The base item–item recommender serves as the foundational model for various other recommendation modules. This core model focuses on analyzing relationships between items to predict which products or content are likely to be of interest to a consumer, based on the items they have already engaged with.
It uses behavioural signals such as product views, cart activity, and past purchases to learn which products are frequently interacted with together. This allows the system to suggest items that are closely related to what a consumer has just viewed or bought.
Inspire's other recommender modules build upon this base model by leveraging its predictions and tailoring the recommendations to specific contexts and use cases.
Key Requirements
Before setting up the Base Item Item Recommender, ensure that:
-
The ENTERPRISE subscription is active.
-
Item engagement data (product views, add-to-cart actions, and purchases) is available via web tracking.
-
A product feed is configured with product IDs that match the product IDs in your transaction/engagement data. This is essential for the model to correctly link interactions to products.
-
If your catalog contains product variants (e.g., different sizes or colors of the same product), ensure these are identifiable via a variant ID, parent ID, or similar identifier so the model can correctly group variant-level interactions.
-
Sufficient engagement history is available for the model to identify meaningful relationships between products.
How it Works
The recommender uses a technique called item–item collaborative filtering. In simple terms:
-
It looks at which items are often viewed, purchased, or added to carts together.
-
If many consumers who bought product A also buy product B, the model learns that A and B are strongly related.
-
When a consumer interacts with product A, the system recommends product B (and other related items).
Engagement Scoring
The model supports multiple ways of scoring item interactions:
-
Count — Each interaction (view, add-to-cart, purchase) is counted equally, regardless of when it happened.
-
Time-discounted — More recent interactions are weighted higher than older ones. This ensures the model adapts to changing product trends over time.
-
Seasonal — Interactions are weighted based on seasonal patterns, accounting for recurring trends (e.g., holiday shopping behavior).
The engagement scorer determines how the model interprets item relationships, and can be configured to best fit the business context.
You can further shape the results with business rules in the Inspire cockpit, by enabling one of the modules above.
Limitations to Consider
-
Popularity bias: Frequently purchased items may dominate recommendations, limiting visibility of niche or new products.
-
Cold start: New products may not yet have enough interaction data to appear in recommendations.
-
Data dependency: The quality of recommendations depends on accurate and complete behavioural data (e.g., product view and purchase tracking).