We need more answers, and therefore answer-writers
There's no escaping that more answers to on-topic questions are required to close the gap.
More active participation in closure queues is a partial solution to our increasing volume of questions. However, a particularly distressing trend in this plot is that the number of answered questions per month is flat. Even if we more aggressively moderate poor questions and close duplicate questions, the plot in my post shows that the number of questions asked per month is linear in time, but the number of answers is basically the same. Moreover, even if we can identify more duplicates and close marginal and off-topic questions, that does nothing to address the fact that solid, on-topic questions often lack answers.
One of the major take-aways from Are we seeing a dramatic drop in answers per question? is that we need more "journeyman statistician" participation. I think this is a great idea, but I don't know how we get there!
At the broadest level, Stack-exchange is undertaking to make the entire suite of websites a more accessible and friendlier place (discussed here We'd like your feedback on our new Code of Conduct!).
Making SE more friendly seems like a good, long-term way to accumulate new users to answer questions to all sites. But is there anything that we, as statisticians and machine learners specifically, can do to gather more of our kind to answer questions?
If someone answers a question in a comment, copy the comment in to an answer
I'm as guilty of leaving answers-as-comments as the next person, but if a question is fully and completely answered by a comment, you might as well leave that same text as an answer.
Moreover, if you see this happen, you can just copy-paste the comment to an answer, with attribution to the original commenter. I have a boilerplate explanation that I append whenever I do this.
I've copied @____'s comment as an answer because the comment is, more or less, an answer to this question. We have a dramatic gap between answers and questions. At least part of the problem is that some questions are answered in comments: if comments which answered the question were answers instead, we would have fewer unanswered questions.
If you don't particularly care about the rep, you could make your answer community wiki.
Don't let the perfect be the enemy of the good
Sometimes we get a question which is completely answered with two or three trite sentences. "How do I know whether weight decay improves my neural network?" is fully and completely answered by pointing out that the question is an empirical one, and a good way to gather empirical information is by conducting an experiment. I'd like to emphasize that the question "How do I know whether weight decay improves my neural network?" is wholly on-topic, (it's about a core topic in neural networks!), so closing it is not the right solution. Instead of closing the question, write the (short, clear, obvious) answer.
Another question asked how to label images for a neural network. The poster had taken photographs of objects of interest and had seen code examples of data loading functions which yield image data and image labels. The user didn't know that, at some point in time, a human had undertaken the tedious effort of telling a computer "this image is a cat" and "this image is a car."
Both of these are on-topic questions with obvious answers. Not all answers need to be, or can be, multi-paragraph efforts with figures, citations, graphs and equations.
If you can completely answer a question with a few clear sentences, do it!
More aggressively close overly broad, unclear and marginal questions
I think that a large portion of the problem is that there are quite a few poor questions, or marginally useful questions, or questions for which there is no known answer, or questions which ask the answerer to outline an entire statistical analysis project. We could do more to aggressively close these kinds of questions.
I specifically want to call out "marginal" as a genre of question. This is a question that asks a question that is, purely in the abstract, answerable. The dullest version of this question is "what's the best model for my data?" followed by a long explanation of how a database is configured, or how a website tracks user behavior. There surely is an answer -- but no one here is equipped to provide that answer without carrying out the actual analysis. This genre of question is the machine learning cousin to "What test should I run?"
Moreover, the strategy of writing a single, canonical answer to address a large audience of future readers is a losing proposition in the face of a rising tide of question for which the interest is entirely parochial. We need to close marginal questions.
As an aside, if anyone's looking for a quick way to get a lot of rep, we could use a canonical answer about back-propagation. I'm willing to award a princely bounty for such an answer (if I don't write one first). (Or, if one exists, we need to use that existing answer as a canonical duplicate target.)
(This seems to be a good place to note that I think we do a good job of directing questions which are purely about programming, math or obtaining data to their respective host sites, so good job team!)
We have a "moderator blind spot" on modern machine learning; we need a specialist
I don't want this to sound like criticism or disparagement. We have an exceptionally knowledgable and experienced group of moderators. However, I think there is a "blind spot" among them with respect to modern machine learning topics such as machine-learning, xgboost and neural-networks. I've adopted these tags as pet projects of mine, but I do notice that VTC efforts often get stuck at 4 votes, rather than the required 5.
I think that stats.SE would benefit from adding at least one moderator who specializes in these topics. (But machine learning is a large field! We might need more than one!)
Likewise, duplicate questions are a problem. It can be hard to passively identify and close duplicate questions without a near-omniscient knowledge of the thousands of question on CV. I often find that I can recall an answer to a question exists somewhere, but struggle to find the relevant thread.
To solve this, I've developed a strategy which I (lamely) call "duplicate patrol." It starts by looking at questions bearing some tag in which I am interested & feel that I am competent, and looking to identify "veins" of unanswered questions in need of an answer. If an answered version of that question can't be found, answering one of these question gives dupe targets for the others. When composing an answer, the goal is to identify common "themes" of questions, so as to address an issue which is of broad interest (and close several threads at once).