This paper presents a decision theoretic ranking system that incorporates both explicit and implicit feedback. The system has a model that predicts, given all available data at query time, different interactions a person might have with search results. Possible interactions include relevance labelling and clicking. We define a utility function that takes as input the outputs of the interaction model to provide a real valued score to the user’s session. The optimal ranking is the list of documents that, in expectation under the model, maximizes the utility for a user session.

The system presented is based on a simple example utility function that combines both click behavior and labelling. The click prediction model is a Bayesian generalized linear model. Its notable characteristic is that it incorporates both weights for explanatory features and weights for each query-document pair. This allows the model to generalize to unseen queries but makes it at the same time flexible enough to keep in a ‘memory’ where the model should deviate from its feature based prediction. Such a click-predicting model could be particularly useful in an application such as enterprise search, allowing on-site adaptation to local documents and user behaviour. The example utility function has a parameter that controls the trade-off between optimizing for clicks and optimizing for labels. Experimental results in the context of enterprise search show that a balance in the trade-off leads to the best NDCG and good (predicted) click-through.

PDF file

In  SIGIR 2008 Workshop on Learning to Rank for Information Retrieval

Details