Traditional inference techniques and infrastructure tools can be illsuited to time-series data, which may be noisy, streaming, multispectral, irregularly sampled, and/or extremely large. This BoF is aimed at identifying promising approaches to time series analysis across a diverse set of use cases and finding projects of common interest for potential future collaborations.
- Share use cases, tooling & pain points around time series analysis and inference
- Identify common tools & difficulties across use cases
- Open issues discussions
- Find potential cross-domain/cross-methodology areas for future collaboration
- How to deal with "too much data" -- sensors generating more data than can bear sent to analysis pipeline (e.g., radio astronomy, high-energy physics)
- Rapid/real-time inference with limited data (seismology)
- Dealing with concept drift; online/incremental models
- Databases/query mechanisms for time-series (e.g., InfluxDB)
- Inference influencing outcomes influencing inference... (e.g., reinforcement learning)
- Anomaly detection
- Handling noisy, uncertain, irregularly sampled data
Another thought or potential topic of conversation: challenges in dealing with scale. E.g., as a neuroscientist when I think "time series" I think on the order of >= 2Hz, and in my specific case of electrophysiology, more like >=100Hz. That seems to be way faster than most other kinds of data out there, and as such many packages that are designed for TSA (e.g. pandas' time series functionality) isn't really useful for neuro analysis. I wonder if that's a fundamental problem that can't be resolved, or if there are more clever package structures that could handle it.