iRecSys25 - Talks by Israeli researchers accepted to RecSys25 conference

iRecSys25 - Talks by Israeli researchers accepted to RecSys25 conference

RecSys is a leading international annual conference in the field of recommender systems. Will follow our tradition of hosting the Israeli speakers that were accepted to the recent RecSys25 conference. It'll be as always a pleasure hosting the speakers, to share the ir work with their local community.

We are grateful to Taboola for their community sponsorship and for hosting this event.

Talks are in English.

Agenda:
18:00 - 18:30 : Gathering
18:30 -1900 : Large Scale E-Commerce Model for Learning and Analyzing Long-Term User Preferences
19:00 - 19:30 : Practical Multi-Task Learning for Rare Conversions in Ad Tech
19:30 - 19:40 : Short break
19:40 - 20:00 : PAIRSAT: Integrating Preference-Based Signals for User Satisfaction Estimation in Dialogue Systems
20:00 - 20:30 : Optimal signals assignment for eBay View Item page

Details:

Talk 1
Title: Large Scale E-Commerce Model for Learning and Analyzing Long-Term User Preferences
Speaker: Yonatan Hadar

Abstract:
A user browses watches today but what if their real passion is restoring vintage cars? Standard systems latch onto recent clicks, missing the bigger picture. We built NILUS to change that. NILUS is a transformer based model trained with contrastive learning over raw item text and a full year of user behavior, producing stable embeddings that reflect long term interests. Designed to predict behavior weeks ahead, not just the next click, NILUS improves accuracy, diversity, and engagement at scale. We also introduce a novel evaluation framework to verify that NILUS captures enduring user preferences, not just short term signals. This talk shares how we built and deployed NILUS in a real world ecommerce system. The paper was presented at recsys 2025.

Bio:
Yonatan Hadar is a data scientist with expertise in machine learning and deep learning across several domains, including natural language processing (NLP), time series analysis, and recommendation systems. Matan holds an M.Sc. in Industrial Engineering with a specialization in NLP. Currently, working as a Senior Applied Researcher at eBay, focusing on large-scale recommendation engines and NLP tasks-specifically training neural networks to generate high-quality embeddings.

Talk 2
Title: Practical Multi-Task Learning for Rare Conversions in Ad Tech
Speaker: Yuval Dishi

Abstract:
We present a multi-task learning (MTL) approach for improving the prediction of rare and high-value conversion events (e.g., purchases) in online advertising.
Conversions are categorized as either rare or frequent based on historical statistics, eliminating the dependency on advertisers’ manual tagging.
The model learns shared representations across all signals while specializing through separate task-specific towers for each type.
This approach was tested and fully deployed in production, demonstrating consistent improvements in both offline and online key performance indicators (KPIs).
The proposed method may also be applicable to other rare-event prediction tasks in large-scale recommender systems, such as low click-through rate (CTR) events or other infrequent user actions.

Bio:
Yuval Dishi is a Senior Machine Learning Engineer and Tech Lead in the AI department at Teads, where he is responsible for developing and optimizing algorithms that power various components of the company’s recommender system.
Before joining Teads, he led research and development in the fields of machine learning and cybersecurity within the Israeli Military Intelligence - Unit 8200.
He holds a Master of Science degree in Computer Science from Bar-Ilan University.
His research interests include deep learning, classical machine learning, recommender systems, and crowdsourcing.

Talk 3
Title: Optimal signals assignment for eBay View Item page
Speaker: Matan Mandelbrod

Abstract:
Signals are short textual or visual snippets displayed on the eBay View-Item (VI) page, providing additional, contextual information for users about the viewed item. The aim in displaying the signals is to facilitate intelligent purchase and to incentivise engagement. In this paper, we present two approaches for developing statistical models that optimally populate the VI page with signals. Both approaches were A/B tested, and yielded significant increase in business metrics.

Bio:
Matan Mandelbrod has been working at eBay for the past 5 years. The Signals optimization problem has been particularly challenging and interesting, and was an opportunity to to get acquainted with causal inference and uplift modeling. Before joining eBay I worked at Liveperson, and before that at BrightSource Energy, Runcom and IBM research labs. I have an Msc in applied mathematics, and I practice (and instruct) meditation.

Talk 4
Title: PAIRSAT: Integrating Preference-Based Signals for User Satisfaction Estimation in Dialogue Systems
Speaker: Adir Solomon, Ph.D.

Abstract:
User satisfaction estimation in dialogue systems is a fundamental measure for assessing and improving conversational-AI quality and user experience. Current approaches rely on users’ satisfaction annotations, referred to as supervised labels. Yet these labels are scarce, costly to collect, and often domain-specific. Another form of feedback arises when a user selects one of two offered responses in a conversation, usually called a preference signal. In this work, we propose PAIRSAT, a new model for user-satisfaction estimation that integrates both satisfaction labels and preference signals. We reformulate satisfaction prediction as a bounded regression task on a continuous scale, enabling fine-grained modeling of satisfaction levels. To exploit the preference data, we incorporate a pairwise ranking loss that encourages higher predicted satisfaction for accepted conversation responses over rejected ones. PAIRSAT jointly optimizes regression on labeled data and ranking on preference pairs using a Transformer-based encoder. Experiments demonstrate that our model outperforms baselines that rely solely on supervised satisfaction labels, demonstrating the value of adding preference signals. Further, our results underscore the value of leveraging additional signals for satisfaction estimation in dialogue systems.

Bio:
Adir Solomon has been an Assistant Professor since 2023 in the Department of Information Systems at the University of Haifa. He received his Ph.D. in 2022 from Ben-Gurion University of the Negev and completed a postdoctoral fellowship (2022-2023) at KU Leuven’s Research Center for Information Systems Engineering. His research focuses on user modeling, recommender systems, and natural language processing. Dr. Solomon uses machine learning and deep learning techniques to enhance personalization, improve communication in healthcare, and mitigate biases in AI systems, with a particular emphasis on morphologically rich languages such as Hebrew.