Recs Algorithm Engine

This section explains how the recommendation engine works, including the data inputs it relies on, the minimum data required for reliable output, and how performance scales with increasing data volume.

Data Inputs: Behavioural Events

Data Inputs are the behavioural and catalogue signals (like clicks, views, purchases, and product attributes) that power the recommendation engine. The quality and volume of these inputs directly determine how accurate, relevant, and personalised your recommendations are.

The recommendation engine is powered by behavioural event data captured via the Unbxd clickstream SDK.

Depending on the algorithm, one or more of the following signals are used:

Event Type	Description	Signal Strength
Click	A shopper clicks on a product	Low
Add to Cart	A shopper adds a product to their cart	Medium
Order / Purchase	A shopper completes a purchase	High
Page / PDP View	A shopper views a product detail page	Low–Medium
Search Query	Text queries entered by the shopper	Medium

Minimum Data Requirements

While the engine can operate with minimal data, stronger behavioural signals lead to more accurate and personalised recommendations. Below certain thresholds, performance may be limited.

Metric	Recommended Minimum	Notes
Lookback Window	7 days (rolling)	Balances recency with sufficient data volume
Total behavioural events	5,000 events	Combined clicks, cart adds, and orders
Unique users	1,000 visitors	Required for user-level personalization
Unique products interacted	500+ stock keeping units	Enables meaningful product associations
Purchase events	500+ orders	Critical for sales-based and collaborative filtering

Graceful Degradation

When data is limited, the engine automatically adjusts its strategy as per below:

Data Volume	Impact
< 1,000 events	Falls back to catalog-based signals (e.g., product attributes, recency)
1,000–5,000 events	Recommendations available but less personalized
5,000+ events	Full algorithm performance with reliable personalization
10,000+ events	High-confidence recommendations across all algorithms

Low-Dependency Algorithms

Some algorithms are available from day one and do not depend on large behavioural datasets.

These include:

Recently Viewed
Recs Based on Recently Viewed
Recs Based on Last Viewed
More Like This
Catalog-driven variants

These rely primarily on session data or product attributes, making them effective even in low-data scenarios.