you finish when you have used several descriptive and predictive techniques on your data sets and you don't gain new insights. It is highly recommended to apply different preprocessing task on your attributtes (discretization, summarization, etc.) and different training and test methods

Before completing any experiments, i find useful to set goals that would describe several levels of achievements. One would be for poor, one for satisfactory and another one for good. 

A lot of time is spent getting to know the data. 

I hope it helps.

Patricia 

The simple answer will be, whenever you have answer all your initial questions and/or you have prove of disprove hypothesis.

From the theory of computing point of view the problem that you presented is undecidable. It seems that you can always get new insights. There should be a limit, however, and I think it could be characterised by measuring entropy understood as the average amount of information contained in your data (Shannon, information theory). The bigger the entropy is, the more "random" is your data is, and hence more rules can/must be discovered.

Dear Mayur, 

Your question that could be rephrase as "when to stop the analysis algorithm?" is a major part of any algorithm that deals with unsupervised unstudied data sets, a part without which the algorithm runs the risks of (a) missing the optimum, (b) wasting big time. 

Since it is imperative to have a definition of stopping indicator, I suggest the following practical EO criterion: 

1. Before starting the analysis, calculate the best results you can get by using common techniques, or just ask the users what is the best results they can get on their own. 

2. The EO criterion would be then, surpassing the existing best result by at least 1 Sigma.

EO criterion ensures an added value to your algo, and can help refine the results by installing the new found optimum as a stop indicator.

Regards,

Edith - Home of GT data mining