Learning to Tag is a Microsoft paper about suggesting tags associated to Flicker images. The authors use traditional textual co-occurences and visual features (such as colour histogram, and moment). These features are combined by using a RankBoost learning process.
The results are compared with a naive linear combination of ranking signals, and with simple tag co-occurences. It seems that Microsoft likes to combine different ranking signals with RankBoost, since I saw many papers describing variations of this idea and the performance they achieve seems quite good.