Data whisperers extract their secrets from advertising

The XING Marketing Solutions division is responsible for advertising marketing within XING SE. One of the core tasks here is native advertising: the inclusion of relevant advertising content that adapts to the appearance of the platform and supplements editorial contributions and contributions submitted by other users.

The department's own software development team had already developed a solution for the implementation of this task over a long period of time, which consisted not only of an architecture but also ultimately of an algorithm for the selection of relevant advertising material. This algorithm was based on anonymous user characteristics and statistics about clicks on advertising material. At the time, XING Marketing Solutions, as the main person entity for online advertising, increasingly focused on expanding the team around its core competencies of data science and data engineering in order to be able to implement set goals independently within the overall company using a data-driven strategy as well as optimization. In addition to new hires and some restructuring, the division's own team was also supported by codecentric AG in these subject areas in order to tackle existing challenges.

One of the main goals was to further increase sales. To achieve this, the product owner and management wanted better control and monitoring options than those offered by the previous solution in order to better link business metrics with in-system optimization through the algorithm. There was also a need to introduce a clearly defined data-driven methodology as activities expanded, in order to be able to demonstrably measure effects and changes in the data-driven process and distinguish them from other causes. In addition to these requirements in the area of methodology and algorithms, there was also a requirement to adapt the system architecture of the ad delivery pipeline. Besides latency targets, the aim was to develop and operate backend components as well as dedicated systems for machine learning that could handle the large amounts of data involved, and all this was to be achieved in a cross-functional team of software developers and data scientists.

A tracking solution already existed at the beginning of the project, and this recorded key data and performance indicators of the algorithm playing the ads and visualized them in a dashboard. One of the first results was an improvement in tracking and the identification of suitable metrics. This made it possible to better distinguish between cause and effect without them being obscured by variations in key performance indicators (KPIs).

The technical capabilities of clearly defined A/B tests were also developed in order to better distinguish between symptoms and causes and to be able to provide measurable evidence of the effects of system changes. This now made it possible to objectively measure the performance of the advertising algorithm and to compare it with other approaches.

As part of implementing these steps, a review of the legacy architecture and implementation, including the data collection process, was performed in collaboration with the client's team with a view to data quality assurance and data readiness. This also helped to create a better shared understanding across the entire team of the existing solution and the causes of problems related to it.

Close collaboration and continuous communication with the team improved understanding of the requirements and processes of a data science process, as well as better prioritization of data science over other development process objectives. This also increased ownership and accountability for the topic within the team.

At this point, the decision was made to gradually replace the existing system with a new one. This also made it possible to integrate a tool stack for data science and machine learning tasks, allowing new algorithms to be tested to achieve improvements in KPIs and, in turn, revenue targets. The basics of working with the tool stack were taught through employee training and team programming.

Solutions were successfully implemented and the foundations laid for the further development of products using intelligent, self-learning algorithms to create new data-driven product features and to further improve existing ones, bring them into production, and evaluate them in a quantifiably measurable way. Building on this, XING Marketing Solutions was able to achieve a double-digit percentage increase in revenue per impression (RPI) and a reduction in latencies.

The example of platform-wide content personalization is representative of a wide range of applications where automated decisions need to be made based on dynamic data and the characteristics of various data points flowing through a self-learning and adaptive system.

codecentric AG supports with its experts in the areas of data science and data engineering in choosing suitable machine learning techniques and in the technical implementation of data processing chains that contain self-learning components. We can work with you to meet the challenges of developing data-driven solutions. It doesn't matter whether it's the core business, supporting services, or a new digital business model with data at its core.