where edge strength quantifies the affinity between two products. Weights are typically derived from conditional probability—the likelihood of purchasing product B given product A is in the cart—or from cosine similarity between product embedding vectors. This structure enables the system to distinguish between strong complements (e.g., a phone and its specific case) and weak associations (e.g., a phone and a generic screen cleaner).

Edges are categorized by relationship type to power distinct merchandising strategies. A complementary edge connects items bought together (e.g., a camera and a memory card), driving cross-sell recommendations. A substitutable edge connects items viewed or bought instead of each other (e.g., two brands of the same cereal), enabling intelligent backfill when a product is out of stock. This classification is often determined by analyzing post-purchase return rates or session co-browsing patterns.

Modern affinity graphs integrate with inventory telemetry to prevent recommending unavailable items. Each node carries a dynamic stock_status attribute. During graph traversal for recommendation generation, nodes with a stockout_probability above a threshold are pruned from the candidate set. This ensures the system only surfaces availability-weighted relevance, directly linking the graph's mathematical structure to operational reality.

Global affinity graphs often fail to capture regional taste variations. Localized Affinity Scoring creates geo-specific graph layers by filtering co-purchase data through a geospatial demand cluster. An edge between craft beer A and craft beer B might be strong in one city but nonexistent in another. This allows the same recommendation engine to serve a micro-merchandising zone with culturally relevant suggestions without manual curation.

New products with no transactional history lack collaborative edges. To solve this cold start problem, the graph is initialized with content-based edges derived from product attribute similarity (e.g., same brand, category, color, or material). These edges are initially weak but provide a starting point for exploration. As interaction data accumulates, a contextual assortment bandit can shift traffic from content-based to behavior-based edges, dynamically re-weighting the graph.

When a primary product is out of stock, the affinity graph drives demand transference modeling. The system traverses the outbound substitutable edges of the unavailable node, selecting the connected product with the highest edge weight that is currently in stock. This predicts which alternative the customer is most likely to accept, minimizing lost revenue. The success of the substitution is fed back into the graph to reinforce or decay the edge weight.

A critical distinction in affinity graphs that determines whether a product edge represents a substitutable or complementary relationship. Misclassifying these leads to poor recommendations—suggesting a competing product when cross-selling is intended.

• Substitutes: Products that fulfill the same need (e.g., two brands of AA batteries). Identified by co-view but not co-purchase patterns, or by temporal separation in repeat purchases
• Complements: Products used together (e.g., a camera and a memory card). Identified by co-purchase within the same session or basket
• Advanced models use session-level temporal analysis and product category taxonomies to disambiguate these edge types automatically

The process of extracting product relationship signals from real-time user session data rather than historical purchase records alone. This captures emerging trends and short-lived affinities that batch processing misses.

• Session windows (typically 30 minutes to 2 hours) define the boundary within which interactions are considered related
• Sequential patterns (e.g., viewing A then B then adding C to cart) provide richer affinity signals than simple co-occurrence
• Streaming architectures using Apache Kafka and Apache Flink enable continuous graph edge updates as new session data arrives, keeping the affinity graph fresh

The mathematical formulation that assigns strength scores to connections in the affinity graph. Raw co-occurrence counts are rarely sufficient; sophisticated weighting accounts for popularity bias, recency, and confidence.

• Pointwise Mutual Information (PMI) normalizes co-occurrence by individual item popularity, preventing best-sellers from dominating all edges
• Temporal decay (e.g., exponential or half-life functions) reduces the weight of older interactions, allowing the graph to adapt to seasonal trends and shifting preferences
• Confidence weighting uses interaction type (purchase > add-to-cart > view) and engagement depth (dwell time, scroll depth) to differentiate weak signals from strong intent