- Through collaboration with data partners and leveraging AdMaster's own data sources, the company creates a 'truth dataset' to decide which mobile-device IDs and cookies can be accurately paired and which cannot.
- Through machine learning, AdMaster further expands the 'truth dataset' in hope of building an ID behaviour network that could improve mapping accuracy across different devices at scale.
- The cross-device solution is tested in real digital campaigns to 'train' the 'truth dataset' to connect more dots.
“What differentiates us from other cross-device measurement providers is by marrying the merits of deterministic and probabilistic models, our technology is a best fit for China’s fragmented reality," said Hong Bei (洪倍), founder and chief technology officer from AdMaster.
A deterministic model is based on matching unique information associated with a user across both mobiles and PCs, allowing advertisers to target, reach and measure against known audiences.
Examples of commonly-used deterministic models include single sign-in information with Facebook and WeChat accounts. However, these options work only if registered customers actually log in.
A probabilistic approach looks at various data points from Wi-Fi networks, locations, device types, and operating systems to make the best possible estimate about the user. IP address identification is the most important anchor of this approach, assuming that someone who always browses from a smartphone or PC with the same IP address can be recognised as a unique user.
“This can work, say, in 70 to 80 per cent of cases in the US market. It, however, bears a high risk of wrong mapping when localising to China's internet environment,” said Hong.
According to Hong, this is attributed to the very complicated technical situation in China. The precondition of probabilistic accuracy is a stable IP address environment, in which each person (or at least each household) can be allocated at least a unique and permanent IP address.
In the US, internet users enjoy 4.62 IP addresses on average. But in China, the average allocation is 0.51 IP addresses, which means each IP address cannot be reliably associated with one individual.
The unstable IP environment in China, combined with other disturbing factors such as pirated cellphone software, could lead to incorrect mapping.
For this reason, targeted advertising based solely on either deterministic or probabilistic models makes it difficult for marketers to consistently follow consumers throughout their path to purchase online, he added.