You are right.

We can address the interpretation problem and the classification problem together.

However, we can also separate the two problems: use models to predict behaviors, and models to interpret these behaviors.

Anyway, one of the claims of "Big Data" is to increase the performances of the predictors. Then, to achieve this goal, the right way is online learning on the data stream or bagging in a Hadoop system ?

It is an important question, because the IT tools are really different. In one hand we have a Data Stream Management System and in the other hand we have an Hadoop cluster.

Although the question seems direct, the answer reveals that it is complex. It strongly depends on the scale and complexity of the problem at hand.

In physics laboratories, linear regression can still be a functional machine learning technique -- "all other factors held constant". But the more correlated, complex, and high-dimensional the problem's input features become, the less likely that simple assumptions (such as linearity) will remain valid.

If you want a catch-all 'starter' ML algorithm to try against large numbers of different problems in different spaces, then I would use random forests -- it's my goto for a baseline when comparing models for a new problem. The solution is always 'good enough' to start with, but sometimes not subtle or versatile enough. I mostly do regression analysis, where RF is "okay"; but RF's are great as a starter solution set for classification.

The real answer is the apply a broad spectrum of models within sub-spaces of the data to understand the implicit correlations between dimensions and the versatility of each ML method to tackle blocks of data or sets of features. Then increase complexity and feature space volume (n_features x n_samples) until the full solution is 'at scale'.

In a sense, you need to 'converge' to a solution over many iterations of both ML methods and feature sub-spaces that asymptotically approach the full solution.