Prospect certainty for data-driven models

Abstract

Abstract Uncertainty in the output of a data-driven model is a natural feature that limits its practical application. Identifying this uncertainty is required to improve the reliability of the model. In this paper, we propose a novel method to explicitly determine the certainty of the model output, by considering the input distributional changes during its deployment. In addition, a new concept of logit masking is introduced to entail the model more probabilistic characteristic which identifies the behavior of each output alternative by evaluating its influence on the distribution of the model output. Finally, it quantifies the prospect certainty of each variant and selects the final output. Experimental results using benchmark and real-world datasets show that the proposed method outperforms the state-of-the-art techniques in the sense of certainty.