Base User Item Recommender

What it is & the business problem it solves

The recommender is the part of Inspire that gives every customer their own ranked list of products, for example for a "Recommended for you" block in a newsletter or on the homepage. You do not need to choose a separate product for it: since June 2026 the recommender is powered by a transformer model, a state-of-the-art technique. It is generally available (GA) and included in the Enterprise package. It is not an add-on and not a separate product.

Earlier recommenders used collaborative filtering: "people who engaged with X also engaged with Y". That approach mostly ignores the order in which a customer did things. A transformer is a type of neural network that is built to read sequences (the same family of models used for language). It reads each customer's shopping journey step by step and predicts which products are most likely to come next. That is how the recommender decides what to show a customer.

You work with the recommender through two modules in the cockpit (Inspire's front-end tooling, where you adjust the business rules of your predictions):

Lists for product pages (Contextual Product Recommendation) and follow-up messages after a behaviour (Event Based Retargeting) are not built by this engine. They use the Base Item Item Recommender.

New customers start with the transformer-powered recommender by default.

How it works (plain language)

  1. Collect journeys. For each customer, Inspire builds a timeline of their product interactions: product detail views, add-to-cart, remove-from-cart and purchases. The model also sees how much time passed between steps, for example "same session", "same day", "within a week" or "older". It does not use channel touchpoints such as email opens or clicks.

  2. Train (on a schedule). On the schedule you set, the model learns patterns from all customer journeys. Training runs on a GPU (a graphics processor used for machine learning). A training run first waits a few minutes for the GPU machine to start.

  3. Predict. After each run of the data load that prepares the journeys, Inspire scores every customer with recent activity and stores their top products (10 by default).

  4. Personalize and deliver. The Personalized product recommendation transformer module takes these raw lists. It removes products your rules exclude, applies boosts and fixed positions, and tops up short lists with a fallback. A fallback is a generic list such as the most popular products. It then delivers the result in the same way as Personalized Product Recommendation: to CDP or CXP.

The engine only recommends products. It does not predict when a customer will buy, and it does not predict churn. Purchase-timing probabilities come from a separate module (see Next Purchase Moment).

Example use cases

Use case

Scenario

Channel

Required products

Homepage "For you" carousel

A home-improvement retailer shows a personal carousel on the homepage. It A/B-tests the transformer lists against the classic recommender.

Website

CDP or CXP, Webtracking, Product Feed, Web Personalization (via the Recommendations API), A/B Testing module (optional)

Personal newsletter block

A baby-products webshop adds a "Picked for you" block of 4 products (illustrative) to its weekly newsletter.

Email

CDP or CXP, Webtracking, Product Feed, MMC Email Campaigns

Two audiences from one engine

A retailer runs two personalization instances on one engine: one for newsletters (exclude bought items) and one for the website (boost new arrivals).

Email + website

As above

What you can configure

You can configure the settings below yourself in the cockpit. Settings marked Expert need expert rights. CM staff and customer admins can change expert settings; to get expert rights, ask your CM contact person.

Setting (cockpit name)

What it does

Default

When to change

Who can change

Training Schedule (Cron)

When the model is retrained (a cron schedule, a standard notation for "run at these times", in Amsterdam time)

Not scheduled: no model is trained until this is set

Always set this at go-live. Weekly retraining is a common start (illustrative).

Expert

Number of Recommendations

How many products the engine stores per customer (1–100)

10

Raise it (for example to 50–100). Otherwise business rules in the personalization module remove items and many lists fall back to generic products.

