if you have no any hípothesises or research questions you should run an explorative analysis first. All the data mining packages offer broad scale of functions for that. Based on the outcomes you can focus towards a promissing direction.
thanks a lot for all of your suggestions. but the thing i really wanted is, i want to train a machine and in supervised mode by giving a data set and later it should able to classify the given input data correctly. but i have used J$* algorithm for this by giving a simple data set of marks data set which it should classify the records into grade wise basing on the data present in a criteria of what we will perform manually.
so i have decided to design a network and make to learn the things as if as i was learning the things in naturally a new thing in my life.
Data set selection completely depends on your problem statement , as in grade prediction where data is suoervised it will be small data set. Now your actual question about which data to choose will complete depend on problem you trying to solve. And if you take social media data: data will be huge and unsupervised. So it will be good to take sensor data and experiment with it first rather than big data analysis (again this completely depends on your problem).
Both are emerging topics with wide area of applications. You will have to look at the nature of the data as well, because every problem required different data. As an example if you try to do community detection you will need different type of data than that of trend detection. While doing the study, please investigate how clean the features are, as they will influence your model heavily.