Schedule: Temporal data and time-series analytics sessions

Arun Kejariwal, Karthik Ramasamy, and Anurag Khandelwal walk you through the landscape of streaming systems and examine the inception and growth of the serverless paradigm. You'll take a deep dive into Apache Pulsar, which provides native serverless support in the form of Pulsar functions and get a bird’s-eye view of the application domains where you can leverage Pulsar functions. Read more.

Go hands-on with Sophie Watson and William Benton to examine data structures that let you answer interesting queries about massive datasets in fixed amounts of space and constant time. This seems like magic, but they'll explain the key trick that makes it possible and show you how to use these structures for real-world machine learning and data engineering applications. Read more.

ARIMA has been used for time series modeling for decades. In practice, most time series collected from human activities exhibit seasonal patterns, but the efficient estimation of seasonal ARIMA ((S)ARIMA) models was inefficient for decades. Meir Toledano explains how Anodot was able to apply the technique for forecasting and anomaly detection for millions of time series every day. Read more.

Observability is the key in modern architecture to quickly detect and repair problems in microservices. Modern observability platforms have evolved beyond simple application logs and include distributed tracing systems like Zipkin and Haystack. Keshav Peswani and Ashish Aggarwal explore how combining them with real-time, intelligent alerting mechanisms helps in the automated detection of problems. Read more.

Anomaly detection may sound old fashioned, yet it's super important in many industry applications. Tony Xing, Congrui Huang, Qiyang Li, and Wenyi Yang detail a novel anomaly-detection algorithm based on spectral residual (SR) and convolutional neural network (CNN) and how this method was applied in the monitoring system supporting Microsoft AIOps and business incident prevention. Read more.

Data-driven software is revolutionizing the world and enable intelligent services we interact with daily. Robert Pesch and Robin Senge outline the development process, statistical modeling, data-driven decision making, and components needed for productionizing a fully automated and highly scalable demand forecasting system for an online grocery shop for a billion-dollar retail group in Europe. Read more.

Criteo’s infrastructure provides the capacity and connectivity to host Criteo’s platform and applications. The evolution of this infrastructure is driven by the ability to forecast Criteo’s traffic demand. Hamlet Jesse Medina Ruiz explains how Criteo uses Bayesian dynamic time series models to accurately forecast its traffic load and optimize hardware resources across data centers. Read more.

Stavros Kontopoulos and Debasish Ghosh explore online machine learning algorithm choices for streaming applications, especially those with resource-constrained use cases like IoT and personalization. They dive into Hoeffding Adaptive Trees, classic sketch data structures, and drift detection algorithms from implementation to production deployment, describing the pros and cons of each of them. Read more.

The application of smoothing and imputation strategies is common practice in predictive modeling and time series analysis. With a technique-agnostic approach, Anjali Samani provides qualitative and quantitative frameworks that address questions related to smoothing and imputation of missing values to improve data density. Read more.

Time series forecasts depend on sensors or measurements made in the real, messy world. The sensors flake out, get turned off, disconnect, and otherwise conspire to cause missing signals. Signals that may tell you what tomorrow's temperature will be or what your blood glucose levels are before bed. Alfred Whitehead and Clare Jeon explore methods for handling data gaps and when to consider which. Read more.

Machine learning (ML) gets a lot of hype, but its classical predecessors are still immensely powerful, especially in the time series space, and classical algorithms outperform machine learning methods in time series forecasting. Anais Dotis dives into how she used the Holt-Winters forecasting algorithm to predict water levels in a creek. Read more.

Failures or issues in a product or service can negatively affect the business. Detecting issues in advance and recovering from them is crucial to keeping the business alive. Join Akshay Rai to learn more about LinkedIn's next-generation open source monitoring platform, an integrated solution for real-time alerting and collaborative analysis. Read more.

Heitor Murilo Gomes and Albert Bifet introduce you to a machine learning pipeline for streaming data using the streamDM framework. You'll also learn how to use streamDM for supervised and unsupervised learning tasks, see examples of online preprocessing methods, and discover how to expand the framework by adding new learning algorithms or preprocessing methods. Read more.

Jeroen Janssens dives into stochastic outlier section (SOS), an unsupervised algorithm for detecting anomalies in large, high-dimensional data. SOS has been implemented in Python, R, and, most recently, Spark. He illustrates the idea and intuition behind SOS, demonstrates the implementation of SOS on top of Spark, and applies SOS to a real-world use case. Read more.