Emotion on the edge: An evaluation of feature representations and machine learning models

Abstract

This paper presents a comprehensive analysis of textual emotion classification, employing a tweet-based dataset to classify emotions such as surprise, love, fear, anger, sadness, and joy. We compare the performances of nine distinct machine learning classification models using Bag of Words (BoW) and Term Frequency-Inverse Document Frequency (TF-IDF) feature representations, as well as a fine-tuned DistilBERT transformer model. We examine the training and inference times of models to determine the most efficient combination when employing an edge architecture, investigating each model’s performance from training to inference using an edge board. The study underscores the significance of combinations of models and features in machine learning, detailing how these choices affect model performance when low computation power needs to be considered. The findings reveal that feature representations significantly influence model efficacy, with BoW and TF-IDF models outperforming DistilBERT. The results show that while BoW models tend to have higher accuracy, the overall performance of TF-IDF models is superior, requiring less time for fitting, Stochastic Gradient Descent and Support Vector Machines proving to be the most efficient in terms of performance and inference times.