Title: Unsupervised Concept Drift Detection from Deep Learning Representations in Real-time

Abstract: Concept drift is a phenomenon in which the underlying data distribution and statistical properties of a target domain change over time, leading to a degradation in model performance. Consequently, production models require continuous monitoring through drift detection techniques. Most drift detection methods to date are supervised, relying on ground-truth labels. However, in many real-world scenarios, true labels are often not available. Despite recent efforts to develop unsupervised methods, they often lack the necessary accuracy or are too complex for real-time implementation in production environments characterized by high data dimensionality and volume. To overcome these challenges, we propose DriftLens, an unsupervised framework for detecting concept drift in real-time. DriftLens works on unstructured data by leveraging the distribution distances of deep learning representations, enabling fast detection. In addition, it can characterize drift by analyzing the impact on each label separately. Our evaluation across deep learning classifiers and data types shows that (i) DriftLens outperforms previous methods in detecting drift in 13 out of 15 use cases; (ii) it runs at least 5 times faster; and (iii) its drift curve closely matches the actual amount of drift with a correlation of 0.85 or higher.

About the speaker: Tania Cerquitelli is a Full Professor at the Department of Control and Computer Engineering at the Polytechnic of Turin, Italy. In addition to her academic role, Tania holds responsibilities for aggregate functions under the Deputy Rector for Society, Community, and Program Delivery, where she supports various initiatives for the Polytechnic University Community. She also serves as a member of the university's trade union relations delegation and actively participates in the Gender, Equality, Diversity, and Inclusion (GEDI) Observatory. Tania's research focuses on several cutting-edge topics within the areas of data science and machine learning, including techniques for explaining black-box models, algorithms designed to democratize data science, and early detection of concept drift. She is a member of the editorial boards of several international journals, including Computer Networks, Future Generation Computer Systems, Expert Systems with Applications, Engineering Applications of Artificial Intelligence (all published by Elsevier), Knowledge and Information Systems (Springer), and IEEE Data Descriptions. Her involvement in the academic community extends to her role on the steering committee of ECML-PKDD and ADBIS. She has served as Program Co-Chair for ADBIS 2022, Co-Chair for the Journal Track of ECML-PKDD 2023, and Co-Chair for the EDI Special Day at ACM KDD 2024. Additionally, Tania has organized over 15 international workshops on data science and machine learning topics in recent years. Her research has been supported by funding from various sources, including the European Union, the Piedmont Region, the Ministry of University and Research, and private companies.

NB: Prof Tania Cerquitelli is also an opponent of Mahtab Shahin's PhD thesis "Efficient and effective association rule mining on big data and cloud computing: a multivariate analysis", and will participate in the thesis defence as a member of the defence committee. The defence will take place on 27 September 2024, starting at 10:00 in room ICT-638 (Akadeemia tee 15a) and will be accessible via the web application Zoom.