Posts by Collection

portfolio

publications

talks

Machine Learning for Spark Streaming with StreamDM

Published:

The main goal of this tutorial is to introduce attendees to big data stream mining theory and practice. We will use the StreamDM framework to illustrate concepts and also to demonstrate how data stream mining pipelines can be deployed using StreamDM.

Big Data Stream Mining using Spark Streaming

Published:

The volume of data is rapidly increasing due to the development of the technology of information and communication. This data comes mostly in the form of streams. Learning from this ever-growing amount of data requires flexible learning models that self-adapt over time. In addition, these models must take into account many constraints: (pseudo) real-time processing, high-velocity, and dynamic multi-form change such as concept drift and novelty. The tutorial was combined with a workshop on the same topic.

Machine learning for nonstationary streaming data using Structured Streaming and StreamDM

Published:

The main difference between batch machine learning implementations in Spark (MLlib and Spark ML) and StreamDM is that the latter focus on algorithms that can be trained and adapted incrementally. This can be a huge advantage in some domains as it enables automatically updating the learning models. StreamDM is currently under development by Huawei Noah’s Ark Lab and Télécom ParisTech.

Streaming Random Forest Learning in Spark and StreamDM

Published:

We present how to build random forest models from streaming data. This is achieved by training, predicting and adapting the model in real-time with evolving data streams. The implementation is on the open source library StreamDM, built on top of Apache Spark.

Machine learning for streaming data: Practical insights

Published:

In many domains, data is generated at a fast pace. A clear example is the Internet of Things (IoT) applications, where connected sensors yield large amount of data in short periods. To build predictive models from this data, you need to either settle for traditional offline learning or attempt to learn from the data incrementally. A significant setback with the offline learning approach is that it’s slow to react to changes in the domain, and these changes can have a catastrophic impact on the model predictive performance, since the patterns in which the model was trained on are no longer valid. An online approach where the model is trained incrementally can potentially fix this; however, the untold story is that the existing challenges for offline learning are still present (and are even maximized) when processing the data online. These challenges include, but are not limited to, raw data preprocessing, efficient incremental updates to models, algorithms to detect changes and react to them, and dealing with lots of unlabeled and delayed-labeled data.

Lecture at the IOT Stream Data Mining course

Published:

Lecture at the IOT Stream Data Mining course (Paris, France) as part of the Data and Knowledge 2nd year Master Program of Université Paris Saclay 2019-2020.

teaching

COMPX523: Data Stream Mining

Undergraduate and MSc course, University of Waikato, School of Computing & Mathematical Sciences, 2020

This paper is an introduction to stream data mining. Data streams are everywhere, from F1 racing over electricity networks to news feeds. Data stream mining relies on incremental algorithms that process streams under strict resource limitations. This paper focuses on, as well as extends the methods implemented in MOA (Java) and scikit-multiflow (Python), two open-source stream mining software suites currently being developed by the Machine Learning group at the University of Waikato. More information.