Songlei Xue

Semi-supervised fine-grained classification suffers from accumulating pseudo-label errors among potentially confusable classes when labeled data are scarce, yet negative-rule information that excludes classes at the attribute level remains unused. We propose AW-RS-ABL, a semi-supervised abductive learning method that fuses intra-class and inter-class attribute importance within rough sets and couples positive-rule confidence matching with negative-rule candidate-set reduction for pseudo-label correction. Samples, attributes, and labels form a decision information system $S=(U,A,F,d)$. Inter-class discriminative power is measured by positive-region dependency ${\gamma }_A\left(d\right)$ and intra-class stability by class-conditional entropy $H(a\mid d)$; the two are normalized and linearly fused into an attribute weight that selects a key attribute subset $A^{*}$. Weighted positive rules and high-confidence negative rules are built on $A^{*}$: negative rules shrink the candidate class set, and positive rules complete the correction via confidence matching. Rule consistency checking is embedded in the semi-supervised pseudo-label loop so that rule verification and perceptual model prediction co-evolve across training iterations. On six datasets spanning SAR target recognition, medical diagnosis, and agricultural classification at a reported 10% label ratio, AW-RS-ABL has the highest reported mean top-1 accuracy in each comparison, with displayed mean differences of 1.36–2.58 percentage points relative to FixMatch-Attr. Exact paired Wilcoxon tests with Holm correction support the differences for five datasets, whereas the difference for Dermatology is not significant. On the imbalanced Nursery test set, a representative seed yields 87.34% accuracy but 73.54% macro-F1 and 18.18% recall for the minority class. The reported formulation relies on conditional attributes, and its results remain conditional on their representation and binarization.