SEA: Search Engines Amsterdam

This meetup session will focus on machine unlearning for retrieval and recommendation, featuring two speakers: Meghdad Kurmanji from IQVIA and Pierre Sicco Lubitzsch from Technische Universität Berlin.

Location: Science Park 904, Room C3.161
Date: Friday, Sep 25
Time: 16:00-17:00
Zoom link: https://uva-live.zoom.us/j/65011610507

We’ll also post our future events on GatherEasy. Here is the registration link for this event: https://gathereasy.rsvp/event/WjetI0ax7

Details below:
Speaker #1: Meghdad Kurmanji

Title: Machine Unlearning: Methods, Limits, and Regulatory Demands

Abstract: Machine unlearning aims to remove the influence of selected training data from a model without the cost of full retraining. The central challenge is to achieve convincing forgetting while preserving performance on the data that should remain. In this talk, I will introduce how unlearning is defined and evaluated, then use examples from my work on SCRUB and the factors that make unlearning difficult to examine where current methods succeed and fail. I will also discuss how the problem changes in federated learning and large language models, where forgetting can be difficult to specify and verify. Finally, I will consider what regulated deployments require beyond promising benchmark results: identifying what must be removed, testing for residual influence, and producing evidence that supports an auditable claim. The talk will close with open research questions at the intersection of unlearning, model behavior, and verification.

Bio: Meghdad Kurmanji is a Senior AI Research Scientist at IQVIA. He previously did a postdoc at Cambridge University and currently stays as a Visiting Fellow there. His research focuses on machine unlearning, AI safety, and scalable (decentralised) machine learning systems. His work has appeared at NeurIPS, ICLR, and SIGMOD. Meghdad co-organized the first NeurIPS Machine Unlearning Challenge on Kaggle with Google. His PhD thesis received a SIGMOD Jim Gray Doctoral Dissertation Honorable Mention.

Speaker #2: Pierre Sicco Lubitzsch

Title: ERASE - A Real-World Aligned Benchmark for Unlearning in Recommender Systems

Abstract: Machine unlearning (MU) enables the removal of selected training data from models to address privacy compliance, security, and liability issues in recommender systems. Existing MU benchmarks poorly reflect real-world recommender settings: they focus primarily on collaborative filtering, assume unrealistically large deletion requests, and overlook practical constraints such as sequential unlearning and efficiency. We present ERASE, a large-scale benchmark for MU in recommender systems designed to align with real-world usage. ERASE spans three core tasks---collaborative filtering, session-based recommendation, and next-basket recommendation---and includes unlearning realistic scenarios such as sequentially removing sensitive interactions or spam. The benchmark covers seven unlearning algorithms, including general-purpose and recommender-specific methods, across nine public datasets and nine state-of-the-art models, producing over 600 GB of reusable artifacts, such as extensive logs and over a thousand model checkpoints. The artifacts that we release enable systematic analyses of where current unlearning methods succeed and where they fail. ERASE shows that approximate unlearning matches retraining in some settings, but robustness varies widely across datasets and architectures. Repeated unlearning exposes weaknesses in general-purpose methods, especially for attention-based and recurrent models, while recommender-specific approaches behave more reliably. ERASE offers an empirical basis for the community to assess, drive, and track progress toward practical MU in recommender systems.

Bio: Pierre Sicco Lubitzsch is a PhD student at TU Berlin and BIFOLD. His research focuses on developing and evaluating machine unlearning methods that efficiently remove the influence of unwanted or harmful training data from recommendation models. He has published work at SIGIR and RecSys.

Counter: SEA Talks #311, and #312.