Tunable structure priors for Bayesian rule learning for knowledge integrated biomarker discovery

ARTICLE HIGHLIGHTS

Research background

Biomedicine is increasingly a data-driven science, owing largely to the explosion in data, especially from the development of high-throughput technologies. Such datasets often suffer from the problem of high-dimensionality, where a very large number of candidate variables can explain the outcome variable of interest but have few instances to support any model hypothesis. In many applications, in addition to the data itself, some domain knowledge is available that may assist in the data mining process to help learn more meaningful models. It is important to develop data mining tools to leverage this available domain knowledge. However, currently, there is a dearth of data mining methods that can incorporate this available domain knowledge.

Research motivation

Developing data mining methods that can incorporate domain knowledge will help learn more meaningful models and will benefit many domains, especially the ones that suffer from data scarcity but have some domain knowledge that can assist with the data mining process (for example - biomedicine).

Research objectives

In this work, our objective was to extend a rule learning algorithm, called Bayesian rule learning (BRL), to make it capable of incorporating prior domain knowledge. BRL is a good candidate because it has been shown to be successful in application to high-dimensional biomedical data analysis tasks. We implemented such a tool, called BRL_p that has tunable priors, which means the user can control the degree of incorporation of their specified knowledge. BRL_p is a novel data mining tool that allows the user to specify their domain knowledge (including uncertain domain knowledge) and incorporates it into the model search process.

Research methods

BRL searches over a space of Bayesian belief network models (BNs) to find the optimal network and infers a rule set from that model. We implemented a way for the BN to incorporate informative priors, a distribution encoding the relative importance of each model prior to seeing the training data. This allowed BRL to incorporate user-specified domain knowledge into the data mining process called BRL_p. BRL_p has a hyperparameter λ that allows the user to adjust the degree of incorporation of their specified prior knowledge.

We evaluated BRL_p by comparing it to BRL (without informative priors) and other state-of-the-art classifiers on a simple simulated dataset, and a real-world lung cancer prognostic dataset. We measured the degree of acceptance of the specified prior knowledge with respect to the hyperparameter λ in BRL_p. We also observed the changes in predictive power using AUC.

Research results

We observed, in both the experiments with simulated data and the real-world lung cancer prognostic data that with increasing values of λ the degree of incorporation of the specified prior knowledge also increased. We also observed that specifying prior knowledge relevant to the problem dataset could sometimes help find models with better predictive performance. When BRL_p is compared to the state-of-the-art classifiers, we observed that it performed better than other interpretable models but the more complex and non-interpretable models achieved better predictive performance than BRL_p.

Research conclusions

BRL_p allows the user to incorporate their specified domain knowledge into the data mining task and allows them to control the degree of incorporation with a hyperparameter. This is a novel rule learning algorithm that we have made available to the general public via GitHub. We anticipate its use in many applications especially the ones suffering from data scarcity but have additional domain knowledge available that may assist in the data mining task.

Research perspectives

In this paper, we explored specifications of simple domain knowledge. We need to further explore the incorporation of more complex forms of knowledge. In this paper, we incorporate domain knowledge from literature. We also want to explore domain knowledge available in other sources. These future directions may motivate further developments to BRL_p.