Using Knowledge Graph to improve Causal Network learning
Inferring a causal network solely from observational data is a challenging task. For a network with n observables (nodes), the number of possible causal graphs grows super-exponentially, with concomitantly growing requirements on the dataset size for accurate network inference. Both due to vastness of the space of possible networks and the degeneracy of Markov-equivalent network structures, using standard inference procedures and a uniform prior on the network structure leads to measurement bias that causes erroneous relationships or relationship orientations.
For every domain, some kind of prior knowledge is always known to the experts which are usually obtained via extensive literature survey and text mining. However, these are mostly in the form of raw disparate data in various standards and formats from heterogeneous sources, which is difficult to integrate into a single structure. Recent development for solving these problems led to the building of strong hyper-graphs called Knowledge Graphs (KG) to organize large-scale complex big data containing all complex biological relationships. These large scale networks accommodate structural information which can be leveraged for reasoning, recommendation, decision making etc. Combining information from these structured databases will result in predictions that are consistent with both the evidence in data as well as the true prior knowledge already known.
In this work, we exploit the potential of knowledge graphs to guide towards a better causal network structure learning. We found that our application tested on several real world data shows improvement from baseline.
Major Advisor: Stephen Ramsey
Minor Advisor: Thomas Sharpton
Committee: Prasad Tadepalli
Committee: Rebecca Hutchinson
GCR: Ren Guo
Friday, October 30 at 1:00pm to 3:00pmVirtual Event