Expert

Min Item Engagement Count

How many different customers must have engaged with a product before the model can learn it

1

Raise it for very large catalogues with a long tail of rarely seen products

Expert

Recommendation Expiration Days

How long a stored list is kept

30 days

Rarely

Expert

Inference User Scoping (Enable User Scoping, Inactive Days, Outdated Recommendation Days)

Limits predictions to recently active customers and skips lists that are still fresh

Off (when on: 31 days inactive, 7 days outdated)

Turn on for large customer bases, so that not every profile is rewritten on every run

Expert

Model and training settings (Sequence Construction, Input Features, Model Architecture, Regularization, Loss Function, Training Configuration, Masked Pretraining (BERT4Rec), Evaluation)

Data-science tuning of journey length, optional features (day of week, price, category), model size and training

Tuned defaults (for example journeys of 3–50 steps; price, category and day-of-week features off)

Only with CM data-science guidance

Expert

Training Resources / Inference Resources

Memory and CPU for training and prediction

16 GB training memory; 2 CPU cores for prediction

When runs fail for lack of memory or are slow

Expert


Prerequisites & required CM products

Requirement

Why

CDP (CM's customer data platform) or CXP (CM's customer experience platform)

The platform that holds your customer data and receives the lists. The Batch Method in the delivery settings decides whether lists go to CDP or CXP.

Standard event types on CDP: the Retail & Ecommerce adapter

Sets up the standard event types (Purchase, Add to cart and so on). It is included in onboarding: there is nothing extra to buy or configure. What matters is that events use these standard event types; custom event types are ignored.

Standard event types on CXP: Templates

The Retail & Ecommerce adapter does not apply to CXP. On CXP you use Templates to set up the equivalent events.

Webtracking (CM's website tracker)

Product views and add-to-cart / remove-from-cart events

Purchase lines (order lines from the webshop or back office)

Training uses only customers who have purchases

Product Feed

Product catalogue. Product IDs in events must match the feed's item ID.

Data module: transformer loader

The Data module must have Enable Transformer Loader switched on and a Transformer Loader Cron set (CM setup). Prediction starts after this loader finishes. The loader also requires the Product Feed to be enabled in the Data module.

Packaging tier

Included in the MMC Enterprise package (both the Transformer Recommender and the Personalized product recommendation transformer). Not available in Basic, Advanced or Pro. It is generally available and not sold separately.

Data requirements

  • Events used: product detail view, add-to-cart, remove-from-cart and purchase lines, in time order. Email, SMS or web-personalization events are not used.

  • Technical minimum requirements (as built into the product; they are minimums for the model to work, not a promise of quality):

    • Training: only customers who have at least one purchase are used, and their journey must have more than 3 events spread over at least 2 separate sessions. Journeys are cut to the most recent 300 events per customer and at most 50 steps per sequence.

    • Products: a product must have at least 1 engaged customer (Min Item Engagement Count) to be learned.

    • Prediction: every customer with at least one event receives a raw list. Customers without events get nothing from the engine, and the personalization module does not create a list for them.

  • New products (cold start): products that appear after the last training run are unknown to the model until the next retrain. Retrain regularly.

  • Delisted products: the stored model can still suggest a product that was delisted after training, until the next retrain. Use Item Exclusion Criteria (for example "in stock = false") in the personalization module.

  • History: Inspire keeps customer events for up to 730 days (about two years).

How to check it works

  • Run history: check that training runs succeed on schedule and that prediction runs after each transformer-loader run.

Limitations & common pitfalls

  • Nothing happens after setup. Both the Training Schedule and the Recommender Schedule start as "not scheduled". The Data module's transformer loader must also be on and scheduled.

  • High fallback rate. The engine stores only 10 products per customer by default. After exclusions there may be too few left, and lists fill up with popular products. Raise Number of Recommendations.

  • Profile flooding. Inference User Scoping is off by default, so every customer with an event is predicted and delivered on every run. Turn it on for large audiences.

  • Delisted products can appear until the next retrain. Combine an "in stock" exclusion rule with regular retraining.

  • Variant-heavy catalogues. If every size or colour is a separate product ID, some categories can be under-represented. Check that event product IDs and feed item IDs use the same level.

  • Training can fail for lack of memory on large catalogues; CM can raise Training Resources.

  • Not a timing or churn model. For purchase-timing probabilities use Next Purchase Moment (shown as Timing Propensity in the cockpit).

  • Lists are not visible instantly. A run must age briefly and API responses are cached, so new lists do not appear immediately.