A knowledge graph model represents a given knowledge graph as a number of vectors. These models are evaluated for several tasks, and one of them is link prediction, which consists of predicting whether new edges are plausible when the model is provided with a partial edge. Calibration is a postprocessing technique that aims to align the predictions of a model with respect to a ground truth. The idea is to make a model more reliable by reducing its confidence for incorrect predictions (overconfidence), and increasing the confidence for correct predictions that are closer to the negative threshold (underconfidence). Calibration for knowledge graph models have been previously studied for the task of triple classification, which is different than link prediction, and assuming closed-world, that is, knowledge that is missing from the graph at hand is incorrect. However, knowledge graphs operate under the open-world assumption such that it is unknown whether missing knowledge is correct or incorrect. In this thesis, we propose open-world calibration of knowledge graph models for link prediction. We rely on strategies to synthetically generate negatives that are expected to have different levels of semantic plausibility. Calibration thus consists of aligning the predictions of the model with these different semantic levels. Nonsensical negatives should be farther away from a positive than semantically plausible negatives. We analyze several scenarios in which calibration based on the sigmoid function can lead to incorrect results when considering distance-based models. We also propose the Jensen-Shannon distance to measure the divergence of the predictions before and after calibration. Our experiments exploit several pre-trained models of nine algorithms over seven datasets. Our results show that many of these pre-trained models are properly calibrated without intervention under the closed-world assumption, but it is not the case for the open-world assumption. Furthermore, Brier scores (the mean squared error before and after calibration) using the closed-world assumption are generally lower and the divergence is higher when using open-world calibration. From these results, we gather that open-world calibration is a harder task than closed-world calibration. Finally, analyzing different measurements related to link prediction accuracy, we propose a combined loss function for calibration that maintains the accuracy of the model.

Publication Date


Document Type


Student Type


Degree Name

Computer Science (MS)

Department, Program, or Center

Computer Science (GCCIS)


Carlos R. Rivero

Advisor/Committee Member

Xumin Liu

Advisor/Committee Member

Peizhao Hu


RIT – Main Campus