Incomplete Multisource Transfer Learning.

Area of Science:

• Machine Learning
• Artificial Intelligence

Background:

• Transfer learning adapts source knowledge to target domains, often with limited labels.
• Multiple data sources are common but may have incomplete or divergent class information.
• Directly merging incomplete sources degrades performance due to data divergence.

Purpose of the Study:

• To develop an effective multisource transfer learning method using incomplete data sources.
• To address challenges posed by missing class information and source divergence in transfer learning.

Main Methods:

• Proposes a two-directional knowledge transfer approach: cross-domain and cross-source.
• Cross-domain transfer utilizes latent low-rank learning with iterative structure learning.
• Cross-source transfer employs unsupervised manifold regularization and multisource alignment.

Main Results:

• The proposed method effectively compensates for missing data in both source and target domains.
• Mitigates marginal and conditional distribution discrepancies across domains and sources.
• Demonstrates superior performance on standard cross-domain benchmarks and synthetic datasets.

Conclusions:

• The developed incomplete multisource transfer learning model enhances knowledge transfer from incomplete sources.
• The bidirectional approach successfully handles data heterogeneity and missing information.
• Validates the model's effectiveness for real-world applications with incomplete data.