Multiclass heterogeneous domain adaptation via bidirectional ECOC projection.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Computer Science

Background:

• Heterogeneous domain adaptation addresses training models for target domains with different feature spaces using source domain data.
• Existing methods often separate common space learning and classifier training, leading to complex models and increased parameters.

Purpose of the Study:

• To propose a novel, unified approach for heterogeneous domain adaptation that integrates common space learning and classifier training.
• To enhance the efficiency and effectiveness of domain adaptation by simultaneously learning shared representations and classifiers.

Main Methods:

• Introduced HDA-ECOC, a bidirectional Error-Correcting Output Codes (ECOC) projection method.
• Projects both inputs and outputs (labels) into a common ECOC coding space for simultaneous learning.
• Utilizes Maximum Mean Discrepancy (MMD) to enforce consistency between projected instances from different domains, incorporating unlabeled target data.

Main Results:

• Demonstrated effectiveness and efficiency in cross-lingual text classification and cross-domain digital image classification tasks with heterogeneous feature spaces.
• The proposed method achieved superior performance compared to existing approaches in heterogeneous domain adaptation scenarios.

Conclusions:

• HDA-ECOC offers a unified and efficient solution for heterogeneous domain adaptation by enabling simultaneous common space learning and classifier training.
• The method effectively leverages unlabeled target data and heterogeneous features for improved cross-domain prediction.