SHL Challenge 2026 — ML Pipeline Tutorial

A four-notebook tutorial series teaching the fundamentals of a classical machine learning pipeline for Human Activity Recognition (HAR), built on the SHL Challenge 2026 dataset.

---

Overview

The series walks through every stage of a HAR pipeline end-to-end — from raw sensor files to trained and evaluated classifiers — using smartphone inertial sensor data collected at four body positions.

Notebook	Topic	Key output
01_data_loading.ipynb	Data loading & preprocessing	Preprocessed .npy arrays
02_visualization.ipynb	Data visualization	Signal plots, FFT spectra, PCA
03_feature_extraction.ipynb	Feature extraction	Feature matrices (N, F)
04_model_training.ipynb	Model training & evaluation	Trained models, results comparison

---

Dataset

The notebooks use the SHL Challenge 2026 dataset. You must download it separately from the official source:

🔗 http://www.shl-dataset.org/

The dataset is not included in this repository.

Expected directory structure

After downloading, place the data so the project root looks like this:

SHL-Tutorial/
├── Data/
│   ├── train/
│   │   └── Bag/
│   │       ├── Acc_x.txt
│   │       ├── Acc_y.txt
│   │       ├── Acc_z.txt
│   │       ├── Gyr_x.txt
│   │       ├── Gyr_y.txt
│   │       ├── Gyr_z.txt
│   │       ├── Mag_x.txt
│   │       ├── Mag_y.txt
│   │       ├── Mag_z.txt
│   │       └── Label.txt
│   ├── train 2/
│   │   └── Hand/  (same files)
│   ├── train 3/
│   │   └── Hips/  (same files)
│   ├── train 4/
│   │   └── Torso/ (same files)
│   └── validation/
│       ├── Bag/   (same files)
│       ├── Hand/  (same files)
│       ├── Hips/  (same files)
│       └── Torso/ (same files)
├── 01_data_loading.ipynb
├── 02_visualization.ipynb
├── 03_feature_extraction.ipynb
├── 04_model_training.ipynb
└── README.md

Dataset format

Each sensor file is a plain text matrix of shape (N_frames, 500):

• Rows — time windows (frames)
• Columns — 500 raw samples per frame (5 seconds × 100 Hz)

Split	Users	Frames	Labels
Train	User 1	196,072	✓
Validation	Users 2 & 3	28,789	✓
Test	Users 2 & 3	92,726	✗ (challenge submission)

Sensors: Accelerometer, Gyroscope, Magnetometer (3 axes each = 9 channels)
Positions: Bag, Hand, Hips, Torso

Activity classes:

Code	Activity	Code	Activity
1	Still	5	Car
2	Walking	6	Bus
3	Run	7	Train
4	Bike	8	Subway

---

Preprocessed Data

Notebook 1 generates a preprocessed/ directory of .npy arrays. This folder is also not included in the repository — run 01_data_loading.ipynb first to generate it.

preprocessed/
├── X_train_Bag.npy         X_val_Bag.npy
├── X_train_Hand.npy        X_val_Hand.npy
├── X_train_Hips.npy        X_val_Hips.npy
├── X_train_Torso.npy       X_val_Torso.npy
├── y_train_{position}.npy  y_val_{position}.npy
│
│   (generated by Notebook 3)
├── X_feat_train_Hips.npy   X_feat_val_Hips.npy
├── y_feat_train_Hips.npy   y_feat_val_Hips.npy
└── feature_names_Hips.txt

Array shapes and approximate sizes:

File	Shape	Size
X_train_{position}.npy	(196072, 9, 500)	~3.5 GB each
X_val_{position}.npy	(28789, 9, 500)	~520 MB each
X_feat_train_Hips.npy	(196072, ~219)	~170 MB
X_feat_val_Hips.npy	(28789, ~219)	~25 MB

Storage note: Full preprocessing of all 4 positions requires approximately 17 GB of free disk space.

