Classroom climate detection using machine learning

• Main Author: Mangesh, Dhamnaskar Ruchi
• Other Authors: Justin Dauwels
• Format: Theses and Dissertations
• Language: English
• Published: 2018
• Subjects: DRNTU::Engineering::Electrical and electronic engineering
• Online Access: http://hdl.handle.net/10356/76349
• Institution: Nanyang Technological University

The project proposes a system for automatically determining whether the classroom climate exerts a positive or negative influence on the learning process of the children. The focus of the current study is to explore various combinations of machine learning techniques for classroom climate detection. Several cues are embedded in the conversations and other sounds that are generated in a classroom. Many studies suggest that such cues from an audio signal can help in inferring whether the classroom environment is positive or negative. Hence, audio recordings of classroom activities are used as the basis of generating the data set required for the machine learning algorithms. The study was conducted based on a limited real-life data set (also referred to as ‘in-the-wild-data set’). The audio files are labelled as having activities with either positive or negative climate by a person skilled in the art such as a counsellor. Further, speaker diarisation is performed manually by annotating the audio files with speakers participating in the interaction. A set of features is extracted from each of the audio files. A feature matrix is generated based on statistical analysis of the set of features, the speaker diarisation, and class labels. The proposed exploratory study was implemented using python language and made use of various modules provided by the scikit and sklearn python libraries. During the process of cross validation, the data set is divided in train and test data sets in multiple iterations. Further, feature selection, hyper parameter tuning of the classifiers, and training of the classifiers is performed within each iteration of a cross validation loop. In the step of feature selection, irrelevant and redundant features are removed using feature selection techniques to output a subset of features. The subset of features is used to tune hyper parameters of each classifier and train the classifier model. The test data set is used to evaluate the performance of the trained classifier model. Finally, the performance of each classifier is evaluated by combining the performances of all the iterations of the cross-validation process.