I read a great deal of technical literature and highly recommend the same for anyone interested in developing their skill in this field. While many recent publications have proven worthwhile (2011's update of Data Mining by Witten and Frank; and 2010's Fundamentals of Predictive Text Mining, by Weiss, Indurkhya and Zhang are good examples), I confess being less than overwhelmed by many current offerings in the literature. I will not name names, but the first ten entries returned from my search of books at Amazon for "big data" left me unimpressed. While this field has enjoyed popular attention because of a colorful succession of new labels ("big data" merely being the most recent), many current books and articles remain too far down the wrong end of the flash/substance continuum.
For some time, I have been exploring older material, some from as far back as the 1920s. A good example would be a certain group of texts I've discovered, from the 1950s and 1960s- a period during which the practical value of statistical analysis had been firmly established in many fields, but long before the arrival of cheap computing machinery. At the time, techniques which maximized the statistician's productivity were ones which were most economical in their use of arithmetic calculations.
I first encountered such techniques in Introduction to Statistical Analysis, by Dixon and Massey (my edition is from 1957). Two chapters, in particular, are pertinent: "Microstatistics" and "Macrostatistics", which deal with very small and very large data sets, respectively (using the definitions of that time). One set of techniques described in this book involve the calculation of the mean and standard deviation of a set of numbers from a very small subset of their values. For instance, the mean of just 5 values- percentiles 10, 30, 50, 70 and 90- estimates the mean with 93% statistical efficiency.
How is this relevant to today's analysis? Assuming that the data is already sorted (which it often is, for tree induction techniques), extremely large data sets can be summarized with a very small number of operations (4 additions and 1 division, in this case), without appreciably degrading the quality of the estimate.
Data storage has grown faster than computer processing speed. Today, truly vast collections of data are easily acquired by even tiny organizations. Dealing with such data requires finding methods to accelerate information processing, which is exactly what authors in the 1950s and 1960s were writing about.
Readers may find a book on exactly this subject, Nonparametric and Shortcut Statistics, by Tate and Clelland (my copy is from 1959), to be of interest.
Quite a bit was written during the time period in question regarding statistical techniques which: 1. made economical use of calculation, 2. avoided many of the usual assumptions attached to more popular techniques (normal distributions, etc.) or 3. provided some resistance to outliers.
Genuinely new ideas, naturally, continue to emerge, but don't kid yourself: Much of what is standard practice today was established years ago. There's a book on my shelf, Probabilistic Models, by Springer, Herlihy, Mall and Beggs, published in 1968, which describes in detail what we would today call a naive Bayes regression model. (Indeed, Thomas Bayes' contributions were published in the 1760s.) Claude Shannon described information and entropy- today commonly used in rule- and tree-induction algorithms as well as regression regularization techniques- in the late 1940s. Logistic regression (today's tool of choice in credit scoring and many medical analyses) was initially developed by Pearl and Reed in the 1920s, and the logistic curve itself was used to forecast proportions (not probabilities) after being introduced by Verhulst in the late 1830s.
Ignore the history of our field at your peril.