Federated open intent classification via granular-ball knowledge representation.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Natural Language Processing

Background:

• Open Intent Classification (OIC) identifies known and unknown user intents in open-world settings.
• Existing OIC methods require centralized data, posing privacy challenges.
• Federated learning offers a decentralized approach but faces issues like class inconsistency and uncontrolled pseudo-unknown sample generation.

Purpose of the Study:

• To propose a privacy-preserving federated learning method for Open Intent Classification.
• To address challenges of class inconsistency, client-specific distributions, and pseudo-unknown sample generation in federated OIC.
• To enable both known intent recognition and unknown intent detection in decentralized settings.

Main Methods:

• Federated Open intent Classification via Granular-Ball representations (FedOC-GB).
• Clients construct granular-ball representations to capture local class structures and generate structure-aware pseudo-unknown samples.
• Server aggregates model parameters and granular-ball knowledge for a unified global model and knowledge base.

Main Results:

• FedOC-GB consistently outperforms state-of-the-art federated OIC methods on benchmark datasets.
• Achieved up to 4.3% higher F1-All and 3.7% higher ACC on average.
• Demonstrated the effectiveness of granular-ball representation for privacy-aware and structure-preserving federated OIC.

Conclusions:

• FedOC-GB successfully enables privacy-aware and structure-preserving federated open intent recognition.
• The granular-ball representation is key to overcoming challenges in decentralized OIC.
• The proposed method offers a robust solution for real-world OIC applications with privacy constraints.