Understanding Recommendation Systems for B2B Construction E‑Commerce

In B2B construction e‑commerce, the traditional "people find goods" model is replaced by a "goods find people" approach, requiring a recommendation system to quickly match projects with suitable products and prevent customers from leaving for competitors.

The system depends on three main data categories:

Basic information: project attributes such as type, structure, status, and region.

Behavior data: collected via tracking, including search keywords, page views, dwell time, and clicks.

Business data: order, add‑to‑cart, and collection records that indicate purchase intent.

These data feed multiple recall strategies (multi‑channel recall): popular items, interest‑topic matching, and collaborative filtering.

Popular recall selects items with high sales or high user interaction metrics.

Interest‑topic recall builds a project‑interest model from project attributes (e.g., building type, structure) to recommend items aligned with those topics.

User‑based collaborative filtering computes similarity between projects using cosine similarity on their preference vectors (e.g., safety helmets, screws, floodlights), allowing the system to recommend items liked by similar projects.

Item‑based collaborative filtering treats items as vectors and calculates item‑item similarity, again using cosine similarity, to suggest products that co‑occur in project preference profiles.

The overall recommendation workflow consists of four stages:

Recall : generate a large candidate pool using the above strategies.

Coarse ranking : apply simple rules or lightweight models to deduplicate and filter out invalid items.

Fine ranking : use more sophisticated models or business rules (e.g., weighting, recent purchase exclusion) to prioritize candidates.

Re‑ranking : ensure diversity by mixing candidates from different recall channels, often using predicted CTR for final ordering.

In summary, the article provides a practical overview of recommendation system fundamentals for the construction B2B market, illustrating data requirements, collaborative‑filtering techniques, and the multi‑stage processing pipeline that turns raw data into personalized product suggestions.