Tutorials

Description & Outline:

Machine learning for data streams (MLDS) attempts to extract knowledge from a stream of non-IID data.

It has been a significant research area since the late 1990s, with increasing adoption in the industry over the past few years due to the emergence of Industry 4.0, where more industry processes are monitored online. Practitioners will presented with challenges such as detecting concept drifts, incremental models models, and efficient approaches to learn from data streams.

Despite commendable efforts in open-source libraries, a gap persists between pioneering research and accessible tools, presenting challenges for practitioners, including experienced data scientists, in implementing and evaluating methods in this complex domain. Our tutorial addresses this gap with a dual focus. We discuss advanced research topics, such as partially delayed labeled streams, while providing practical demonstrations of their implementation and assessment using Python. By catering to both researchers and practitioners, this tutorial aims to empower users in designing, conducting experiments, and extending existing methodologies.

In this tutorial, our objective is to familiarize attendees with applying diverse machine-learning tasks to streaming data. Beyond an introductory overview, where we delineate the learning cycle of typical supervised learning tasks, we steer our focus towards pertinent challenges seldom addressed in conventional tutorials, such as: Prediction Intervals for regression tasks, Concept drift detection, visualization and evaluation, Anomaly detection using incremental algorithms, and the trade-offs of runtime and predictive performance when using the parallel bagging ensembles.

Outline

Part 1: Machine Learning for Data Streams

• Supervised learning

                 – Learning cycle

                 – Evaluation procedures

                – Classification algorithms

                  * Incremental decision trees

                  * Ensemble methods [Gunasekara et al., 2024]

                 – Regression algorithms [Sun et al., 2022]

                 – Prediction Intervals [Sun et al., 2024]

  – Practical examples

• QA (5 minutes)
• Concept drifts

                 – Basic concepts (definitions and categorizations)

                 – Detecting concept drifts

                 – Evaluating and visualizing detections

                 – Using drift detectors in beyond data streams

                 – Practical examples

Part 2: Unsupervised learning (Clustering and Anomaly Detection)

• Clustering

                 – Micro-clusters

                 – Online and offline steps

        – Algorithms

                  * CluStream

                 – Evaluation procedures

                 – Practical examples

• Anomaly Detection

                 – Assumptions and limitations

                 – Algorithms

                  * Half Space Trees

  * Online Isolation Forest

                  * Adaptive Isolation Forest*

                 – Evaluation procedures

                 – Practical examples

Part 3: Parallel ensembles using mini-batch

• Assumptions
• Data locality
• Algorithms

                 – Mini-batching for bagging ensembles

• Evaluation procedures
• Practical examples

Target Audience:

This tutorial’s target audience includes researchers and practitioners, especially those interested in learning from data streams, evolving data, and/or IoT applications. No previous experience in machine learning for data streams is required, but familiarity with traditional machine learning concepts and frameworks (like scikit-learn) is expected.

Tutorial Organisers: Guilherme Weigert Cassales & Yibin Sun (University of Waikato), Heitor Gomes (Victoria University of Wellington

Guilherme earned his PhD in ”Improving data locality of bagging ensembles for data streams through mini-batching” from the Federal University of S ̃ao Carlos, and he currently works as a postdoctoral fellow at the AI Institute of Waikato university. His research interests primarily revolve around StreamLearning, with interests that permeate Anomaly Detection,Explainability, and High Performance Computing. He has been the convener of the University of Waikato’s Data Stream Mining (COMPX523 Masters) on 2023 and 2024. Guilherme has contributed to this field by working on the optimization streaming bagging ensembles, co-developing an online method based on Isolation Forests for Data Streams, and being part of the Accelerated WEKA development team. Guilherme has served a few conferences as a PC member (i.e., ECAI, KDD, ECML, CCGrid), and is part of CapyMOA’s development team.

Heitor is a senior lecturer at the Victoria University of Wellington (VuW) in New Zealand. Before joining VuW, Heitor was a senior research fellow and co-director of the AI Institute at the University of Waikato where he taught from 2020 to 2022 the "data stream mining" (COMPX523) course. 
Heitor's main research area is the application of machine learning for data streams in a variety of tasks. 
In this field, he has contributed to ensemble learning for both regression and classification tasks, worked on unsupervised drift detection, and in 2023, he was awarded a grant to conduct research on developing novel theories and algorithms for partially delayed labeled streams.
Besides participating as a PC member of a multitude of conferences (KDD, IJCAI, ECML, PAKDD, ...) Heitor is also an active contributor to open-source projects like MOA (Massive Online Analysis), StreamDM (a real-time analytics open-source software library built on top of Spark Streaming), and river (where he supervises students and postdocs since the inception of the project). 

Yibin is a fourth-year Ph.D. student at the University of Waikato in New Zealand, where he previously earned his Master’s degree in Information Technology. Yibin’s research centres on advancing machine learning algorithms for streaming data scenarios, with a primary focus on regression and classification tasks. Yibin’s interest extends to addressing challenges like concept drifts and implementing anomaly detection techniques for data streams