In the introductory chapter of "Process Mining: Data Science in Action" (2016 - Van der Aalst, pag 11) the author says that :

Although data science can be seen as a continuation of statistics, the majority of statisticians did not contribute much to recent progress in data science. Most statisticians focused on theoretical results rather than real-world analysis problems. The computational aspects, which are critical for larger data sets, are typically ignored by statisticians. The focus is on generative modeling rather than prediction and dealing with practical challenges related to data quality and size.

The bold phrase is not clear to me. In fact, since a generative model is the model that generates the data, once we obtain it we can do predictions. So, to me generative modeling and prediction are not opposing concepts. What do you think ?


I think what the author is speaking about is the time/memory complexity of algorithms that statisticians may don't care about. Make a model which is mathematically well proven may be more important to statistician eyes than making approximation to render a model feasible in real life.

I encourage you to look at the complexity of frequent mathematical operations.

Often operation like Singular Value Decomposition, Matrix Inversion, Matrix transpose are used and their cost is way higher than the "upper bound scalable time complexity limit" $O(n.log(n))$ which prevent any utilization on massive datasets.

As an example, you can easily imagine than we are rapidly limited with time complexity because we are not able to increase computational power -- or wait -- more than approximately linearly with the increase of problem size. Take a very common $O(n^2)$ complexity, you can't afford to wait $1000000$ more longer -- or multiply your computational power by $1000000$ to keep same app duration -- when you multiply your dataset size by $1000$ (except if your operations are really fast...)

  • $\begingroup$ This ""upper bound scalable time complexity limit" $nln(x)$ it's interesting, i didn't know about it. Do you know some material about it ? $\endgroup$ – Koinos Jul 19 '18 at 17:30
  • 1
    $\begingroup$ I edit my answer with an example to be more expressive but i don't have some literature saying explicitly that $n.log(n)$ is a "upper bound scalable time complexity limit". It's experience and common sense. Depending on how you think you can say that every algorithm is scalable but in practice you cannot increase your computational power or wait as much as the problem size increase... $\endgroup$ – KyBe Jul 20 '18 at 8:59

This is quite theoretical. I will try hard to simplify. Lets say one egg cost 3 bucks at an egg store

To get 3 eggs you need 9 bucks in your pocket. -- But (if you are a cashier) -- Someone handing 9 bucks tells he/she want 3 eggs

Buyer need to have an engineer mind and "build" a decision of how much money is needed, while cashier need to have a scientist mind and break down the 9 bucks scenario that this person want 3 eggs.

Data sciences became more with the data availability explosions and being (the data) copy of real world (here you are understanding the real world as a scientist) while generative mind is more about engineering to build a solution for the real world.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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