Reducing Human-Induced Label Bias in SMS Spam with Context-Enhanced Clustering (CEC)

Short Message Service (SMS) is a widely used text messaging feature available on both basic and smartphones, making SMS spam detection a critical task. Supervised machine learning approaches often face challenges in this domain due to their dependence on manually crafted features, such as keyword detection, which can result in simplistic patterns and misclassification of more complex messages. Furthermore, these models can exacerbate human-induced bias if the training data include inconsistent labeling or subjective interpretations, leading to unfair treatment of specific keywords or contexts. We propose a Context-Enhanced Clustering (CEC) approach to address these challenges by leveraging contextual metadata, adaptive thresholding, and modified similarity measures for clustering. We evaluate our approach using the English SMS spam dataset source from UC Irvine’s Machine Learning Repository. CEC identifies representative samples from the SMS dataset to fine-tune LLMs such as ChatGPT-4, improving the robustness and fairness of spam classification. Our approach outperforms traditional clustering techniques such as K -means and DBSCAN in mitigating bias, as demonstrated through experiments measuring a balanced accuracy of $85 \%$ and a treatment equality difference (TED) of precisely zero. When used to identify representative samples to fine-tune ChatGPT-4, the CEC achieves a balanced accuracy of $98 \%$, an equal opportunity of difference (EOD), and a treatment equality difference (TED) of zero. These results significantly reduce human-induced bias while maintaining high classification accuracy.