How does the model work?

Our model is a deep-learning model.

The model tracks everything that a user does on site. We collect 5 types of events from customers, which we convert into over 800 model inputs. These 5 types of events are click events, time beacons, page view, add to carts and conversions. Each event passes core data (e.g. time, URL, device type, referral information), and event specific data (e.g. for click event - information about that click).
​

We also feed website specific inputs into the model. For example, what is the price range of the products, or how many steps are in the checkout.

The model then returns an output on every event a user does. Everything returned is a probability; context and labels (such as buying stages, or high intent to convert) are derived based on the output.

Why do you use a universal model over a site-specific model?

The Intent model is a single model and is trained on all our online retail clients. We found that a universal multi site trained model worked best across retailers than individual site specific approach.

All our websites are continuously fine-tuned to our models to get the optimal performance for all our clients. Our models are continuously iterated on and optimised for accuracy and performance. This is only made possible by only having to do this for all sites at once.

A universal model trained across multiple retailers increases the accuracy over any model trained only for a specific website due to the sheer volume of data availability. It also means a client benefits from user behaviours not yet experienced on their site specifically. For example, imagine launching a new product range. How can we understand users’ intent if this is totally new behaviours and sequences? The universal model has seen hundreds of millions of interactions and would have no problem in interpreting users’ intent, even if it's new to your site.

Is my dashboard and scoring specific to my website(s)?

The short answer is yes. Despite our models being trained across all our retailers the model predictions are specific to the selected site in the platform, and the model takes in website specific inputs (such as the price range of products). You will only ever see the data that’s specific to you, benefitting from both the universal model’s power but specific training on your data.

Is my data still private in this universal model?

Yes. All client data is partitioned in our data stores. Then anonymised and scaled prior to being sent for model training.

Is the model up to date? How often is it trained?

We go through a quarterly training cycle with our models to ensure they are continuously fine tuned and monitored for optimal performance.

How live is the data in my dashboard? How often is it updated?

Data is currently available up to the previous day within the Intent Performance Score and our Segments. However our real-time data health check provides you with a view of the metrics we are receiving, in real-time