Feedback | SoME

A Hitchhiker's Guide to Data Science

Audience:

This video describes data science as being concerned with using a sample of data to infer something of interest in one of the 2 unseen and theoretical universes out there - the causal universe and the dependence universe. It walks towards making the point that insufficient data is at some level the only thing blocking our view of the dependence universe, and it builds the sharp contrast with the causal universe in which there are far more formidable things blocking our view of the causal universe. In so doing, I try to intuitively introduce the concept of an equivalence class of causal models, without ever introducing the term. The overarching goal of the video is to give students who are beginning their journey into data science a big picture description of what data science is about. One particular thing I am trying to contribute to is informing the latest crops of machine learning students about the fact that there is another entire universe out there besides the prediction/dependence universe. The typical machine learning graduate of today tends to be uninformed of the causal universe and of how very different causal inference and prediction tasks are. When a freshly minted ML student critiques a linear regression (with the relevant variables appropriately transformed) intended to estimate (an otherwise properly) identified causal effect for having an R^2 that is too low, chances are that he/she does not understand the difference between prediction and causal inference. Even ML teachers/pioneers are at times confused about causal inference. Take the classic ISL with R or ISL with Python books as an example. The fact that the following paragraph has survived several rewrites suggests that it is a not due to a typo but is a reflection of a misunderstanding on the part of the authors of what causal inference is and how it is different than prediction "... Therefore, if we determine that there is an association between advertising and sales, then we can instruct our client to adjust advertising budgets, thereby indirectly increasing sales. In other words, our goal is to develop an accurate model that can be used to predict sales on the basis of the three media budgets" (page 15 in ISL with Python and ISL with R). Every machine learning student needs to spend some time to understand causal inference and how it differs from prediction, because causal questions will not cease to be of interest to us as a society; ever. And the machine learning community should not use "linear regression" as a punching bag example only to dismiss it as the ever-useless model, in spite of it being often too simplistic for prediction, because social scientists still use some form of linear regression in most top journal published articles exploring a causal (as opposed to a "prediction") question.

Analytics

5.07 Overall score*

115 Rank

14 Votes

8 Comments

Comments

5.8

The takeaway is made very clear by introducing those two universes (causal/dependence).

I think the video could be slightly more focused.

The data wants to present a introduction to data science and causal relation, and it might be better to focus on subsections here on introducing causal inferencing more and the main point which is that causal inferencing is worth studying and not easy for other reasons. Causal graphs aren’t also very much introduced very fully here, which could cause some confusion on newcomers to the area and it might help to have a concrete real lift example to tie this down.

Probability part is a bit confusing to me, didnt quite follow well

7.5

Excellent video! Great intro for people getting into the field, and nice refresher for those already into it! I feel like bits of the first third of the video drag on, but the pacing getting into the second half feels much tighter. I think I wish that some of the expression was a bit more dynamic, but otherwise the content as a whole is genuinely quite excellent.

The discussion here is fairly abstract throughout - which is great for a undergrad/graduate level audience to clearly split apart the ideas of dependence and causality. However, a few explicit examples with real-life variables and models could help make the concepts clear at a very fundamental level to a wider audience.

8.5

I was surprised by how quickly the video went by and it left me wanting more. The example of how the three causal models lead to the same distribution was of course telegraphed, but satisfying. Some of the visualisations could be developed graphically and made more dynamic for clarity. I think the pacing could be faster (I watched at 1.5x speed and that solved it for me), but this is a strong entry.

I loved the dip into causal inference — it’s generally something that’s shunned as something to be completely avoided, and the pitfall example at the end was a nice visualization.

That being said, though, this probably would be better as a blog post. It was very hefty in words (good and well-phrased words, but hefty nonetheless) and generally I was looking at a static screen; I would have prefered to read at my own pace. As a video, it could have at least used some dynamic highlighting to indicate what you were focusing on.

My biggest problem withal regards your target audience: an incoming freshman interested in studying data science would understand almost none of this. This would be appreciated by graduate students or undergraduates well into their degree, but it’s definitely scaring anyone else off. (That’s not good or bad, it’s just probably not what you intended!)

Overall, I like it! The discussion on causal graphs was a little new to me and definitely something I’ll remember.

5.1

I was quite excited about the explanation of causality and correlation. But, there is no real conclusion to this video… Other than don’t confuse correlation with causality