Model selection: where to start? For a general modeling problem, there are literally at least a dozen choices of statistical and algorithmic models to choose from. Off the top of my head, choices could be: regression (and its variants), neural networks (and its variants), SVM (and its variants), random forests (and its variants), etc. 
So what I'm asking about isn't much about what is referred to 'model selection'. I'm looking for a guide that can help me (intelligently) choose candidate models. Or is this process more of an art? I'm not looking for models specialized for certain data types. I'm looking for models that generically model tabular data.
I tried consulting with Google but it wasn't of much help. Surely, there must be a guide on how to begin the model selection process.
 A: Start with your substantive non-statistical research question.
Be comfortable with what question(s) you are trying to answer, and what the form of a satisfactory answer would look like. Then go looking for models.
If you commit to an analytic model (statistical or otherwise) and then fit your research question and study design into that, you are constraining a priori the form of answer that your analysis will produce in ways that are deeply ontological both in terms of expectation (what the world looks like) and ideological (how you value it). To give a simple and cartoonish example: committing to 'flat' (non-hierarchical/non-multilevel) regression models biases your analysis away from considering group context. A raft of logical fallacies might follow, and certainly ideological positions (such as Margaret Thatcher's oft-quoted direction of where to gaze "they are casting their problems on society and who is society? There is no such thing! There are individual men and women and there are families and no government can do anything except through people and people look to themselves first") could be reified in the seeming of a mathematical neutrality.
In the sciences, all our models are false, being convenient fictions that employ until we run into the boundaries of their convenience. To paraphrase Richard Levins, we will find the truth at the intersection of independent lies. So triangulate with (radically) different forms of model building: game theoretic models versus regression models versus differential equation models versus loop analysis models etc.
See this chapter for some deliciousness: Awerbuch, T., Kiszewski, A. E., and Levins, R. (2002). Environmental change, climate and health: issues and research methods, chapter Surprise, nonlinearity and complex behavior, pages 197–219. Cambridge University Press, New York, NY.
Also, Levins, R. (1966). The strategy of model building in population biology. American Scientist, 54(4):421–431 is worth a read, as is the series of critical responses to it.
