11-11, 11:00–12:30 (America/New_York), Winter Garden (5th floor)
Effective forecasting and classification of time series data is critical in a wide variety of industries. Deep learning has made impressive strides with respect to time series forecasting and classification in a academic settings. However, we seldom see deep learning employed by businesses on time series data. This tutorial will will explore and educate Python developers on how to utilize existing open source deep learning (Python) frameworks to forecast and classify their real world temporal data and power business solutions.
In almost all industries the ability forecast future trends, accurately predict events, or detect deviations is vital. For instance, in healthcare you might want to predict patient mortality to better triage critical patients, in retail you might want to forecast the demand for certain items to optimize the supply chain, in cyber security you might want to detect anomalous sequences of actions on a network to triage threats in a timely manner, in climatology you might want to predict a flash flood to evacuate relevant areas in time. The list goes on and on. Historically, many of these industries have relied on simple statistical models, complex physical models, or human intuition to make these decision.
Recently, deep learning has emerged as viable alternative for solving complex time series forecasting, classification, and anomaly detection problems. In academic papers deep learning has achieved state of the art results on many datasets yet it has not been widely adopted in industry. Part of the reason for this stems from the high barrier to utilizing deep learning on time series data. Models often have to be tuned exactly right or they will fail spectacularly. The nuances of training, testing, and deploying deep time series models can take years to master. However, recently many open source tools have emerged to ease this process and make deep learning on time series data more accessible.
In this tutorial we will delve into using several of these tools and frameworks on real world time series datasets. We will look at using Flow Forecast (an end to end deep learning for time series framework), PyTorch Geometric Temporal and several others to forecast and classify climate, healthcare, and sales data. We will also look at interpretability tools like SHAP, ELI5 and others to show what DL models look at when forecasting/classifying time series data. Participants will leave with a solid understanding of tools in the deep learning for time series Python ecosystem and how to use them on their real world data.
Previous knowledge expected
Isaac Godfried is a data scientist at SimSpace. Isaac specializes in utilizing deep learning on real world problems in cybersecurity, climate, healthcare, and agriculture. He is the author and principal maintainer of Flow Forecast a deep learning for time series framework in PyTorch.