By using machine learning, the platform can create predictive product sets (items likely to be purchased next) for individual customers. Unlike general recommendations (e.g. best sellers), which are a one-to-many product set, predictive recommendations are personalised, one-to-one product sets.
Our predictive product recommendations are powered by machine learning models that extract and analyse your customer orders, preferences, and behaviours.
Accessing personalised recommendations
Go to Campaign > Product recommendations, click Create recommendation, mouse over the 'Personalised' type and click on Learn more.
The sidebar will allow you to request the feature is enabled on your account. We will first review your data and then confirm with you that the feature is ready to use, or if there are data dependencies to be resolved first.
Personalised recommendations are available on a 90-day free trial.
Types of model
We don't have a one-size-fits-all machine learning model. Instead, our data science team will recommend and optimise an appropriate model based on the following factors:
- Breadth of data available
- Volume of data
- What you sell
- Who you sell it to
Our models fall into two categories and either model type can be trained and optimised for your data. Data can also be enhanced by our AI product data enrichment to give more accurate recommendations.
Content-based filtering model
The emphasis of this model is on the products, not the customers. Your product catalogue is analysed, and similar products are grouped together. Next, by looking at a contact’s purchase history, product sets are selected based on similarity with previously purchased items. The machine learning algorithm then attempts to predict how a contact would rate these selected products and will rank them accordingly. The ranked set of products forms the basis of a contact’s recommendations.
This model is most effective when the quality and depth of product data is high. It has a narrower scope and so can make very accurate recommendations, in particular with niche purchases.
Collaborative filtering model
With collaborative filtering, recommendations aren't based on direct product analysis. Instead, the model is interested in customer behaviour. The purchase history of a contact is analysed against the context of all customer purchases. From this, we can identify ‘lookalike’ contacts. Product recommendations are then made based on products these lookalike customers have purchased.
This high performing model gives a wider range of relevant recommendations because it learns against the underlying behaviour rules of your customers.
Model training rules
Depending on the volume of data given to the model, it can take several hours to complete training. Your account’s data can train one model at a time and there is a three-hour window between re-training.