It may actually give them negative attention; suppose I don't message anyone I rate a 4 (I don't) and by raising their rating I make others less likely to message them (because their average rating is higher). (I thought there was a way to determine another user's average rating, but I'm not seeing it from a quick check of the site, so this may not be the case.)

To the best of my knowledge, though, the coefficient for m4 and m2 aren't "relative to m3" but absolute; if someone gets 10 5s, they're expected to get 9 messages. If they got 9 4s and a 5, they're expected to get no messages. (Of course, what would be interesting is looking at clusters rather than just linearly regressing the data.)