I feel that a properly compensated FR graph can confirm some subjective findings for a person, or persons. In otherwords, I feel that there is correlation between the frequency response and subjective findings. I know Tyll does this, others as well, they'll point to a certain part of the frequency response, square wave, etc. and explain why that shows what they are hearing. Going the other way, reading a graph, is a much harder feat however. At this point in time, we understand enough to go one direction, from our subjective findings to the graph (properly compensated) to confirm what we hear. Going the other direction, graph to subjective (blind, we don't know the subjective up front), isn't entirely possible as we don't understand this direction as much.
All my use of FR graphs is relative. I'll take two FR graphs from the same measurement setup, one of a headphone I know, and one of the headphone I'm curious about, and use the knowledge I have about the known headphone to discern useful things about the new one. To me, compensation isn't really needed, although it helps to keep track of a baseline in any measurement setup. Tyll's data is really great for this, by the way.
The only gotcha is resonances. If a headphone has them, it might not be obvious from the FR graph, and even standard EQ won't fix them because they'll still ring over time even if you lower their amplitude. I recently came across a dynamic headphone that acutely suffered from this problem. Fortunately the high-end planar headphones typically don't have those kinds of resonances. :)