Automation risk

Question

Answer

Can you export job risk at an attribute level?

It's not possible to pull out the contribution of attributes (skills, knowledge, abilities, activities, content and style) to our prediction of automation risk. In machine learning we are working in a high multi-dimensional space: meaning all relationships between each data point are calculated within the model to identify combinations of all attributes to produce a prediction.

In contrast, linear regression is highly explainable, as you can see the direct impact of each attribute within the linear algorithm (measured as coefficients). These coefficients describe both the magnitude and direction of each attribute and therefore, could directly relate a skill to impact.

Machine learning algorithms are far more accurate compared to linear regression, but less interpretable. This is because they create a signal from all combinations of features. At Faethm, we apply a Support Vector Machine (SVM), which is one of the most accurate machine learning algorithms available, but the least interpretable - i.e. it is only possible to provide a direct link of a skill to impact on an aggregate basis. Prediction tests show a success rate of 95% for our SVM.

A level of trust is required when using machine learning algorithms - that trust is gained from rigorous testing of the algorithm over test sets. At Faethm, we evaluate the suite of our models through using a combination of diverse, robust methodologies. Our methodology includes 1) isolation of model components – to test validity of individual parts; 2) ensemble of alternative methods – to explore uncertainty from internal variability; 3) observation based – comparison to real outcomes; and finally, 4) by scientific peer review.