Conclusion

Abstract

In this book, we have studied multi-mode prediction problems. The focus was on problem settings with large categorical domains and high sparsity. Due to the sparsity and usually high imbalance, we are not interested in classification but in ranking the entities of one of the modes. Instead of creating one global ranking, the rankings should be context-aware – i.e. we create many rankings that depend on a given context. Important applications for this setting are recommender systems. There exist several recommender tasks, two of the most well studied ones are personalization and tag recommendation. Our developed method of context-aware ranking subsumes both of them and includes also other settings like time-awareness. Moreover other well-known applications like web search or multi-label classification (e.g. annotation, wikipedia categorization) can be seen as an instance of context-aware ranking.