Content Personalisation for The Sun + The Times

Our impact

• Developed state-of-the-art recommender systems to contribute to the titles’ digital viewership, user engagement and reader satisfaction
• Developed cutting-edge features, such as automated real-time systems for model retraining, and implement custom metrics to drive meaningful engagement
• Improved transparency with explainable AI methods to help to understand why each article recommendation is being served

 

The challenge

News Corp declared in their 2019 Annual Report that their mission was to “deliver content in a more engaging, timely and personalised manner”. As a key step in this direction, News UK requires greater control in personalising article recommendations for The Sun,  The Times and The Sunday Times. These recommendations will be used in key parts of the user experience – alongside editorial curation – to help readers discover relevant stories. 

Their key requirements are:

1. Develop a high-performing model according to business priorities
2. Process large volumes of data in real-time and cost-effectively
3. Provide explainability and appropriate controls for each recommendation being served.

“We have experimented with personalisation in the past, however, we were always left frustrated with the opaque and inflexible nature of the solutions. We need to cater for a broad range of user experiences across our different titles, where different levels and types of personalisation are appropriate.” – Dan Gilbert, Director of Data, News UK

 

Our solution

The team kickstarted the engagement by conducting exploratory data analysis on the user, context and content trends for The Sun and The Times and The Sunday Times. As a result, a suite of features were engineered to use in the Machine Learning models. Advanced feature engineering was additionally tested, such as image processing with the Google Cloud Vision API and text processing with natural language processing techniques.

In the design phase, the models were constructed to maximise business objectives such as readership, user engagement and satisfaction catering to advertisement and subscription-based revenue models for The Sun and The Times and The Sunday Times respectively. For this purpose, a list of the latest Machine Learning methods were surveyed and tested, such as deep learning, the attention mechanism and generalised matrix factorisation.

The most significant features were prioritised to speed up model evaluation – for example, reader interests and reading behaviours were more relevant than demographics. For the final recommender, adjustable levers were included to prioritise factors such as popularity and diversity.

To automate the model re-training mechanism used for the recommender system, Datatonic used a set of bash scripts. In addition to this, and considering the multitude of components and tools constituting the architecture (such as Dataflow, Pub/Sub, cloud functions, Compute Engine and Memorystore), Terraform was a tool of choice to build, change, and manage pipeline versions safely and efficiently throughout the entire of the infrastructure lifecycle.

News UK is now experimenting with the recommendation capabilities Datatonic helped to develop on The Sun and will shortly be rolling out the first experiment on The Times and The Sunday Times. Once the initial positive results have been assessed, the capability will be rolled out more widely across Sun and Times products. The News UK team is also looking at how they can bring personalisation into their Wireless radio station products such as Virgin Radio, Times Radio and talkSPORT.

“Datatonic brought a wealth of expertise and experience to our team, and collaborated with our engineers and data scientists, working to their standards, so that we can continue to evolve the system they developed.” – Dan Gilbert, Director of Data, News UK