Reproduce “Are You A Risk Taker? Adversarial Learning of Asymmetric Cross-Domain Alignment for Risk Tolerance Prediction” on OpenResearch

About this paper

Zhe Liu, Lina Yao, Xianzhi Wang, Lei Bai, Jake An

Researchers at UNSW and UTS, in collaboration with Raiz Invest Limited, developed an Asymmetric Cross-Domain Generative Adversarial Network (ADGAN) to predict consumers' financial risk tolerance. This system integrates sparse survey responses with abundant, unlabeled transaction data, effectively addressing data asymmetry, class imbalance, and data quality challenges to achieve higher prediction accuracy and Macro F1-scores.

Key insights

• Incorporating domain-specific knowledge through hierarchical feature engineering and embedding architectures significantly enhances the extraction of meaningful behavioral insights from raw consumer transaction data.
• A generative adversarial framework, carefully designed for asymmetric data scales and non-pairwise samples, can effectively bridge disparate data domains to improve prediction accuracy for a latent variable like financial risk tolerance.
• Fusing Gaussian noise with both input activity data and real survey data during training is crucial for increasing sample diversity, improving model robustness, and mitigating issues like the 'multivalued function phenomenon' in real-world datasets.

Summary by alphaXiv.