Introducing uncertainty in predictive modeling--friend or foe?

Area of Science:

• Computational chemistry
• Cheminformatics
• Machine learning in chemistry

Background:

• Chemical descriptors are crucial for predictive modeling in cheminformatics.
• Uncertainty in descriptor values can arise from experimental or computational limitations.
• The impact of descriptor uncertainty on predictive model performance requires thorough investigation.

Purpose of the Study:

• To assess the effect of varying degrees and types of uncertainty in chemical descriptors on the predictive performance of decision tree ensembles.
• To evaluate different strategies for handling uncertainty within decision tree ensemble models.
• To determine if complex uncertainty handling methods offer advantages over simpler approaches.

Main Methods:

• Systematic introduction of uncertainty into chemical descriptors across 16 public datasets.
• Application of state-of-the-art decision tree ensemble methods for predictive modeling.
• Evaluation of various strategies for managing uncertainty in descriptor data.
• Comparison of model performance using single-point samples versus full uncertainty distributions.

Main Results:

• Predictive performance of decision tree ensembles remained largely unimpaired despite significant uncertainty introduced into chemical descriptors.
• Practical utility of predictive models was not substantially compromised by descriptor uncertainty.
• Ensemble models performed comparably whether trained on single-point samples or full distributions of uncertain values.

Conclusions:

• Decision tree ensembles exhibit robustness towards uncertainty in chemical descriptors.
• There is no practical benefit to employing complex uncertainty handling techniques for chemical descriptors when using decision tree ensembles.
• The findings suggest that simpler data handling strategies are sufficient for maintaining predictive accuracy in the presence of descriptor uncertainty.