1
$\begingroup$

Do people in Machine Learning (or, more generally, in Data Mining) realise that no causal link can be inferred from a correlation?

My impression of the ML/AI community is that most people involved have a Computer Science background. I think it uncontroversial to point out that this is not a Science in the academic sense (think Physics, Chemistry, Psychology). Therefore, most of these people have no formal training in the Scientific Method, in Statistics or even in Mathematics.

"Correlation does not equal causation" is lesson nr. 1 in Statistics. In order to establish causal links, experiments are required (under some very special circumstances, other techniques are available). To me, this seems a fundamental and disastrous limitation to Machine Learning. ML will do fine with describing data (and classifying it), but it will never substitute an experiment. The application of ML is, therefore, limited in scope.

My question is about the atmosphere within the ML community.

  • Do people within the field recognise this limitation? Or do people think the causal-inference problem is not fundamental and will be solved in the future by better CPU's and better algorithms?
  • Do they restrict the use of ML to descriptive analyses? Or do they, erroneously, assume ML can help us understand and influence the world?
  • Is there a lively debate on this topic within the community? And are there articles/blogs related to these issues?
$\endgroup$
6
  • $\begingroup$ Real applications show that influence of ML methods is evident. Further, dimensionality reduction methods help to visualize huge data, therefore helps to understand the world. Many people use ML methods which do not have proper education about statistics, true. Insufficient education yet is not a problem restricted to the ML community - which anyhow is not a homogenous community at all. Therefore ANY statements about ML community knowledge is a long shot $\endgroup$ Commented Jan 9, 2017 at 14:08
  • $\begingroup$ Most Psychologists are educated in Psychology, most Biologists are educated in Biology, yet most ML'ers are not educated in Statistics. Is my impression wrong? I want to be proven wrong. $\endgroup$
    – LBogaardt
    Commented Jan 9, 2017 at 14:42
  • $\begingroup$ ML and statistics are not equal. Further, inference is not the main target of ML methods. In my university, most people applying ML methods are either computer scientists or domain experts (for example: engineers). Most people researching ML methods have a background in mathematics or physics. Therefore the ML community is rather heterogeneous regarding their statistics education. $\endgroup$ Commented Jan 9, 2017 at 15:05
  • 1
    $\begingroup$ Machine learning tends not to care much about causality, but only to solve specific problems in the form of classification, regression, etc. On the other hand, psycology, biology, etc are fields where it is frequent for scientists to do p-value hacking to suggest cause-effect relationships. Also, computer scientist do have formal training on statistics and maths. $\endgroup$
    – noe
    Commented Jan 9, 2017 at 15:20
  • 2
    $\begingroup$ Good computer scientist are thoroughly educated in information theory and statistics. In fact, if it is true that there are many machine learners that do not grasp these concepts, that would be quite worrisome for the domain as such. Installing R makes you as much of a data scientist as picking up a scalpel makes you a surgeon. I personally do not encounter professionals that elevate correlation to cause, I do however find many people that reject correlations that include no casualties. That would be just a foolish. $\endgroup$ Commented Jan 9, 2017 at 16:11

1 Answer 1

6
$\begingroup$

I think you're a bit confused about the role machine learning plays in the order of things, and I find your impression of the ML community to be a bit bizarre.

To begin, ML experts are often not the people who use ML algorithms to solve science problems - those people are usually called "data scientists" these days, and in principle they are supposed to have some sort of background in science, statistics, and/or mathematics. And yes, (good) data scientists spend a lot of time worrying about causal inference - they have to because there are big companies with lots of money riding on their predictions. Contrary to your impression ML experts are usually aware of these issues too, but their job is to build and analyze modeling tools rather than actually apply them to real data. Of course there are lots of ML and data science people which straddle the line between the two fields, but my point is that your question is a bit like asking why mathematicians don't worry about experimental error.

That said, many (most?) ML algorithms are more oriented towards classification problems which are not quite as vulnerable to your concerns. You still have to worry about sticky statistical problems like overfitting and building good training / validation datasets, but the focus is more on understanding the structure of the data that you have than predicting the future.

$\endgroup$

Not the answer you're looking for? Browse other questions tagged or ask your own question.