Abstract:
The increasing deployment of sensors, connected devices, and cyber-physical systems is generating continuous, high-speed data streams in dynamic, evolving environments. In such settings, learning systems must move beyond static models trained once on finite datasets, toward adaptive models that learn continuously, detect change, and forget outdated information. This talk explores the role of learning and forgetting in data stream mining, with emphasis on real-time decision-making under limited computational resources. The first part presents a case study on interconnect bypass fraud detection in telecommunications. Traditional blacklist-based approaches are limited in their ability to detect new fraud patterns and adapt to behavioural changes. We discuss a data-driven online approach based on streaming algorithms, including Heavy Hitters and Hierarchical Heavy Hitters, and introduce a fast-forgetting extension of the Lossy Counting algorithm. Experimental results on large-scale call datasets show that forgetting improves the detection of recent abnormal activity while reducing memory usage and execution time. The second part addresses rare-event explanation in industrial sensor streams. We present a neuro-symbolic architecture for real-time fault detection and explanation in air compressor units. The detection layer uses an LSTM autoencoder trained on normal behaviour, while the explanation layer learns online regression rules that map sensor inputs to reconstruction error. A Chebyshev-based sampling strategy focuses learning on rare, high-error observations. This architecture provides both global explanations through learned rule sets and local explanations through the rules triggered by individual anomalies. Together, these examples illustrate how online learning, adaptive forgetting, and interpretable modelling can support robust, efficient, and explainable intelligence over data streams.

Biography:
João Gama is an Emeritus Professor at the University of Porto, Portugal. He received his PhD in Computer Science from the University of Porto in 2000. He taught Informatics and data sciences at the School of Economics for more than 30 years. He was the Director of the Master's in Data Analytics for 12 years. He is an EurAI Fellow, an IEEE Fellow, and a Fellow of the Asia-Pacific AI Association. He is a member of the Academia das Ciências de Lisboa. He is a member of the board of directors of LIAAD, a group affiliated with INESC Tec. His main scientific contributions are in the field of learning from data streams, for which he has an extensive record of publications. He is the Editor-in-Chief of the International Journal of Data Science and Analytics, published by Springer.

Abstract:
Modern scientific discovery and decision-making necessitate agents capable of navigating both complex datasets and multifaceted, often conflicting, goal landscapes. This talk explores the intersection of conversational data exploration, an incremental process where agents help users articulate needs through data interaction, and multi-objective sequential decision making. While traditional Reinforcement Learning (RL) has advanced task-specific agent training, we are witnessing a paradigm shift toward zero-shot orchestration. By leveraging a spectrum of techniques, from specialized RL policies to general-purpose Large Language Models (LLMs), we can reuse existing single-objective policies to solve complex multi-objective problems without training from scratch. We will discuss empirical evidence from the field of Education showing that this orchestrated approach can achieve competitive Pareto quality while reducing computational cost. The talk concludes by examining open research questions regarding how LLM context richness and reflective foresight enable agents to bridge the gap between incremental data-driven insights and optimal sequential actions.

Biography:
Biography: Sihem Amer-Yahia is a Silver Medal CNRS Research Director and Deputy Director of the Lab of Informatics of Grenoble. She works on exploratory data analysis and algorithmic upskilling. Prior to that she was Principal Scientist at QCRI, Senior Scientist at Yahoo! Research and Member of Technical Staff at at&t Labs. Sihem served as PC chair for SIGMOD 2023 and as the coordinator of the Diversity, Equity and Inclusion initiative for the database community. In 2024, she received the 2024 IEEE TCDE Impact Award, the SIGMOD Contributions Award, and the VLDB Women in Database Award.