This is a response to the question, “Is performance review data too biased to use in machine learning?” posed by Kate Bischoff
The very definition of culture is that it’s biased. Cultures are defined by their customs, mores. boundaries and histories. The bias (lederhosen are beautiful, for example) makes one culture different from another.
That bias may be the most predictable thing about an organization. Like water to a fish, cultural bias is extremely difficult to recognize. It is easiest to see when you try to change it. The change itself feels like the problem.
If you are not trying to change the culture, reinforcing its biases (strengthening) is a smart thing. You can use algorithms to reinforce and amplify existing bias. That’s more or less what is meant by ‘it learns as you use it’ or ‘it gets better with age.’ Algorithms and machine learning isolate and purify the things that are unique to a culture.
But, if you are trying to ‘improve’ the culture (and reduce its biases), it’s a different story. It is much more difficult to apply machine learning to a situation in which you want different results than you are currently getting. There’s probably a name for learning software that rehabilitates a culture. I haven’t run across it.
All data contains bias. It is created by a culture for a purpose. You can see the bias by noticing what isn’t measured or managed. All data is biased.
The data from performance reviews are highly biased (towards the existing status quo). Whether or not you choose to use it depends on what you are trying to accomplish. If you want to reinforce the current state of things, use data from the current state of things. Your AI/ML/BigData/Predictive Analytics tools will always give you more-better results.
If you do not want to continue to perpetuate the bias. use external data. While it will contain the bias of the open market, it won’t perpetuate your existing biases necessarily. If you want to move your culture in a specific direction, use data that carry that point of view.
I’m imagining a business that labels the quality, toxicity and probable effects of data by source.