---

Installation

Requirements

• Python ≥ 3.10
• Conda or virtualenv recommended

Setup

# Clone the repository
git clone https://github.com/<your-username>/SHL-Tutorial.git
cd SHL-Tutorial

# Create and activate environment
conda create -n shl-tutorial python=3.11
conda activate shl-tutorial

# Install dependencies
pip install -r requirements.txt

# Launch Jupyter
jupyter notebook

Dependencies

numpy>=1.26
pandas>=2.0
matplotlib>=3.7
scipy>=1.11
scikit-learn>=1.4
tqdm
jupyter

Or install directly:

pip install numpy pandas matplotlib scipy scikit-learn tqdm jupyter

---

Notebook Guide

01 — Data Loading & Preprocessing

Estimated runtime: 15–30 min (full dataset, all 4 positions)

Loads raw .txt sensor files, checks for NaN/Inf/zero-variance frames, applies per-sample z-score normalization, and saves stacked (N, 9, 500) arrays.

Key concepts taught:

• Sliding window representation of time-series
• Per-sample vs. global normalization
• Majority-vote label assignment for transition windows

Quick run: Set MAX_FRAMES = 5000 in the relevant cells to test the pipeline in seconds before committing to a full run.

---

02 — Data Visualization

Estimated runtime: 5–10 min (loads from preprocessed arrays)

Eight visualization sections building intuition about the sensor signals before any modeling. All plots are implemented as reusable parameterized functions.

Section	What you see
Raw signal plots	Per-class waveforms for any sensor axis
Intra-class variability	Frame overlays showing within-class spread
Sensor comparison	Acc vs Gyr vs Mag for the same activity
Position comparison	Same activity across all 4 phone positions
Statistical boxplots	Mean, Std, RMS, Peak-to-peak per class
Correlation heatmap	Inter-axis Pearson correlation by class
FFT spectra	Dominant frequency content per class
PCA projection	2D separability of classes from simple features

---

03 — Feature Extraction

Estimated runtime: 5–10 min (full dataset)

Transforms each (9, 500) raw frame into a flat feature vector using three feature groups:

Group	Count	Examples
Time-domain	13 × 9 = 117	mean, std, RMS, skewness, kurtosis, ZCR
Frequency-domain	10 × 9 = 90	dominant frequency, spectral entropy, band energies
Cross-axis	12	SMA, vector magnitude, Acc–Gyr correlation
Total	~219	(after zero-variance removal)

Also includes feature quality checks (NaN/Inf imputation, zero-variance removal, outlier clipping) and a feature importance ranking via a lightweight Random Forest.

---

04 — Model Training & Evaluation

Estimated runtime: 30–90 min (dominated by SVM; see note below)

Trains and evaluates four classifiers on the feature matrices from Notebook 3.

Model	Train data	Notes
Random Forest	Unscaled	200 trees, balanced class weight
KNN	Scaled	k=11, distance weighting
SVM	Scaled (subsample)	RBF kernel, C=10 — subsample to ~30K for tractability
HistGradientBoosting	Unscaled	300 rounds, early stopping

Results reported: Accuracy, Macro F1, Weighted F1, per-class F1, confusion matrices, train/inference time.

SVM warning: sklearn.svm.SVC with RBF kernel scales as O(n²)–O(n³). Training on the full 196K samples is not feasible on most laptops. The notebook includes a stratified subsample cell — use it.

---

Notes on Evaluation

The SHL Challenge 2026 test set has no labels — it is intended for official leaderboard submission only. Throughout this tutorial, the labeled validation set (Users 2 & 3) serves as a proxy test set.

This means:

• Reported validation metrics are slightly optimistic, as some modeling choices (hyperparameters, feature set) were guided by observing validation performance
• For a production pipeline, hyperparameter tuning should be performed on a held-out slice of the training data, with the validation set reserved for final evaluation

---

Results (Hips position, full train → validation)

Indicative results — exact numbers will vary with hardware, NumPy/sklearn versions, and whether SVM is trained on a subsample.

Model	Accuracy	Macro F1	Train Time
Random Forest	~75–80%	~70–75%	~3–5 min
KNN (k=11)	~65–70%	~60–65%	~2 min (inference)
SVM (subsample)	~65–70%	~60–65%	~5–10 min
HistGradientBoosting	~78–83%	~72–78%	~3–6 min

Vehicle transport modes (Car, Bus, Train, Subway) are the hardest to distinguish — they share similar posture and differ mainly in vibration frequency profile.

---