Research on a Weakly Supervised Algorithm Based on Few-Shot Learning

Abstract

To address the issues of weakly supervised learning (WSL) such as limited feature diversity, poor adaptability to new scenarios, and the incomplete intra-class feature coverage and low accuracy in complex scenarios inherent to few-shot learning (FSL), this paper proposes a weakly supervised algorithm based on few-shot learning. The algorithm first constructs a rule-based system through five core modules to generate approximately accurate pseudo-labels for large-scale unlabeled data. Subsequently, an intra-class feature expansion optimization strategy is designed to mine effective features using methods such as weighted distance metrics, optimizing the feature space and reducing the risk of model overfitting. By fully exploiting the effective information in data and optimizing the structure of the feature space, the algorithm compensates for the lack of intra-class feature diversity in FSL, effectively alleviates overfitting, and improves classification accuracy and generalization in complex scenarios. Experimental results demonstrate that, compared with existing literature, the proposed few-shot feature expansion algorithm under weak supervision achieves at least a 1.18% improvement in accuracy, 2.30% in precision, 1.19% in recall, and 1.76% in F1 score. For strongly supervised algorithms, the improvements are at least 4.55%, 4.49%, 4.60%, and 4.55%, respectively. The proposed approach provides a novel solution for model training under data scarcity conditions.