Research
Human Activity Recognition (WISDM 2.0)
My MSci dissertation: a comparison of classical machine-learning and deep-learning approaches to recognising human activity from smartphone sensor data, using WISDM 2.0.
Result
Gradient Boosting reached F1 0.981; an LSTM reached 0.980.
01
Problem
Smartphone accelerometer data is noisy, varies a lot from person to person, and is easy to overfit. I compared classical and deep-learning approaches using an evaluation set-up meant to reflect how well a model works on someone it has never seen.
02
How it was evaluated
Every model was tested on users who weren't in the training data. That matters: with a random split, the same person's data ends up on both sides, and results can look much better than they really are.
03
Results
- Gradient Boosting — F1 0.981
- LSTM — F1 0.980
- LSTM-CNN — F1 0.958
04
Features
Time- and frequency-domain features: mean, standard deviation, minimum and maximum, signal magnitude, skew, kurtosis, energy, signal magnitude area, dominant FFT frequency, power spectral density and peak-based features.
05
What I found
A classical pipeline with careful feature engineering matched the best deep-learning model. It's also easier to inspect and cheaper to run, which makes it the more practical choice.