Abstract Novel spam with rapidly evolving content faces a scarcity of labeled data in its early stages. Yet, current detection models rely heavily on large datasets and high-dimensional features, leading to poor generalization and opaque decisions when data is scarce. This opacity hinders error tracing and limits their use in early threat detection and response. The belief rule base (BRB), as an expert system, demonstrates effective learning under small-sample conditions, and its rule-based reasoning mechanism provides decision interpretability. However, high-dimensional features may cause combination explosion. To address these issues, a BRB spam detection model based on the Discriminative term frequency-inverse document frequency (TF-IDF) method (DTI-BRB) is proposed in this paper. By discriminating whether terms are more indicative of ham or spam, the Discriminative TF-IDF method converts raw text into low-dimensional features, thereby effectively resolving the combination explosion problem inherent in the traditional BRB model. Through two case studies under small-sample conditions, the effectiveness of the proposed model is validated. With only 200 samples, it achieves accuracies of 91.5% and 95.5% in the two cases, respectively, exhibiting excellent predictive performance and interpretability.
Yang et al. (Wed,) studied this question.