Do features remain informative across domains and predicted variables? For example, consider a feature like "wait_time_in_queue" is important (informative) in predicting whether a customer will return to a restaurant. 
Will the same feature be important for another predicted var like whether or not customer will return to a doctor, or movie theater, or grocery store? 
Also will the feature be informative across 2 different datasets e.g. for example across datasets for 2 different restaurants or even same restaurant across different times?
I guess my question is, if a feature is only informative for a particular dataset, how is it possible to make generalizations across domains/predicted variables?
Also what if the feature is generated using the same process across domains e.g. assume "wait_time_in_queue" is always sampled from a poisson distribution. Will it then be informative across different domains?
Is it possible to have "world knowledge" direct which features are important to which predicted variables without looking into the dataset?
 A: The "world knowledge" is actually called "domain knowledge" or "experience". It enables experts do make good predictions without using machine learning since the dawn of mankind ;).
Given that a feature has importance to predict a certain target in a certain domain, it is reasonable to expect that this particular feature will have also importance to predict comparable targets in comparable domains. However, the devil is in the detail:


*

*How has the data been collected ? Wait-Time-in-Queue seems to be pretty obvious, but it is just a name, so it could reflect waiting time in seconds, average waiting time in minutes on the last 30 days, average of the median waiting time at all stations within the restaurant, average waiting time at the last day it has been collected (which weekday ?) etc.. Another factor is measurement errors, the feature data might be reliable or too noisy to use.

*Restaurant may differ in such a way, that the relative importance of features does change. For example, drive-through at a highway without regular customers versus a restaurant at a fixed location. In the first problem the number of returning customers is really sparse, even more, if the restaurant is only visited by tourists

*How long a customer is willing to wait is also very important and may depend on the domain and circumstances of the particular instance. If the doctor is really good (or the only one within miles) and it is for example winter, then customers are willing to wait basically because they need help and going elsewhere won't help.

*Another example are amusement parks. For reasons I do not understand, a lot of people are willing to wait 1h+ at certain attractions within the park for a single ride. This problem is also sparse (1-2 visits a year for a family) and waiting time cannot be avoided, because there is only time at weekends and during the holidays. 


In summary, the feature "waiting-time" is important in all these cases, but the degree of the importance varies a lot.
So it is a matter of experience, domain knowledge and collected data. Speaking for myself, when facing a new challenge / data set, I will first try approaches which have worked most of the times in the past in comparable domains, but already look for clues which might indicate the opposite. Afterwards I will go with the approaches which have worked sometimes, carefully selecting those which are most promising. In the end, new approaches are tried and this is where new knowledge is "created". I guess this is how human experience works in general.
(Fun) exercise: The next time you are doing a (partially) new task you are not an expert in (for example repairing stuff at home, cooking, etc.), reflect why you are choosing particular actions. I think this is both interesting and refreshing ;). 
