Human-Activity-Recognition

Introduction

This is my bonus project for the course Pattern Recognition and Machine Learning (2022 Winter Semester). The problem statement is:
Human activity recognition is the problem of classifying sequences of data recorded by specialized harnesses or smartphones into known well-defined human activities. The goal of this machine learning project is to build a classification model that can precisely identify human fitness activities. Working on this machine learning project will help you understand how to solve multiclass-classification problems.

Dataset

The Human Activity Recognition database was built from the recordings of 30 study participants performing activities of daily living (ADL) while carrying a waist-mounted smartphone with embedded inertial sensors. The objective is to classify activities into one of the six activities performed:

• Walking
• Walking Upstairs
• Walking Downstairs
• Sitting
• Standing
• Laying

The obtained dataset has been randomly partitioned into two sets, where 70% of the volunteers were selected for generating the training data and 30% for the test data.
Each row consists of 561 features, 1 label (Activity) and one column denoting the subject ID. The dataset is available at Kaggle.

Pipeline

My task for this project was to implement an end-to-end pipeline that could take the dataset and output predictions for the test data along with some metrics to evaluate its performance. The pipeline consists of the following components:

• Importing Data: Train and Test datasets can be input either as URLs (relative or absolute) or as Pandas DataFrames.
• Preprocessing: NULL values are removed from the dataset and the features are scaled using the Scaler chosen by the user (StandardScaler/MinMaxScaler/MaxAbsScaler). The column containing subject IDs is removed since it is not required for making predictions.
• Training the Model: Using the model parameters I found while exploring the task, the model is trained on the training data. I found LinearDiscriminantAnalysis and XGBoostClassifier to be the best performing models for this task, hence I have given the user the option to choose either one of these models.
• Evaluation: The model is evaluated on the test data and predictions are stored in the working directory under the filename predictions.csv. Accuracy on the test dataset along with a classification report containing different metrics is printed.

Instructions for running the notebook

• Use git clone https://github.com/sawmill811/Human-Activity-Recognition.git in your terminal to clone the repository. If you wish to run the file on Google Colaboratory, then here is the link to the notebook.
• To run the pipeline, simply run the first cell in the HAR.ipynb notebook - this will import the necessary libraries.
• Next, skip all cells under the heading "Exploring the Task" and run the cells under the heading "Creating a Pipeline".
• The predictions for the test data will be stored in the working directory under the filename predictions.csv.

Pipeline Performance

• LDA: Predictions are obtained in less than 10 seconds on my PC as well as Google Colaboratory.
• XGBoostClassifier: Predictions are obtained in about 30 seconds on my PC (using GPU) and 1.5 minutes (using CPU). It takes about 4 minutes to run the pipeline on Google Colaboratory.

Troubleshooting

If you are getting an error message then try removing the tree_method = "gpu_hist" parameter from the XGBClassifier model. This is used to speed up the training process using GPU resources. Other than that, the code should run fine on most systems.