HomeAbstract of 8th-areconfDistilTab–Multimodal Fusion with Distilbert-based Tabular Transformer for Learning Style Classification

DistilTab–Multimodal Fusion with Distilbert-based Tabular Transformer for Learning Style Classification

  • Mar 1, 2026
  • Posted by:
  • Category: Abstract of 8th-areconf
No Comments


Abstract Book of the 8th International Conference on Applied Research in Education

Year: 2026

[PDF]

DistilTab–Multimodal Fusion with Distilbert-based Tabular Transformer for Learning Style Classification

Amruth John, Binu Thomas

ABSTRACT:

Recently, an adaptive learning system personalised education to each student by identifying the most effective learning approach to enhance understanding and retention. A crucial component of this system is recognising individual learning styles, including Visual, Auditory and Kinesthetic (VAK), to maximise every student’s learning potential. Traditional models rely on textual indicators such as linguistic cues or behavioural traces from the learning management system. Therefore, a Distil Bidirectional Encoder Representations from Transformer (DistilBERT) with Tabular Transformer (TabTransformer) is proposed as DistilTab in this research by integrating textual and behavioural information. The novelty of the proposed DistilTab lies in the usage of DistilBERT and Tabtransformer models to achieve deep multimodal alignment while enabling context-aware fusion of textual and behavioural data for accurate VAK learning style classification. Unlike previous fusion-based approaches, DistilTab achieves deep modality alignment through contextualised projection instead of shallow concatenation, thereby improving cross-modal representation learning and improving overall classification accuracy. The DistilBERT model functions as a transformer encoder in TabTransformer, which generates semantic embeddings from the VAK dataset and contextual embeddings from the OULAD dataset. A network projection is employed to generate network embeddings, which are then concatenated to form a unified feature representation. Finally, Multi-Layer Perceptron (MLP) was utilised for multi- class classification. From the results, the proposed DistilTab model achieved higher accuracy (98.52%) and F1-score (98.54%) than the existing Convolutional Neural Network with Natural Language Processing (CNN-NLP).

Keywords: Distil Bidirectional Encoder Representations From Transformer; Learning Style Classification; Modality-Specific Embeddings; Multi-Layer Perceptron; Tabular Transformer