only data I collect is characater match-ups on specific levels with specific players. IE i have data on player1 vs player2 using char1 and char2, respectively, the map and the winner. I can extrapolate actual character relative ranks from that sample data alone (if its isn't skewed and there is enough). To set up a ranking system with players would require a ranking system of characters seeing as (if true) certain match-ups favor the user. Also. I believe I mentioned it before but for each instance (player1 wins, player2 losses, char1, char2, map) the better players stat is inflated (this is what you mean by having equitably ranked players must be used) but then the same character match-up with players using opposite characters (player1 wins, player2 losses, char2, char1, map) would be equally unweighted. Perfect example would be boom or Isai. If Isai's record is 10-0 against you when he is link and you are pikachu, then it is also likely that your record with link against his pik with also be a no loss ratio.
So say I win x10 time as much as the other player. Then a character vs character ratio of 20-1 (20) for my mario against his luigi, and a character vs character ratio of 2-20 (.1) would show that even though it is apparent that I am better than my opponent (the fact my mario-luigi ratio is higher than the inverse of his mario-luigi (ratio ^ -1) implies that even though neither total distinguishes either, the combination of the oppositely correlated ratios evens out the difference in skill levels.
Also, in my quest for usable data there are some standards that would eliminate those risks. Any useable ratios for wins and losses that are to be applied to a character as a whole must contain a minimum number of games played (i'm thinking 10 min) and have at least 1 win and loss. This is extremely important when using ratios to eliminathe the risk of having to integrated undefeated characters. Another example, if Isai is part of the data and never loses, I can't use his character statistics because I have no data that differentiates his characters and thus they provide no useful insight in the character rankings overall. Like you said, the players need to be at least reasonable close in skill (such that in any match-up there is enough probably of either player winning that the event has actually occurred).