Building a time series model using CNNs and GANs
Time series anomaly detection and classification problems have existed and there are various existing solutions to tackle such kind of problems. However, all/most of the solutions are for ideal cases (having enough labelled data or developing a model with average precision/recall) and does not take into account the practical constraints of implementing and deploying a highly generalizable solution.
In this talk, I intend to cover brief overview of how time series anomaly detection problems can be tackled with an unconventional approach which captures the multi-spatial relationships between time and various features, augment the dataset using deep generative models (GANs) and train a classifier to achieve state of the art results. I also intend to briefly touch upon an evaluation framework for measuring GAN performance by evaluating on explicitly parameterized, synthetic data distributions which can be applied to any dataset.
- Approaches that did not work relative to our benchmark of precision/recall
- Data transformation - an innovative way to transform time series data so as to capture the multi-spatial relationships between time and features.
- Building a deep generative model – configuration parameters for the model which worked in practical implementation and are reproducible
- Discuss the evaluation framework
- How to make this approach work in other domains
- Intermediate understanding of CNN
- Basic understanding of the nature of IoT data
Prasenjeet is a Senior Data Scientist working in Ericsson Global AI Accelerator(GAIA) Team. He has 10 years of industry experience in applying data science and advanced machine learning techniques to diverse sectors including telecom, finance, manufacturing and equity research for quantitative hedge funds. Married to Data Science (but dating Data Engineering). He is also helping a Silicon Valley start-up in finding its feet in the industry. He has keen interest in Open Source and its contributors and trying to be one of them. Has contributed to Scikit-learn and other python libraries.