Local regression transfer learning with applications to users' psychological characteristics prediction

Area of Science:

• Computational psychology
• Machine learning
• Data science

Background:

• Accurate prediction of web users' psychological characteristics is crucial.
• Supervised learning models face generalization challenges due to data distribution shifts between training and testing sets.

Purpose of the Study:

• To develop novel local regression transfer learning methods to enhance the generalization capability of computational models for predicting user psychological characteristics.
• To address the limitations of existing supervised learning models in handling distribution differences in web user data.

Main Methods:

• Proposed local regression transfer learning techniques, including k-nearest-neighbor and clustering reweighting, to assess training instance importance.
• Developed a weighted risk regression model for psychological characteristic prediction.
• Introduced an adaptive parameter-setting method for scenarios with unlabeled test data.

Main Results:

• Experimental validation demonstrated significant improvements in model generalization.
• The proposed methods effectively predicted user personality and depression across diverse datasets (e.g., different genders, districts).

Conclusions:

• Local regression transfer learning offers a robust solution for improving the generalization of psychological characteristic prediction models.
• The developed methods successfully mitigate the impact of data distribution discrepancies, leading to more reliable predictions.