Note , Part of the coursework for the Data Mining and Machine Learning module of MSc Data Science. A Kaggle competition was run as part of the module, I finished in 2nd place.
Objective
A solution to human activity recognition (HAR). The data was sampled from the WISDM (Wireless Sensor Data Mining) Labs dataset, which collects sensor readings from smartphones and modern mobile devices and mines them for useful knowledge. The dataset captures 36 users performing six human activities, ascending and descending stairs, sitting, walking, jogging, and standing, over fixed periods, using accelerometers.
The study walks through exploration and analysis of this data, then applies a range of machine-learning models to recognise the six activities. Being a labelled classification problem, several supervised ML approaches were tested, starting with classical models (Logistic Regression, SVM, Random Forest) and moving to neural networks (deep dense networks, CNNs), while varying the input features and tuning hyperparameters.
Models were evaluated for accuracy, with the best results obtained from a combination of multiple convolutional layers and a deep dense network with carefully tuned hyperparameters.