Presented By O’Reilly and Cloudera
Make Data Work
September 11, 2018: Training & Tutorials
September 12–13, 2018: Keynotes & Sessions
New York, NY

Machine learning for nonstationary streaming data using Structured Streaming and StreamDM

Heitor Murilo Gomes (Télécom ParisTech), Albert Bifet (Télécom ParisTech)
3:30pm–4:10pm Thursday, 09/13/2018
Data engineering and architecture
Location: 1E 07/08 Level: Intermediate
Secondary topics:  Temporal data and time-series analytics

Who is this presentation for?

  • Data scientists and machine learning engineers

Prerequisite knowledge

  • A working knowledge of Structured Streaming in Apache Spark
  • An intermediate understanding of machine learning

What you'll learn

  • Understand how StreamDM can be used alongside Structured Streaming and the relevance of addressing concept drift when learning from streaming data
  • Learn how to implement active and reactive strategies to address this problem using StreamDM

Description

Adapting StreamDM to the novel Structured Streaming engine simplifies both its use and development. Currently, the open source StreamDM library provides the largest collection of data stream mining algorithms for Spark, including both supervised and unsupervised learning algorithms that can be updated online. 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.

There is a vast literature on the topic of addressing concept drift and learning from streaming data. Still, these methods can be complex to implement and integrate. Heitor Murilo Gomes and Albert Bifet explain how to use StreamDM and Structured Streaming to develop, apply, and evaluate learning models specially for nonstationary streams (i.e., those with concept drifts). Adapting StreamDM for Structured Streaming is a natural step that facilitates future integration with major technology improvements, such as continuous processing. Heitor and Albert also introduce a simple yet powerful methodology to address concept drifts using active strategies like combining ensemble models and drift detectors and reactive strategies and reactive strategies like forgetting mechanisms and periodical resets (windowed approaches).

Photo of Heitor Murilo Gomes

Heitor Murilo Gomes

Télécom ParisTech

Heitor Murilo Gomes is a researcher at Télécom ParisTech focusing on machine learning—particularly evolving data streams, concept drift, ensemble methods, and big data streams. He co-leads the StreamDM open data stream mining project.

Photo of Albert Bifet

Albert Bifet

Télécom ParisTech

Albert Bifet is a professor at LTCI and head of the Data, Intelligence, and Graphs (DIG) Group at Télécom ParisTech, and a scientific collaborator at École Polytechnique. A big data scientist with 10+ years of international experience in research, Albert has led new open source software projects for business analytics, data mining, and machine learning at Huawei, Yahoo, the University of Waikato, and UPC. At Yahoo Labs, he cofounded Apache SAMOA (Scalable Advanced Massive Online Analysis), a distributed streaming machine learning framework that contains a programing abstraction for distributed streaming ML algorithms. At the WEKA Machine Learning Group, he co-led MOA (Massive Online Analysis), the most popular open source framework for data stream mining, with more than 20,000 downloads each year. Albert is the author of Adaptive Stream Mining: Pattern Learning and Mining from Evolving Data Streams and the editor of the Big Data Mining special issue of SIGKDD Explorations in 2012. He was cochair of the industrial track at ECML PKDD 2015, BigMine (2014, 2013, 2012), and the data streams track at ACM SAC (2015, 2014, 2013, 2012). He holds a PhD from BarcelonaTech.