Don't have any interesting references, sorry. My reasoning is mainly one of simplicity and power. In SQL you need to think in terms of tables, inner joins, outer joins, foreign keys etc. whereas datalog you do everything with relations as in prolog.

Not only is it conceptually much simpler, it's also a "pit of success" situation as thinking in terms of relations instead of tables leads you towards normal forms by default.

Add the ability to automatically derive new facts based on rules and it just wins by a country mile. I recommend giving Soufflé a try.

I haven't worked with GraphDBs enough to comment on that.