I have been on this stackexchange long enough now to see there is quite a lot of questions along the lines of:

Help, I'm a data science intern/amateur statistician and I've been given a load of data from my boss/professor and told to look for insights/trends, and I've spent half a day browsing wikipedia/stackexchange and now i'm really confused. What do I do?"

General features of these questions:

  • The asker is a first time user with 1 rep point.
  • The asker has been tasked with this by someone high up who has noticed the buzz surrounding Machine Learning and thinks this is definitely something their company ought to be doing, but doesn't know how to go about it.
  • The asker is not a professional statistician and previously had no idea how many different techniques there were.

As "Machine learning" as a corporate buzzword continues to gain widespread use and as neural networks and other AI techniques continue to produce miraculous and highly visible results for big tech companies, we will continue to see an ever-increasing influx of these questions, no matter how punitively we downvote or close them.

As such, I think there should be at least one example of a well-worded question of this nature that has a good answer that we can point to, involving some general rules of thumb about how to do data analysis - stuff about where to start, checking the data is clean, graphing it to get a sense of what's going on, a couple of basic pitfalls, etc.

After that, a series of links on specific topics, or where to go to get help. The trickiest part to answer is the dreaded "what projects should I try with this data?". This part can probably be answered only anecdotally, with specific examples/links of worthwhile DS undertakings at other companies, along with general platitudes about trying to understand the company's needs and difficulties.

Downvoting and closing as "too broad" without giving any useful feedback to the asker will not curtail the influx of these sorts of questions. It merely discourages the new user from asking questions here in future.

There is a specific question I have in mind that I want to have a go at answering, but I can't because it is put on hold. I am happy to have a look for a good "What do I do?" question and develop a generic answer that will hopefully be useful for future searchers.

My general point is, these questions will be asked whether we want them to or not, so we ought to at least make some minimally helpful generic answer that we can use to point them in the right direction.

  • 5
    These questions will continue to be asked if you do this (that doesn't make it a bad suggestion). They will continue no matter what. – gung Jul 12 at 6:32
  • 1
    That's my point. There's too many of these sorts of questions to attempt to tease out some form of an answerable question and then provide a detailed answer. Having one good,detailed answer that helps them at least gain a foothold, I think, would be beneficial to the asker. – Ingolifs Jul 12 at 6:51
  • 4
    See stats.meta.stackexchange.com/questions/3119/… and stats.meta.stackexchange.com/questions/3175/… for similar discussions. Moreover, please notice that putting questions on-hold is not meant to stop them, but rather to encourage OP's to edit & improve. If we wanted to get rid of such questions, we would simply delete them (yes, we can). – Tim Jul 12 at 7:01
  • 6
    Upvoted, because this is an important question, even if I disagree with the proposal. – Stephan Kolassa Jul 12 at 7:30
  • 1
    @ Tim I suspect that regardless of the actual purpose of putting questions on hold, new users will only ever take the closure as a rebuke. "Go away, you're not smart enough to post questions Here". They simply move on and ask the question somewhere else. It would be interesting seeing the proportion of questions that get edited by the OP after being put on hold - I suspect it is quite low. – Ingolifs Jul 12 at 8:31
  • 13
    I'll be heartless here. The "on hold as unclear what you're asking" message seems clear enough to me, and it contains links to the help center. If a poster interprets this as "Go away, you're not smart enough to post questions Here", I am somewhat at a loss what to do. In addition, I doubt that someone prone to this kind of misunderstanding would be, ahem, mentally equipped to profit from a "general" Q&A of the kind you are proposing. – Stephan Kolassa Jul 12 at 8:57
  • 2
    I just pasted by the question that you were referring to, and I commented somewhat myself, but as you noted yourself ("You'll have to narrow your question down substantially to get it into an answerable state"). Yes, general advice is possible but should this really be part of a q&a that is dedicated to statistics? I think yes-maybe, but it is a gray area. In my comment to Stephan's answer I mentioned a canonical case which I believe is more useful. The question that you mention should I believe be deleted if it does not change. It is only meta-discussion (and broad) about data-science. – Martijn Weterings Jul 12 at 10:06
  • 1
    We got such questions long before machine learning started gaining popularity (I was regularly answering questions just like it in the mid 80s in person or via email, and by the early 90s was regularly answering them on the internet - before I started using a browser. I recently encountered a lengthy answer I wrote in 1992). It's not a phenomenon specifically related to machine learning though it's certainly helping to generate more of this type of question. – Glen_b Jul 12 at 23:19
  • 1
    Following up on @Glen_b's comment: this is indeed not confined to ML and people having to "learn something from your data, just do ML!" The other prevalent flavor is people doing a study, collecting lots of data, and only afterwards thinking about how to analyze it. (I'm baffled by thinking that collecting data without an idea of how to analyze it is a good idea.) We used to get more of that flavor; they still come, but are crowded out by the "ML flavor". – Stephan Kolassa Jul 17 at 8:11
  • 3
    @StephanKolassa Here is an example of how I just handled a probably "too broad" question: stats.stackexchange.com/a/358441/130869. I think it is helpful enough, given how broad of a topic they are asking about, and it took me about 5 minutes to write. – Mark White Jul 22 at 17:24

I'm sorry, but I don't think this is a useful approach.

The questions you are referring to are essentially "please teach me data science". There are just far too many possible flavors. If you set up a general question and answer that future questions of this type can be directed to, this "generic" Q&A will be either a full book-length treatment, or so vague as to be meaningless:

  • "Plot your data" - yes, but what should I plot if I have a credit scoring task? Or am trying to forecast a million time series? Or want to cluster stuff in 37 dimensions?
  • "Clean your data" - how do I clean a time series? Categorical data with two levels? Categorical data with 42 levels? Should I remove? Impute?
  • "Distinguish between statistical significance, importance and predictive power" - what is this "statistical significance" thing I keep hearing about? Should I do stepwise variable selection? What quality KPI should I be using for my classification task? For my forecasts, should I use or or ?

