This classification problem was tackled as part of a Science Sprint held at UT Austin (February 2019). The provided data was accelerometer and gyrometer inputs caputured from a Samsung Galay II as people performed various tasks, and the objective was to classify the action to which each input belonged. The dataset can be found here.
I achieved this with an voting classifier ensemble consisting of a logistic classifier, support vector machine, and multilayer perceptron. The Jupyter notebook summarizes my feature exploration and model selection process. My final test accuracy (only measured after my model was entirely complete) was 96%.
I would like to thank the university, the College of Natural Sciences, Dr. Beasley, and all the volunteers for organizing the event.