Prospect certainty for data-driven models

The inherent nature of uncertainty in the inputs of data-driven models can lead to incorrect outputs. Such outcomes are difficult to ascertain due to the lack of reference data during the deployment, which hinders their acceptance in practical applications. This highlights the need to evaluate the degree of certainty of the output of a model to improve its robustness. In this paper, we present a new method for quantifying the output certainty of data-driven models, considering changing probability distributions of input data during the deployment. We achieve this by introducing the concept of logit masking to mitigate the deterministic nature of the model and build multiple alternatives for each output logit. Then, we propose a weighted probability function to provide an initial insight into the certainty of these alternatives. Moreover, we define a behavior function to describe the degree to which these alternatives affect the output distribution pattern. By combining these weights and behaviors, we determine the prospect certainty of these variants and finally choose the one with the highest certainty as the refined output of the model. Experimental results on benchmark and real-world datasets show that our proposed method outperforms state-of-the-art techniques in determining output certainty for data-driven models.