There is a reason why Chemistry.SE does not have a "generic" Q&A on "How do I do a chemical analysis?"

And everything in-between "a full 1000-page tome" and "so vague to be meaningless" has been covered about 1e9 times in MOOCs, tutorials, blog post series and, yes, textbooks that are somewhat shorter than 1000 pages.

So, sorry, I will continue to VTC such questions as "too broad".

  • 7
    So this excellent answer to a broad question is the sort of thing I had in mind: stats.stackexchange.com/questions/7815/… It's not possible to impart 3+ years of experience into an answer, but it is possible to craft a decent overview with multiple links to more detailed answers/articles, and have that be an acceptable springboard from the 'unknown unknowns' where they are currently, to the 'known unknowns'. – Ingolifs Jul 12 at 8:57
  • 1
    Also, to be clear, I'm not advocating for not closing such questions. However, they shouldn't be closed without any other form of feedback or a link to an answer for a similar question. If such answers already exist, I think we should be linking the OP to them. – Ingolifs Jul 12 at 9:01
  • 3
    It might be interesting if we could slice of a bit of the broad or basic but big questions by referring them to some canonical Q&A that relates to the topic. I agree that this can not be too broad and it might be difficult to find good topics for which this can be done, but I believe that the one by stats.meta.stackexchange.com/questions/5273/… Sycorax is a good example. – Martijn Weterings Jul 12 at 9:37

The endless variations on "How do I analyze these data?" are not answerable questions. Full stop.

When I first started my career, I worked as a consultant specializing in data science. I didn't know much about math and machine learning, but I learned. (I was a "unicorn" that had some statistics background and some subject matter knowledge which made me qualified for a specialized role.) Even though I was picking up programming and math skills which were critical to my job, the absolutely most important lesson that I ever learned on that job was to ask this question, early and often:

What problem are we trying to solve?

The answer to that question will provide immediate and obvious direction. The answer to that question is also why it's impossible to produce a catch-all thread to answer "How do I analyze these data?"-type questions. "How do I analyze these data?" is only a question in the strictly grammatical sense. It's not a statistical question, it's not a data science question, it's not a machine learning question. It's a category error masquerading as a technical issue.

But note also that "What problem are you trying to solve?" also isn't a statistics question. It's a business question, for a broad definition of business. We can't answer these questions, either. But consultants can!

  • 5
    +1. Can I bounty Meta questions? – Stephan Kolassa Jul 12 at 20:57
  • 4
    @StephanKolassa Not to my knowledge, but I do accept fine distilled spirits as a gratuity. – Sycorax Jul 12 at 20:59
  • 3
    I have... connections... to some not-too-bad vineyards. We need to set up a secret handshake in case we ever meet. – Stephan Kolassa Jul 12 at 21:00
  • 7
    Ya'll will recognise eachother due to your "what would whuber do?" t-shirts. – Matthew Drury Jul 13 at 4:11
  • 1
    Judging from the profile picture @Sycorax would be easy to spot at any statistics or machine learning meeting. Mind you, I've never been to a machine learning meeting so I might not be making enough allowance for robots or cyborgs in fancy dress. – Nick Cox Jul 13 at 7:44
  • @NickCox Just look for the pine grove imprisoning users asking off-topic questions. – Sycorax Jul 13 at 14:51

Here is another possible approach.

I still stand by my opinion that any answer to "I have a dataset, please help me analyze it" will either be so vague as to be meaningless, or require a book-length treatment.

Happily, this does not need to be the end of the story. Because such book-length treatments exist! For instance, our How best to use the review queue? thread contains boilerplate that I added as something that should be helpful for many questions:

This question is very broad, and I believe you would profit
from reading an introductory level textbook, e.g., the free online
[*Forecasting: Principles and Practice* by
Hyndman & Athanasopoulos](https://otexts.org/fpp2/).
If after reading this you still have more specific questions, then
please do ask them here. If you already *have* read such a textbook,
please edit your question to make it more specific. Thank you!

So: if anyone knows of a good textbook that would be helpful for this kind of question, we could direct people there. Actually, we could use the existing Free statistical textbooks thread to collect (more) books.

And then close questions like this one not necessarily as "too broad", but as duplicates of Free statistical textbooks. Or close them as "too broad" and insert a link to that thread, e.g.:

This question is very broad, and I believe you would profit from 
reading an introductory level textbook. [We have a helpful list of
free statistical
textbooks.](https://stats.stackexchange.com/q/170/1352) If
afterwards you still have more specific questions, then please do ask
them here. If you already *have* read such a textbook, please edit 
your question to make it more specific. Thank you!

When the question is about neural networks, we can supply a more specific set of references.

This question is very broad, and I believe you would profit from 
reading an introductory level textbook. [We have a helpful list of 
textbooks and courses about neural networks.] 
references-textbooks-online-courses-for-beginners) If
afterwards you still have more specific questions, then please do ask
them here. If you already *have* read such a textbook, please edit 
your question to make it more specific. Thank you!
  • 2
    Yes, this was the sort of thing I was thinking of. Links to introductory resources and 'So you're a junior data scientist and don't know what data science is' - type blog posts and links to sets of good answers to broad-ish questions on CV. I'll admit, my push for this is because I was in this exact position a year ago - transitioning from a chemistry PhD to data science in a tech environment. I had only a vague idea of what a data scientist does, and actually thought one of the main functions of a statistician was fitting distributions to things. – Ingolifs Jul 17 at 8:54

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .