This talk was recorded in London on October 30, 2018 and can be viewed here: https://youtu.be/EGVY7-Spv8E
Time series is a unique field in predictive modelling where standard feature engineering techniques and models are employed to get the most accurate results. In this session we will examine some of the most important features of Driverless AI’s newest recipe regarding Time Series. It will cover validation strategies, feature engineering, feature selection and modelling. The capabilities will be showcased through several cases.
Bio: Marios Michailidis is now a Competitive Data Scientist at H2O.ai He holds a Bsc in accounting Finance from the University of Macedonia in Greece and an Msc in Risk Management from the University of Southampton. He has also nearly finished his PhD in machine learning at University College London (UCL) with a focus on ensemble modelling. He has worked in both marketing and credit sectors in the UK Market and has led many analytics’ projects with various themes including: Acquisition, Retention, Recommenders, Uplift, fraud detection, portfolio optimization and more.
He is the creator of KazAnova, a freeware GUI for credit scoring and data mining 100% made in Java as well as is the creator of StackNet Meta-Modelling Framework. In his spare time he loves competing on data science challenges and was ranked 1st out of 500,000 members in the popular Kaggle.com data competition platform. Here is a blog about Marios being ranked at the top in Kaggle and sharing his knowledge with tricks and ideas.
Finally, Marios’ likendin profile can be found here, with more information about what he is working on now or past projects.
https://www.linkedin.com/in/mariosmichailidis/
Bio: A Kaggle Grandmaster and a Data Scientist at H2O.ai, Mathias Müller holds an AI and ML focused diploma (eq. M.Sc.) in computer science from Humboldt University in Berlin. During his studies, he keenly worked on computer vision in the context of bio-inspired visual navigation of autonomous flying quadrocopters. Prior to H2O.ai, he as a machine learning engineer for FSD Fahrzeugsystemdaten GmbH in the automotive sector. His stint with Kaggle was a chance encounter as he stumbled upon the data competition platform while looking for a more ML-focused platform as compared to TopCoder. This is where he entered his first predictive modeling competition and climbed up the ladder to be a Grandmaster. He is an active contributor to XGBoost and is working on Driverless AI with H2O.ai.
Linkedin: https://www.linkedin.com/in/muellermat/
11. Date Sales
1/1/2018 100
2/1/2018 150
3/1/2018 160
4/1/2018 200
5/1/2018 210
6/1/2018 150
7/1/2018 160
8/1/2018 120
9/1/2018 80
10/1/2018 70
Lag1 Lag2
- -
100 -
150 100
160 150
200 160
210 200
150 210
160 150
120 160
80 120
Moving Average
-
100
125
155
180
205
180
155
140
100
Feature Engineering (cont.)
• Exponentially Weighted Moving Averages (EWMA) of n-th order differentiated lags
• Aggregation of lags (mean, std, sums, etc.)
• Interactions of lags (e.g. Lag2 - Lag1)
• Linear regression on lags (taking slope and/or intercept as new feature)
12. • Ranking based on autocorrelation
• Pre-defined intervals (based on estimated frequency)
Daily data
• [7, 14, 21, …]
• [14, 28, 32, …]
• …
Weekly data
• [2, 4, 6, 8, …]
• [4, 8, 12, 16, …]
• …
…
Candidates for Lag-Sizes
13. • Lower bound for considered lag sizes
• Dropout
• Random replacement of actual lag-values by „n.a.“
• Align frequency of available lag information between train and validation/test
• Target binning
• Decrease of possible amount of splits GBM can perform
Regularization of Lag-Features
15. Time Series Roadmap
• Data augmentation at time of prediction
• Signal processing and classification
16. Signal Processing and Classification
Time TG Value Target Feature 1 .. Feature n Target
1 A A(1) 0 F1(A) .. Fn(A) 0
2 A A(2) 0 F1(B) .. Fn(B) 1
3 A A(3) 0
4 A A(4) 0
5 A A(5) 0
1 B B(1) 1
2 B B(2) 1
3 B B(3) 1
4 B B(4) 1
5 B B(5) 1
Original Frame Transformed Frame
• 1 label associated to each individual time series (signal)
• Feature engineering w.r.t. time (frequency domain features, sliding window aggregates,
zero crossings, peak patterns, etc.)
17. Time Series Roadmap
• Data augmentation at time of prediction
• Signal Processing and Classification
• Additional algorithms (Prophet, „classical“ methods, …)
• Ensembling
• Iterative learning (propagation of predictions to create lag-features)
• …