DEEM Lab at BIFOLD showcases two papers addressing data quality and privacy in recommender systems at ACM RecSys 2025
From September 22–26, the 19th ACM Conference on Recommender Systems (RecSys) will take place in Prague, Czech Republic. RecSys is the premier international venue for recommender systems, bringing together theoretical advances, infrastructure challenges, and real-world applications. It is widely regarded as the leading annual forum for this area, attracting both academic groups and industry leaders from e-commerce, media, and beyond. The BIFOLD research group DEEM Lab, led by Prof. Dr. Sebastian Schelter, will present two contributions at this year’s conference:
Spotlight Oral Presentation (Main Conference Track)
Scalable Data Debugging for Neighborhood-based Recommendation with Data Shapley Values
This work addresses the critical issue of data quality in large-scale recommender systems. The authors introduce KMC-Shapley, a scalable algorithm to compute Data Shapley Values—a principled way to quantify the importance of individual training data points. By making such calculations feasible on datasets with millions of interactions, the paper demonstrates how impactful data points can be identified, enabling practitioners to preempt harmful recommendations, such as low-quality or unsafe products, and to improve sustainability in e-commerce settings.
Title: Scalable Data Debugging for Neighborhood-based Recommendation with Data Shapley Values
Authors: Barrie Kersbergen, Olivier Sprangers, Bojan Karlaš, Maarten de Rijke, Sebastian Schelter
Read the paper
Workshop Contribution (FAccTRec Workshop on Responsible Recommendation)
Towards a Real-World Aligned Benchmark for Unlearning in Recommender Systems
This paper tackles the emerging field of machine unlearning, which ensures compliance with privacy regulations such as the GDPR’s right to be forgotten. The authors critique existing benchmarks for their lack of realism and propose a new framework that captures real-world unlearning demands: multiple recommendation tasks, sequential deletion requests, and strict efficiency constraints. Their preliminary experiments demonstrate that unlearning can be applied effectively to sequential recommendation models, with latency as low as a few seconds.
Title: Towards a Real-World Aligned Benchmark for Unlearning in Recommender Systems
Authors: Pierre Lubitzsch, Olga Ovcharenko, Hao Chen, Maarten de Rijke, Sebastian Schelter
Read the paper
