Hildesheim researchers win prestigious web technology price!

Hildesheim researchers win prestigious web technology price!

On Thursday, June 12, 2025, the International World Wide Web Conference in Sydney, Australia, researcher at the University of Hildesheim was awarded the renowned Seoul Test of Time Award. The winners, including Prof. Dr. Dr. Lars Schmidt-Thieme and the former doctoral students Dr. Steffen Rendle and Dr. Christoph Freudenthaler were honored for their pioneering research in the field of sequential recommendation systems. Your work has achieved over 2,900 citation since 2010, which underlines the international relevance of your findings. This award has been awarded by the International World Wide Web Conference Committee and the Association for Computing Machinery (ACM) since 2014, the aim of which is to network teachers and researchers and to promote dialogue as well as lifelong learning.

The research of the Hildesheim scientists deals with the modeling of series sequence effects and latent properties of users and items. In particular, the use of their results in the areas of e-commerce and technology enhanced learning is emphasized. A special feature of your studies is the use of data from a third -party fund in cooperation with the drugstore chain Rossmann.

innovations in recommendation systems

The research of Schmidt-Thieme and his team aims to improve existing recommendation systems. Traditional approaches often concentrate on the prediction of a single article. In their studies, on the other hand, the focus is on so-called TOP-K recommendations. This means that several items are predicted at the same time with which users could probably interact. The challenge here is to develop models that not only create a single element, but also a whole, ranked list of articles.

A proven approach that is used is the modification of existing models to enable multiple forecasts. For example, the GPT-2 model is used, which was specially trained in a user sequence for the prediction of the next element. Step-by-step generation takes into account already recommended articles and produces innovative generation strategies such as greedy decoding, Beam Search and Tempertensampling, along with new methods such as recursive ranking.

measurement and results of research

The validation of these methods was based on various data records, including Mo many 20m and Yelp. The experiments carried out ensured that only users were used for analysis with sufficient interactions. NDCG, Recall and Mean Average Precision were used as standard metrics for the performance assessment of the recommendation systems.

The results show that the author -compressive generation strategies offer significant advantages in the quality of long -term predictions. In particular, the relevance aggregation strategy proved to be a superior method compared to traditional approaches. Although the generation of several sequences causes additional computing costs, this process can be parallelized in order to manage the latency and further increase the quality of the recommendations.