Domain Space Transfer Extreme Learning Machine for Domain Adaptation.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Computer Vision
• Natural Language Processing

Background:

• Extreme Learning Machines (ELM) excel in classification and regression but fail when training and testing data distributions differ.
• Domain adaptation is crucial in real-world scenarios where data distributions often vary between source and target domains.
• Existing ELM methods perform poorly in unsupervised domain adaptation due to violated identical distribution assumptions.

Purpose of the Study:

• To develop a novel ELM-based algorithm, Domain Space Transfer ELM (DST-ELM), for unsupervised domain adaptation.
• To enable accurate predictions on target domain data without available labels by learning a domain-invariant feature space.
• To maintain computational efficiency through a low number of hidden nodes in the ELM components.

Main Methods:

• DST-ELM reconstructs source and target data into a shared, domain-invariant space.
• A feature space learning network preserves target domain structure and discriminative information.
• Source data is projected to minimize distribution distance, followed by an adaptive ELM classifier trained on transferred source labels.

Main Results:

• The proposed DST-ELM method effectively handles unsupervised domain adaptation problems.
• Experiments on real-world image and text datasets demonstrate superior accuracy compared to existing methods.
• The algorithm maintains high computational efficiency due to its low complexity ELM architecture.

Conclusions:

• DST-ELM offers a robust solution for unsupervised domain adaptation by creating a domain-invariant feature space.
• The method successfully bridges the distribution gap between source and target domains, enhancing predictive performance.
• DST-ELM presents a computationally efficient and accurate approach for real-world machine learning applications with domain shift.