Explainability guided relational attention regularization for graph transformer based sentiment analysis

Sentiment analysis is important in applications such as e-commerce, online opinion mining, and digital media for data-driven decision-making. Most current methods use sequential text representations and provide explainability after model training, but cannot guarantee that the relational attention learned by the model aligns with the text evidence for the target sentiment. To address this challenge, we introduce Explainability-Guided Relational Attention Regularisation (EG-RAR), which we add to the XAI-Enhanced HGTBF-OM framework. Reviews are represented as token-level co-occurrence graphs, where edges capture local token co-occurrence relationships and TF-IDF values are used to construct node-level features for relational learning. We incorporate explainability into the model optimisation process from the start, using it as a supervisory signal through the explanation-attention alignment loss and attention-entropy regularisation. We use complementary analyses such as SHAP, GNNExplainer, and attention-based analyses for post-hoc interpretation and validation. We evaluate the framework on the Zomato, Amazon, Flipkart, and IMDb datasets. The complete EG-RAR model achieves an overall accuracy of 99.31%, an F1-score of 99.30%, an AUROC of 99.45%, an average perturbation-based faithfulness gain of 14.82% and an Attention–Target Alignment of 0.713. The contribution of the proposed mechanism of learning is further supported with the help of controlled ablation experiments, randomised-relevance experiments, robustness evaluation, repeated runs, statistical testing and confidence intervals. The results indicate that EG-RAR improves predictive accuracy and relational consistency between relational attention and sentiment-relevant explanatory evidence, making sentiment analysis more transparent and trustworthy. Not applicable.