2019-09-16 –, Assembly Room
How to automatically identify, and describe, interesting patterns in timeseries data, such as trends, change-points and periodic behaviour.
There has significant progress in recent years in developing new, exciting, flexible, predictive ML models. However! Considerable expertise is still required to choose appropriate features/models AND the output of a model, whilst accurate, can be difficult to understand! When dealing with lots of timeseries data it would be useful if we had a system to automatically group and describe similar timeseries and fit a state-of-the-art predictive model all at once...
In this talk I outline an elegant piece of work called the Automatic Statistician (based on Gaussian Processes) and how I implemented a simple Python API which can automatically detect interesting structure in timeseries data (trends, change-points, periodicity, etc) as well as providing a plain English description of the result.
Joe is a Lead Data Scientist at JPMorgan and is currently working on applications of Natural Language Processing for media monitoring and content recommendation for various teams within the firm. Before that he worked on a big data framework timeseries anomaly detection. He has been at JPMorgan for over 5 years working in pure Python development roles, big data, machine learning and data science.
His background is an undergraduate in Avionics (Glasgow) and a PhD in Reinforcement Learning and Control Engineering (Cambridge) in which he wrote a lot of MATLAB code, and hand calculated a lot of gradients - he is now very thankful for Python and autograd!