Description
Feature extraction is the umbrella for transforming raw CSI windows into a vector representation suitable for downstream models — either by hand-crafted statistics (subcarrier-features) or by a learned encoder (cnn, transformer-attention). The thesis uses the term to distinguish "what representation are we feeding" from "what classifier consumes it", because the representation choice dominates cross-domain accuracy.
When it's used
- Pipeline-design discussions of CSI representations
- Comparing hand-crafted vs learned features
- Defining the boundary between sensing and inference layers
Limitations
- Hand-crafted features under-perform learned encoders on rich CSI data
- Learned features rarely transfer cleanly without
domain-adaptation - Feature-importance interpretation is hard for deep encoders
Source Papers
- hou2023_bf83 ↗ — feature-extraction comparison for few-shot CSI
- wang2026_2758 ↗ — feature-extraction strategies in WiFi sensing
- bin2008_8bb7 ↗ — early feature extraction for crowd analysis
- logah2026_c3bb ↗ — feature-extraction for CSI HAR
- chen2023_5cbd ↗ — feature-extraction in generalisation taxonomy