Personalisation has become a core component of digital marketing and the guiding principle behind a range of new products. At insurance giant Aviva, it provided the conceptual foundations of ADA, an algorithmic decision agent that powers omni-channel, hyper-personalised marketing by predicting a customer’s “next best action” – the most effective way that a company can meet their individual needs in a specific moment.
ADA applies this concept by turning the focus of marketing from solving the problems of the business to those of the customer. It uses a supervised machine learning model that trawls through Aviva’s enormous volume of historical customer data to predict what each individual will want to in the future.
The insights it provides have boosted a range of marketing services. Personalising the offers and news in email, direct mail and display advertising has helped boost sales of certain products by three figures, while tailoring the ads on Aviva’s customer portal to individual visitors has resulted in a double-digit increase in click-through rates on the offers.
These results have been achieved through analysing a combination of transactional data on the products customers purchased, demographic data on who the customers are, behavioural data on their contact history, and clickstream data on what they do when they visit the company and its web site.
Aviva also added a variety of new features to ADA. These include the timing of customer actions, which gives the system an understanding of seasonality and how this is reflected in the product lifecycle, which has proven particularly powerful for products that are renewed at certain times of the year, such as car insurance.
“It means that within that renewal window, ADA is going to prioritise sending messages to the customer for whom it’s most relevant,” said Damian Rumble, a senior data scientist in Aviva’s customer science team, Dataiku’s EGG LDN conference.
“And it also means that when you’re outside that window there’s no point us contacting a customer about something they already own or if they’re not in the renewal window, so it depresses that score and allows other products to have that chance to be promoted.”
Combining complex datasets
Aviva traces its heritage all the way back to 1696, when the Hand-in-Hand fire insurance office was founded in London. The company has grown from then to become the UK’s biggest insurance company, with 31 million customers using a wide range of products picked up through acquisitions, from Friends Life’s life insurance to RAC’s car insurance. Here, the company provides pension and investment management services, as well as an extensive insurances portfolio.
This long history and diverse collection of products and customers give the company an enormous volume of data with powerful potential, but the legacy systems struggled to turn it into a single view of a customer.
The range of products also made it tricky to get each prediction right. Aviva’s biggest product category is the general insurance class composed of car, home and travel insurance, but it also derives significant revenue from its savings and retirement and life and health products. This creates an algorithmic bias towards the targets with a lot of information.
Rumble’s team introduced a number of machine learning techniques to balance the bias out and reduced the multi-classification problem by removing some of Aviva’s most minor products, such as pet insurance, from the model, as these didn’t have enough data to create a reliable prediction.
All of this data is integrated into the ADA next best action system, which assesses the information to predict a customer’s propensity to purchase for each of Aviva’s 12 major product categories. It was built using the Dataiku data science platform, which allows Aviva to seamlessly turn the models into production. APIs were built into Dataiku and Hadoop and then integrated with the Adobe Marketing Cloud to distribute campaigns through all Aviva’s marketing channels.
“Our first versions were weekly, and that was already a big improvement,” said Rumble. “But then we were able to do daily and now we’ve got our first our hourly campaigns based on scores produced by ADA, and all other kinds of data and campaigns. Now we push from Hadoop into Adobe campaign on an hourly basis, which is a massive step forward.”
Future plans
Rumble now wants to move the system to Amazon Web Services (AWS) to boost the processing speed to close to real-time and to make better use of clickstream data that arrives in an hourly feed.
For data scientists who have only just started exploring algorithmic marketing, he assures them that the systems can be quick and easy to implement. Rumble proved this during a six-week secondment to Aviva’s Singapore office, where he built a functioning next best action system for the local market from scratch.
“All that was within a six-week period,” he said. “It was actually relatively simple to build and put into production and operate this system using next week’s action and Dataiku.”
IDG News Service
Ltd
