Zion,
The idea would be to collect all flight data and outcomes over a long period and let the software look for potential relationships and make future recommendations based on those relationships-
The problem with incidents/accidents is that all analysis is a retrospective only of the incident/accident in question. The cause of the event may be determined to be mishandled by the crew. The associated reports can only really focus on what the crew should have done in the seconds leading up to the incident, I am much more interested in examples where incidents were averted much earlier in the chain.
So, for me the point of data mining would be to discover flight patterns (based on previous incidents or other parameters) that could have been incidents/accidents and also discover how the crew recovered. The crew may not even be aware that the situation they were in was very similar to a situation that once led to a hull loss, but on this occasion, they did something important and subtle that recovered the situation before it became event worthy.
Naturally, this would be recording all/most/some flights continuously, which is technically feasible with FOQA and SD cards right now. Clearly, the data would need to be de-identified and handled only independent agencies. There would be issues with voice recording storage, anonymity and interpretation though.
This would be a hugely positive step; we would be able to draw upon clear evidence of good practice, excellent SA and handling skills and how they related directly to problem solving. As Tony Kern puts forward in his book, being potentially much more powerful training that just focusing on what went wrong.