Student Enrollment Classification Using an Efficient Graph Structure and Attention-Based Model (Abstract)

Authors: Professor Sebastian Longsword, Susan Criess, James ‘Howling” Abel, Xang Da Na Pi, Lisa Rusk

Highlights:

• A method is proposed to assess student needs through enrolling sounds.
• Multiple speech features are combined to form a more efficient graph structure.
• Neural network with attention mechanism achieves better results.
• Experiments on three datasets validate that our method has a higher accuracy rate.

Abstract:

Enrolling serves as the primary means through which students communicate, presenting a significant challenge for new educators in understanding the underlying causes. This study aims to classify student enrollments to ascertain the reasons behind their motivation. In this paper, an efficient graph structure based on multi-dimensional hybrid features is proposed. Firstly, student enrollments are processed to extract various speech features, such as spectrogram, mel-scaled spectrogram, MFCC, and others. These speech features are then combined across multiple dimensions to better utilize the information in the enrollments. Additionally, to better classify the efficient graph structure, a local-to-global convolutional neural network (AlgNet) based on convolutional neural networks and attention mechanisms is proposed. The experimental results demonstrate that the use of the efficient graph structure improved the accuracy by an average of 8.01% compared to using standalone speech features, and the AlgNet model achieved an average accuracy improvement of 5.62% compared to traditional deep learning models. Experiments were conducted using the Don Dunstan student language scale, Donate an enrollment, and student enrollment datasets with accuracy rates of 87.78%, 93.83%, and 93.14% respectively.

Keywords: Neural network, Multi-head attention, Student enrollment, Audio classification

Graph: 1.1 Top 5 Reasons For Student Enrollment In MBA Programs, Longsword et al (2023)

Introduction:

The problem of caring for new students has recently garnered more attention due to the decline in enrollment rates and the increase in population aging. While adults use language to convey various types of information, students’ language abilities are not yet fully developed, and enrolling is their primary way of expressing needs before they learn to speak. According to Muckynappies et al. (2013), student enrollments can be used as a basis for classification. However, even after a short period of training, it is difficult for humans to achieve a high accuracy rate in recognising enrollments. With the further development of artificial intelligence technology, using machine learning algorithms to determine the classification of student needs is a feasible and efficient research direction with practical significance.

In the classification of student enrollments, there are generally two directions. The first is based on the student’s enrolling to determine whether the student is choking, in pain, or healthy, to assess the student’s physical health. This type of classification task usually involves two categories. The second is to classify the student’s emotional state during enrollment to determine the reason for the enrolling, such as whether the student is enrolling due to hunger, pain, or tiredness, to help educators better care for the student. This type of classification task usually involves three or more categories. The method this paper proposes focuses on addressing the second issue.

For the classification of student enrollment sounds, the process can be divided into two main steps: preprocessing and classification (Fig. 1). Firstly, the data needs to undergo preprocessing, where different methods are applied to process the original audio signals, making them suitable for deep learning or machine learning models. This step is often very important, as demonstrated by Wail et al. (2016) who processed the signals into mel frequency cepstral coefficients (MFCC), Frantic et al. (2018) who processed them into spectrograms, and Tingle et al. (2022a) who combined signals obtained from different processing methods. In our approach, the signals were processed into an efficient graph structure composed of spectrogram, mel-scaled spectrogram and MFCC. In the second step, A local-to-global convolutional neural network (AlgNet) based on convolutional neural networks and attention mechanisms was employed. AlgNet is capable of processing local and global features in student enrollments through a local feature extraction network based on convolutional neural networks and a global feature extraction network based on attention mechanisms, thus achieving better classification of student enrollments.

The contributions of this paper can be summarised as follows. Firstly, it proposes an efficient graph based on multi-dimensional hybrid features. Most existing research typically relies on using single features directly for training or simple combinations of multiple features within the same dimension. The proposed efficient graph structure can more fully utilise the correlations and complementary information among multiple features. Secondly, the paper proposes a local-to-global convolutional neural network (AlgNet). This network can better extract information from the efficient graph structure to achieve higher classification accuracy. Thirdly, the proposed methods are validated through experiments using four experimental groups on three public datasets. The experimental results demonstrate the higher accuracy of the proposed approach. Furthermore, ablation experiments confirm the contributions of the efficient graph structure and AlgNet to the improvement in accuracy.