Introducing Two Criteria for Future Data Prediction with ML Models

on YouTube ↗01h 09m

" summary

Defines two concrete pass/fail rules for next-step prediction: distance-domain criterion requires model metrics to beat average differences of random pairs, and time-domain criterion requires beating average differences of adjacent pairs. Provides Jupyter/Colab/Kaggle code, trains bidirectional LSTM (input length 10 → predict 11th) on synthetic 10,000-sample data and on NASA MODIS AOT images. Reports concrete numbers: prior RMSE 0.32/CS 0.85 (criteria: RMSE 0.46/0.42), data-prep change yields RMSE 0.06/CS 0.998 (criteria: RMSE 0.09/0.066).

AI-written summary · about

" chapters (12)

1. 0:00Tutorial introduction
2. 1:50Evaluation criteria theory
3. 8:41Satellite image case study
4. 16:49Practical demonstration
5. 18:17Synthetic data generation
6. 24:53Deep learning training
7. 31:53Data splitting
8. 36:57LSTM architecture
9. 44:07Model performance report
10. 49:52Regularization methods
11. 54:28Model test evaluation
12. 1:01:08Q&A

" related talks

• 23mHow to detect silent failures in ML modelsWojtek Kuberski · Conf42 Machine Learning 2022
• 39mForecasting With Classical and Machine Learning MethodsJonathan Bechtel · PyData NYC 2023
• 27mMaking Disk Failure Predictions SMARTer!Tirthak Patel, Devesh Tiwari, Weisong Shi, Sidi Lu, Bing Luo, Yongtao Yao · FAST '20
• 27mModel Evaluation for HumansDante Gates · PyData Miami 2019
• 31mML time series forecasting methodsAnna Warno, Pawel Skrzypek · Conf42 Machine Learning 2021
• 17mMulti-view Feature-based SSD Failure Prediction: What, When, and WhyYuqi Zhang, Wenwen Hao, Ben Niu, Kangkang Liu, Shuyang Wang, Na Liu, Xing He, Yongwong Gwon, Chankyu Koh · FAST '23: 21st USENIX Conference on File and Storage Technologies
• 25mTimeSeries Forecasting with ML Algorithms and there comparisonsSonam Pankaj · PyData Global 2020
• 29mMachine Learning: Better than Traditional Techniques?Hannah Higgins · PyCon UK 2018