Worst-case scenario model evaluation

 We are facing many complex modeling problems due to issues with training data sets. For example, it is often difficult to maintain the inductive learning principle in play if there is a lot of drift, poor distributional robustness #DRO, f-divergence, and other problems related to data quality. As a result, we are noticing performance degradation across demographic, temporal, and geographic groupings daily. One of the approaches we are exploring right now is based on synthetic data-based enrichment, carefully evaluated to avoid introducing unwanted biases into the model. On the other hand, the dual reformulation solution presented in this paper is a great way to mitigate the performance drop in specific populations. My approach to such problems in the past involved using singular value decomposition #SVD. The objective was somewhat similar - changing the dimensionality of the underperforming population. 

https://openreview.net/pdf?id=nehzxAdyJxF

#distributionallyrobustoptimization #fdivergence #dualreformulation #singularvaluedecomposition