Seminar M.Sc Students
What Do Fidelity Benchmarks Really Measure? A Controlled Study of Sequential Explanations
Matan Lange, M.Sc. student in the School of Industrial & Intelligent Systems Engineering
Advisors: Prof. Noam Koenigstein
Abstract: Attribution-based explanations are a natural fit for sequential rec-ommendation because each input token corresponds to a human-interpretable item in the user’s history. Yet, despite growing interestin explaining sequential recommenders,
empirical comparisons re-main fragmented across architectures, datasets, and evaluation pro-tocols, making it unclear what current fidelity benchmarks actually reveal. In this work, we present the first controlled benchmark of sequence-based attribution fidelity for sequential recommendation.
We evaluate eight attribution methods across three major backbone families, including SASRec, BERT4Rec, and GRU4Rec on seven real- world datasets under a unified preprocessing, training, and evalua-
tion pipeline. Our results demonstrate that fidelity rankings are not universal, but depend strongly on the recommender backbone and on whether the method is allowed to refine or select explanations at
inference time. We further show that a simple recency-based heuris- tic is surprisingly competitive on causal models, while SpinRec and LXR-ITF consistently define the strongest learned tier, mirroring
their strong standing in collaborative filtering. At the same time,a Brute-Force search reference remains stronger in many settings, indicating substantial headroom for future progress. Overall, our findings indicate that fidelity benchmarks in sequential recom-mendation capture not only attribution quality, but also backbone inductive bias and evaluation design.
Bio: Matan Lange is an M.Sc. student in the School of Industrial Engineering at Tel Aviv University, supervised by Prof. Noam Koenigstein. He is also a Data Scientist at Maccabi Healthcare Services, where he works on applied machine learning, deep learning, and medical AI. His research focuses on explainability in sequential recommender systems, particularly attribution methods and fidelity-based evaluation.
Effective Counter-Narrative Generation Against Misinformation and Hate Speech: A hybrid human-AI approach
Carmel Kronfeld, M.Sc. student in the School of Industrial & Intelligent Systems Engineering
Advisors: Prof. Irad Ben-Gal
Abstract: Online misinformation and hate speech increasingly spread through emotionally charged narratives amplified by social media dynamics and coordinated influence operations. In democratic societies, direct suppression of such content raises concerns about censorship, polarization, and public distrust. Counter-narratives offer an alternative approach by responding to harmful or misleading claims through reframing, with the goal of reducing their persuasive and emotional appeal.
Our work presents a multi-stage LLM-based framework for generating and refining counter-narratives against harmful and misleading pro-Russian narratives related to the war in Ukraine. The methodology combines human evaluation with an iterative refinement process that adapts counter-narratives to specific target narratives and communication goals. Results show that this process substantially improves counter-narrative quality across evaluation dimensions, with improvements further validated by human judgments.
Finally, we evaluate the intervention in a field experiment on X, where tweets were randomly assigned to control and treatment conditions, with treatment tweets receiving a counter-narrative reply. The findings suggest a mixed but informative effect: while replies increase exposure to the target tweet, consistent with platform engagement dynamics, they also suggest reductions in likes and shares. This pattern points to both the risks and potential of counter-narrative interventions in real online environments.
Bio: Carmel Kronfeld is an M.Sc. student in the School of Industrial Engineering at Tel Aviv University, specializing in Data Science and AI. His research, supervised by Prof. Irad Ben-Gal, head of LAMBDA Lab, focuses on counter-narrative generation against misinformation and hate speech.
A Framework for Prediction-Only Expert Identification in Classification via Disagreement Geometry
Gal Gussarsky, M.Sc. student in the School of Industrial & Intelligent Systems Engineering
Advisors: Prof. Irad Ben-Gal
Abstract: Modern machine learning systems increasingly operate in heterogeneous ecosystems of pretrained models that vary in competence, calibration, and specialization. In many practical settings, however, only model predictions are available — without access to labels, validation data, retraining, or model internals. This raises a fundamental question: can expert models be identified and effectively aggregated from predictions alone?
We propose a prediction-only framework for unsupervised, class-conditional expert identification in classification, based on disagreement geometry. For each class, we construct pairwise model distances from prediction disagreement patterns, inducing a geometry over the model pool, and identify compact, well-separated expert subsets using a Dunn-index criterion. The selected experts are then aggregated through a geometry-aware weighted voting scheme that leverages each model's relative centrality within the expert subset.
We evaluate the framework on 24 datasets spanning vision, tabular, and text modalities, using heterogeneous pools of modern convolutional, transformer-based, and classical models. Across datasets, expert-restricted aggregation achieves the best macro-F1 on 18 of 24 benchmarks, outperforming competing unsupervised, prediction-only baselines. Beyond aggregate performance, the experiments reveal structured success and failure regimes, and we develop prediction-only diagnostics that help anticipate when expert identification is reliable.
Overall, the results show that disagreement geometry can contain actionable latent information about model competence and specialization, enabling effective prediction-only expert identification and aggregation under identifiable structural conditions, while also highlighting fundamental limitations of geometry-based approaches in fully unsupervised settings.
Bio: Gal Gussarsky is an M.Sc. student in the School of Industrial Engineering at Tel Aviv University, specializing in Data Science and AI. His research, supervised by Prof. Irad Ben-Gal, head of LAMBDA Lab, focuses on the reliability of AI systems and on prediction-only methods for identifying and combining expert models, using disagreement geometry between predictions to reveal competence across model pools.
