Knowledge graph embedding for experimental uncertainty estimation

Purpose: Experiments are the backbone of the development process of data-driven predictive models for scientific applications. The quality of the experiments directly impacts the model performance. Uncertainty inherently affects experiment measurements and is often missing in the available data sets due to its estimation cost. For similar reasons, experiments are very few compared to other data sources. Discarding experiments based on the missing uncertainty values would preclude the development of predictive models. Data profiling techniques are fundamental to assess data quality, but some data quality dimensions are challenging to evaluate without knowing the uncertainty. In this context, this paper aims to predict the missing uncertainty of the experiments. Design/methodology/approach: This work presents a methodology to forecast the experiments’ missing uncertainty, given a data set and its ontological description. The approach is based on knowledge graph embeddings and leverages the task of link prediction over a knowledge graph representation of the experiments database. The validity of the methodology is first tested in multiple conditions using synthetic data and then applied to a large data set of experiments in the chemical kinetic domain as a case study. Findings: The analysis results of different test case scenarios suggest that knowledge graph embedding can be used to predict the missing uncertainty of the experiments when there is a hidden relationship between the experiment metadata and the uncertainty values. The link prediction task is also resilient to random noise in the relationship. The knowledge graph embedding outperforms the baseline results if the uncertainty depends upon multiple metadata. Originality/value: The employment of knowledge graph embedding to predict the missing experimental uncertainty is a novel alternative to the current and more costly techniques in the literature. Such contribution permits a better data quality profiling of scientific repositories and improves the development process of data-driven models based on scientific experiments.