Siamese meta-learning network for social disputes based on multi-head attention

Area of Science:

• Machine Learning
• Artificial Intelligence
• Computer Science

Background:

• Few-shot learning is crucial when labeled data is scarce.
• Siamese networks are common for meta-learning based few-shot classification but have limitations.
• Existing Siamese networks struggle with label quality dependence, long-distance data dependencies, and performance on complex or unbalanced datasets.

Purpose of the Study:

• To propose a novel Multi-Head Attention Siamese Meta-learning network (MASM).
• To overcome the limitations of traditional Siamese networks in few-shot classification.
• To enhance the model's ability to handle label quality issues, capture long-range dependencies, and perform well on complex/unbalanced data.

Main Methods:

• Developed a Multi-Head Attention Siamese Meta-learning network (MASM).
• Employed synonym substitution to mitigate prototype vector computation's dependence on class label quality.
• Integrated a multi-head attention mechanism to capture long-distance data dependencies via global perception.

Main Results:

• MASM demonstrated strong performance across four benchmark datasets.
• The model achieved good results when applied to a novel social dispute dataset.
• The proposed methods effectively addressed the identified limitations of standard Siamese networks.

Conclusions:

• The MASM model offers a significant improvement over traditional Siamese networks for few-shot classification.
• The integration of synonym substitution and multi-head attention enhances robustness and performance.
• MASM shows promise for applications in diverse fields, including social dispute analysis.