No-Code, Low-Code Machine Learning Platforms Still Require People
No-code, low-code (horizontal) machine learning platforms are useful at scaling data science in an enterprise. Still, as many organizations are now finding out, there are so many ways that data science can go wrong in solving new problems. Zillow experienced billions of dollars in losses buying houses using a flawed data-driven home valuation model. Data-driven human resources technology, especially when based off facial recognition software, has been shown to bias hiring decisions against protected classes.
While automation is a great tool to have in your arsenal, you need to consider the challenges before utilizing a horizontal ML platform. These platforms need to be flexible, configurable, and monitorable to be robust and consistently add value over time. They need to allow data to be weighted flexibly in user-controlled ways and have data visualization tools to detect outliers and contributors to noise. They also need automated model parameters and data drift monitors to alert users to changes. As you can see, we haven’t evolved beyond the point where algorithms outmatch human intelligence.
So, don’t be fooled by AI/ML/low code … you still need people. Let’s take a closer look at the reasons why.
Trying to replace human data scientists, domain experts, and engineers with automation is a hit-or-miss proposition which could lead to disaster if applied to mission-critical decision-making systems. Why? Because human beings understand data in ways that automated systems still struggle with.
Humans can differentiate between data errors and just unusual data (e.g. Game/Stop/GME trading in February) and align unusual data patterns with real-world events (e.g. 9/11, COVID, financial crises, elections). We also understand the impact of calendar events such as holidays. Depending on the data used in ML algorithms and the data being predicted, the semantics of the data might be hard for automated learning algorithms to discover. Forcing them to uncover these hidden relationships isn’t necessary if they aren’t hidden to the human operator.
Aside from semantics, the trickiest part of data science is differentiating between statistically good results and useful results. It’s easy to use estimation statistics to convince yourself you have good results or that a new model gives you better results than an old model, when in fact neither model is useful in solving a real-world problem. However, even with valid statistical methodologies, there is still a component to interpreting modeling results that requires human intelligence.
When developing a model, you often run into issues about what model estimation statistics to measure: how to weight them, evaluate them over time, and decide which results are significant. Then there is the whole issue of over testing: If you test too frequently on the same data set, you eventually “learn” your test data, making your test results overly optimistic.
