time-series-activity-recognition

name	time-series-activity-recognition
description	End-to-end workflow for labeled-stream activity recognition: prepare or download data, run window-size selection, evaluate a windowed classifier, and produce plots + a short report. Use when you need a reproducible CLI workflow artifact or evaluation bundle.
compatibility	Designed for ts-agents CLI. Also usable by coding agents that can run shell commands.
metadata	{"domain":"time-series","tasks":["classification","activity-recognition","windowing","evaluation"],"ts_agents":{"tool_category":"classification","preferred_workflow":"activity-recognition","preferred_tools":["select_window_size_from_csv","evaluate_windowed_classifier_from_csv"],"artifact_checklist":["window_selection.json","eval.json","report.md"]}}

Activity recognition workflow (labeled stream + window-size selection)

When to use

Use when you need a reproducible end-to-end activity-recognition workflow artifact (dataset, window-size search, evaluation metrics, plots, and short report).

Goal

Given a CSV with columns:

• x,y,z (values) and
• label (per-timepoint activity label)

produce:

• best window size (by balanced accuracy)
• final evaluation metrics + confusion matrix
• a plot of score vs window size
• a short report (Markdown)

Minimal artifact checklist

• <output-dir>/window_selection.json
• <output-dir>/eval.json
• <output-dir>/window_scores.png
• <output-dir>/confusion_matrix.png
• <output-dir>/report.md

Workflow

1. Prepare a labeled-stream CSV (synthetic default or real data).
2. Run the activity-recognition workflow.
3. Save plots and a short markdown report.

Fast path: synthetic dataset (no downloads)

Run from repo root:

uv run python data/make_synthetic_labeled_stream.py \
  --scenario gait --seconds 40 --seed 1337 \
  --out data/demo_labeled_stream.csv
uv run ts-agents workflow run activity-recognition \
  --input data/demo_labeled_stream.csv \
  --label-col label \
  --value-cols x,y,z \
  --output-dir outputs/activity-recognition

This:

1. generates data/demo_labeled_stream.csv
2. selects a window size and evaluates the chosen classifier
3. writes plots into outputs/activity-recognition/
4. writes a short report to outputs/activity-recognition/report.md

If signals and labels live in separate files, the workflow can prepare the labeled stream directly:

uv run ts-agents workflow run activity-recognition \
  --input signals.csv \
  --time-col ts \
  --value-cols x,y,z \
  --labels-input segments.csv \
  --labels-start-col start \
  --labels-end-col end \
  --label-col activity \
  --output-dir outputs/activity-recognition

Customize the workflow

uv run ts-agents workflow run activity-recognition \
  --input data/demo_labeled_stream.csv \
  --label-col label \
  --value-cols x,y,z \
  --window-sizes 32,64,128 \
  --classifier minirocket \
  --metric balanced_accuracy \
  --labeling majority \
  --stride 16 \
  --n-splits 5 \
  --output-dir outputs/activity-recognition

Use the lower-level tool run select_window_size_from_csv and tool run evaluate_windowed_classifier_from_csv only when you need more manual control than the workflow surface provides.

Real data option: WISDM (UCI, CC BY 4.0)

If you want a real-world dataset:

python data/make_demo_labeled_stream_wisdm.py \
  --subject 1600 --device watch --sensor accel \
  --activities walking,jogging,sitting,standing \
  --trim-policy per_class_seconds \
  --per-class-seconds walking=180,jogging=60,sitting=180,standing=180 \
  --out data/demo_labeled_stream.csv

Then rerun:

uv run ts-agents workflow run activity-recognition \
  --input data/demo_labeled_stream.csv \
  --label-col label \
  --value-cols x,y,z \
  --output-dir outputs/activity-recognition-wisdm

What to say in the demo (15-30 seconds)

• "We have a long labeled sensor stream; window size matters."
• "We search a small set of candidate windows."
• "We pick the best by balanced accuracy (better under class imbalance)."
• "We evaluate and output confusion matrix + plots for a quick report."

Guardrails / common failure fixes

• If n_windows is very small: reduce window_size, reduce stride, or increase dataset length.
• If one class dominates: keep balance=segment_cap and lower max_windows_per_segment.
• If performance is unstable: run multiple seeds and report mean/std.

Outputs to include in a blog post or repo README

• 1 plot: window_scores.png
• 1 plot: confusion_matrix.png
• 1 short table: best window + balanced accuracy
• 1 command block: uv run ts-agents workflow run activity-recognition --input <CSV> --label-col label --value-cols x,y,z

Report generation standard (Quarto PDF)

For polished deliverables, convert workflow outputs into a Quarto report:

1. Create/update outputs/reports/activity-recognition.qmd with:
   • dataset summary
   • window-size sweep figure
   • confusion matrix figure
   • key metrics table and conclusions
2. Render to PDF:

quarto render outputs/reports/activity-recognition.qmd --to pdf

Use clear sectioning, professional figure captions, and reproducible command snippets.