≡ Menu

Protecting the most vulnerable #pawcon


John McConnell of Analytical People presented on the use of analytics in the UK’s transition to digital TV. This is being rolled out region by region, trying to improve the process each time. Several regions have been converted already. There is a group of people, in the UK as in other countries, for whom TV is very important and yet who are not guaranteed to notice or manage the switchover. The UK has created a company to manage a help scheme for those over 75, who could get disability support, lived in care or who have sight disabilities. The BBC Trust has given this organization a mission to make sure no-one loses their TV service in the switchover while maintaining value, encouraging appropriate takeup and identifying and helping the most vulnerable.

The BBC Trust commissioned some research before the program started and found, as expected, that most would have no problems but that there is a significant group who would have problems making the transition unassisted. Of this 20% about 5% are likely to be left behind no matter what because they simply don’t respond to marketing or outreach. To reach people the program is using generic and targeted marketing as well as local and regional press. But it is adding social networks, charities, outreach and other less traditional channels for "marketing". To begin they use geospatial analysis first to map people to the transmitters because these are what get converted. While many people are clearly served by one, some are in a gray zone where a probability of using one of several transmitters has to be modeled.

But, beyond the geospatial issues, the project develops response models for managing take-up and scoring models to score the most vulnerable. The awareness of predictive analytics came about from both the GIS consultants (who understood the potential) as well as more general awareness. Critically there were senior executives who "got it" and who became executive level business sponsors for the project.

Building these models took data and, while lots of data is publicly available, there are lots of concerns about data privacy and protection. Awareness of data losses and the potential impact of this has become a real headline topic. To address this they used black box models – neural networks. They were developed, shipped to the UK Department of Work and Pensions (who own the most sensitive data) and executed there with just the score being returned. Curiously the value of an opaque, hard to understand model was promoted (the reverse of the usual approach). An illustration of the value of a score that obscures the details of a person while still being able to be useful.

The first model had 42% non responders which was predicted about 73% of the time by the model. More recent regions have higher non-responders, more than 65%, but higher rates of prediction (over 95%). The model is based on a decision tree. The biggest contribution to non-response is post code, probably a sign that something is missing from the model that has a geographic distribution. Age is the second most predictive variable. Young people had the highest rates of non-response and Asian communities also showed very poor response.In the UK there is an Experian service called Mosaic Origins that uses your full name and derives your likely origins. Mosaic was used in later models and confirmed that Asian groups were being particularly missed. So lots of work was targeted on Asian ethnic groups as well as younger groups. The rates have gone up and down, largely depending on the proportion of these less well targeted groups in the region being converted.

Lessons learned include:

  • Geospatial analysis and predictive analytics are very complementary
  • Need to invest time in communicating and proving the point
  • Predictive analytics is only a component of these projects, it must fit the project
  • You need sponsors who get it and vendors who can help them do so