by Jelani Harper 28 January 2020
Machine learning typifies the statistical varieties of Artificial Intelligence: statistically significant results of machine learning models either indicate correlation or causation of factors that ideally involve an assortment of vectors.
Understanding just how this powerful element of AI works—and how to optimize its outputs for business value—requires an understanding of each of these terms and their role in any action derived from machine learning.
Those not quite clear on these concepts run the risk of struggling with explainable AI, and producing models that don’t deliver results successfully impacting business objectives. “Causation is about finding the real factors: the pathways that lead from effect A to B, which can go through effect A1, A2, A3, and A4 before you get to B,” explained Jans Aasman, CEO of graph database developer Franz.
Correlation vs. causation
Since machine learning results reveal correlation or causation between data points—for example, which customer service activities produce gainful upselling opportunities—it’s crucial to learn how to distinguish these concepts. When machine learning outcomes are statistically significant, they express some degree of correlation between the input data and the model’s results. Aasman names a good example of correlation: “It’s a known fact that the rise or decrease in temperature changes murder rates.” In this example, the warmer or colder weather doesn’t necessarily cause more or less killings, but reflects that causative agents for murders have some relationship to the weather.
Once correlation is identified, it’s then necessary to manipulate the inputs and outputs of models to find causation. “If you have a phenomenon and then you do active interference, or an experimentation or something, and then you look at the results, now you can start talking about causation because you either did, or did not do something,” Aasman explained. In the scientific community, such experiments involve using controls. In the data-centric business community, they involve tweaking the input data, or machine learning factors.
The danger of mistaking correlation for causation is that correlation does not directly explain the outcome of machine learning models. In highly regulated verticals like finance or healthcare, it’s vital to be certain of how statistical models are functioning. The key to identifying causation from correlation revolves around understanding the impact of machine learning factors. “When you have a correlation between two phenomena, what you actually want to find out is what are the intermediate factors that make the correlation go either up or down,” Aasman said.
Factors are the essence of machine learning or, perhaps, its starting point. They’re the different dimensions of the input data that affect the model’s output. When building models to reduce churn, for example, factors might include various data about a company’s prices, those of its competitors, job prospects for customers, and any other data points that could relate to customer churn. “Everything in machine learning starts with a factor,” Aasman noted. By manipulating the factors involved in model results, organizations can go from correlation to causation. “The whole point is you still use the same statistical technique to look at correlation,” he added. “But, because now you know that you have done an active interference, you can actually say if I do the interference and then something happens, then I can start talking about causation.”
Aasman defined vectors as “a series of numbers that you give as an input to machine learning.” Factors, the different types of data ingested into machine learning models, are comprised of different vectors. When training models, it’s often advantageous to not only have large amounts of data, but also differentiated data types. That way, organizations are able to encompass as many factors as possible.
Aasman referenced a health care use case in which an organization leveraged machine learning to determine whether or not a person would need to be intubated to avoid respiratory failure. “We would have to vector the weight of the person,” Aasman said, as well as other relevant information like their age, height, blood pressure, etc. These vectors are then “loaded as the input, and the output is whether or not the person is healthy,” he said. “When I talk about vectors, these are things that you get in the database that you can put in the machine learning model.”
Thus, vectors are the numerical inputs of different factors that go into machine learning models that, when producing statistically significant results, indicate correlation. Further refinement of the factors input into models can determine causation. However, there are some times when it’s advisable to act on mere correlation before clarifying causation. “If you do an experiment where you very closely track changes in temperature and how many murders are committed, if you find a high correlation and the temperature goes up, then you’d better put more police people on the street even if you don’t know what the cause is,” Aasman cautioned. Organizations with similar parallels to machine learning results and mission critical business processes should do the same, although at some point it will be necessary to identify causation to explain model results.
Jelani Harper is an editorial consultant servicing the information technology market, specializing in data-driven applications focused on semantic technologies, data governance and analytics.