I note there\'s a \"suggested\" mode in Amarok, which presumably is based off last.fm tags (I\'m not quite sure, this doesn\'t appear to work on my collection yet, since I haven\'t listened enough).

However, it seems to me this isn\'t necessarily terribly useful, and you could do a lot better by observing the user\'s own habits. Here\'s a simple scheme to glean useful information:

Play songs. When a song has played sufficiently, say 2/5 of the way through (or whatever), the user is obviously \"listening\" to it (i.e. they haven\'t skipped). Then look at the list of previous five (or however many) songs. These songs are likely to \"work well\" with the currently-playing song; the mood they\'ve set hasn\'t made the user skip over them. E.g., if I\'m in a mood for hip-hop, and some Chet Baker comes up, I\'m probably going to skip ole Chet, but if a Mos Def/Massive Attack mix comes up, I probably won\'t. So: we can create a network of all our songs, with edges weighted by how closely-connected they are by mood. We can update this based on positive observations and some Bayesian scheme I haven\'t really thought about yet. Subsequently we can use this network to more intelligently choose neighbors. How to do that is debatable - random walks on the graph is one notion, but maybe not the best.