Opening the ‘black box’ of machine learning

Media and telecommunications company

Our client is a provider of subscription television, internet and mobile services. It has an international reach totalling over 22 million customers, and a history of developing innovative products and entertainment content.

Machine Learning models played an important role in the way our client marketed to its huge audience (in part thanks to previous work by Kin + Carta). The task this time was to open the ‘black box’ and visualise this machine learning so it could be understood more clearly, including the SHAP values driving AI decisions.

Our aim was also to enable the client’s data scientists to create their own visualisations without the need to understand front-end development. A ‘Train the trainer’ approach would help embed this tooling across the business.

To deliver the project, we shortlisted and prototyped Python and JavaScript-based front-end frameworks, evaluated them and made final selection. Our team then optimised back-end data access to provide rapid search and retrieval of cached and live data from the AI Platform.

We developed a set of custom libraries to act as a semantic layer over the data access and visualisation to simplify the process for end users. The visualisation app was hosted on App Engine and a dashboard framework used as a base for the custom library. Datastore was used to cache query results and common queries to reduce hits on the AI Platform.

The new framework trialled with data scientists to very positive feedback. The data science team can now better understand the personalised recommendations provided to customers.

New visualisations can be also developed within a couple of days to examine Machine Learning models across the business.