A Review of Geometric Deep Learning Technique for Emotional Detection Systems in the Context of Assistive Learning

Abstract

The study of Emotional Detection Systems (EDS) has gained significant attention in recent years due to its importance in understanding human reactions during interactions with both other humans and machines. EDS involves the use of computer systems to identify human emotions, which can be expressed through various modalities such as facial expressions, text, and speech. While humans can naturally recognize emotions, enabling computer systems to do so remains a complex challenge. Although emotion detection technologies have been in development for decades, there is a continuous need for research to enhance their accuracy and effectiveness, as many existing systems still fall short of optimal performance. Recently, Geometric Deep Learning (GDL) techniques have emerged as a promising alternative to traditional machine learning approaches. By considering the geometric properties of data, GDL enables more efficient learning beyond conventional Euclidean spaces. In every human being, emotion varies from individual even when he or she have the same intonation and facial expression but which may not correspond to their internal state. Also, current EDS may find it difficult to detect the mind or state of mind such as culture, situation, intent, emotion and suppression. More so, some emotion categories are under-represented which makes it hard for models to generalize and also, many EDS are trained on the data sets which may not provide enough sufficiently diverse in culture, age, gender or rely on spontaneous emotions.  However, its application to emotion detection remains relatively unexplored. This paper provides a critical review of various modalities used in emotion detection systems. It also analyzes different techniques and approaches, with a particular focus on assessing whether GDL techniques have been integrated into emotion detection models. Additionally, existing emotion detection datasets and performance results from recent studies are discussed. The report clearly states that the adoption of GDL is still partial and not yet universal across all commonalities and there is room for maturation in term of graph-construction best practices. The paper concludes with insights on current challenges and directions for future research. The future direction of GDL in EDS will integrate GDL’s structural intelligence with multimodal, interpretable, and privacy-preserving machine learning paradigms. Such integration will enable adaptive, human-centric systems capable of perceiving and responding to emotional states with better accuracy, empathy, and ethical responsibility.