New Library Adds Causality to ML Models
A new open source library is designed to help data scientists and domain experts jointly develop machine learning models based on causal relationships rather than just data correlations. The developers of the new CausalNex library argue that running machine learning projects without considering causality can lead to faulty conclusions.
QuantumBlack, a data analytics unit of McKinsey & Co., said CausalNex is its second open source release after Kedro, a library aimed at production ML code. Its new machine learning project is designed to help data scientists infer causal relationships in data. Dependencies in data can be expressed in a network graph that can then be inspected by domain experts. The collaborative approach is designed to eliminate spurious correlations in ML models, QuantumBlack said.
The library for causal reasoning applies “what-if” analysis to Bayesian networks on the assumption that the probabilistic model is more intuitive in describing causality than traditional ML frameworks based on correlation analysis and pattern recognition.
“CausalNex is built on our collective experience to leverage Bayesian networks to identify causal relationships in data so that we can develop the right interventions from analytics,” the developer said.
Bayesian networks have been used previously to build causal models, but the process often requires multiple libraries used to test how different data points relate to or influence each other. The requirement for multiple libraries often hindered subject objects who could otherwise spot relationships among variables with relative ease.
A single CausalNex library allows data scientists to develop network graphs that can be amended by subject experts. That collaboration builds trust in the resulting models. “Causal relationships are more accurate if we can easily encode or augment domain expertise in the graph model,” developers noted.
The company began work on CausalNex in 2018 as clients’ machine learning projects ran into causality issues. In February 2019, researchers combined data gleaned from internal projects into the software that underlies the new library.
Prior to its release to the open source community, QuantumBlack developers refined the library for implementation of Bayesian networks.
CausalNex is available now on GitHub.