The second day at the 5th Data Donation Symposium at the Weizenbaum-Institut in Berlin starts with Rianne Nienke Visscher, whose focus is on the consistence of the digital download packages used in data donation projects. This emerges from the WHAT-IF project, which aims to create a digital twin of the digital political information environment across a number of European projects. Very quickly, this project found that the data structures across individual donations were often slightly different.
This pointed the project to issues with the completeness, structure, changeability, and measurement of variables in data download packages. Such variations apply within platforms, and across platforms: within platforms there are variations due to differences in user habits, and due to changes in platform and download packages designs, for instance. These need to be addressed by comparing JSON paths in the packages for a single platform, and relate them to each other; some of this can be done via LLMs, and a canonical data structure can be designed from this. Across platforms, this has now resulted in a 90%+ success rate for the major platforms from which data donation packages are available, at least for the data fields WHAT-IF is interested in.
Through pairwise comparison, it is also possible to systematically assess the variations between different packages, and variations for specific data types within the package; an average pairwise similarity also enables the project to assess the broader data design variation rate. This also points to differences based on varying user habits – e.g. active and passive usage. This approach has already been tested by comparing human and LLM-based classifications, which is showing some promising results. In a next step, there is then also the possibility of assessing between-platform variations in similar ways.
