Combating Negative Transfer From Predictive Distribution Differences.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Computer Science

Background:

• Domain adaptation (DA) uses labeled source data for target domains with scarce labels.
• Differences in predictive distributions between source and target domains can cause negative transfer, degrading performance.
• Negative transfer is particularly severe in multiclass classification problems.

Purpose of the Study:

• To introduce the concept of positive transferability for assessing domain synergy.
• To propose a criterion for measuring positive transferability between sample pairs across domains.
• To develop a predictive distribution matching (PDM) framework to mitigate negative transfer and enhance DA.

Main Methods:

• Introduced positive transferability to quantify the synergy between source and target domains.
• Developed a criterion to measure positive transferability based on prediction distributions.
• Proposed a predictive distribution matching (PDM) regularizer and framework to learn target classifiers by prioritizing source data with high positive transferability.

Main Results:

• The PDM framework demonstrated robust performance across benchmark datasets (Sentiment, Reuters, Newsgroup) in both binary and multiclass settings.
• The proposed method effectively mitigated the negative impact of domain shift, outperforming state-of-the-art DA methods.
• Validation in a real-world water cluster molecule identification task confirmed the framework's practical efficacy.

Conclusions:

• Positive transferability offers a valuable metric for assessing domain synergy in DA.
• The PDM framework successfully addresses the challenge of negative transfer, leading to improved predictive performance.
• This approach provides a robust solution for DA tasks, especially when domain distributions differ significantly.