Guessing the relevance of delivered search results is one of the biggest issues for today’s search engines. The particular problem is that it’s difficult to obtain explicit statements from users about whether they found what they were searching for. Clicks are commonly used to guess relevance (using so-called “click models”) but they are far from being a perfect indicator. Particularly, a user might click a search result, but then return to the results page because the visited webpage was useless. Also, it’s possible that no clicks happen at all if the desired piece of information is already shown on the results page (e.g., in terms of an info box).
To tackle the above shortcoming, we have investigated the suitability of implicit feedback in terms of mouse cursor interactions for predicting the relevance of search results. For this, we developed StreamMyRelevance!—a system that receives streams of interactions and relevance judgments and trains statistical models from these in near real-time. The models can then be used to infer relevance from interactions in the future. The relevance judgments we’re using to train our models can either be implicit (e.g., a completed booking process in the case of hotel search) or explicit (e.g., statements by paid quality raters/crowdworkers).
Analysis of a large amount of real-world interaction data from two e-commerce portals showed that StreamMyRelevance! is able to train good models that show the tendency to perform better than a state-of-the-art click model solution1 that is successfully used in industry. Our results particularly underpin the benefit of using interaction data other than clicks for guessing the relevance of search results.
We have summarized the design and evaluation of our system in a full research paper2 that will be presented at the 2014 International Conference on Web Engineering (ICWE). The conference proceedings will be published by Springer and the final version of our paper will be available at link.springer.com. Special thanks go to Sebastian Nuck, who helped with development and evaluation of StreamMyRelevance! in the context of his Master’s Thesis at Leipzig University of Applied Sciences.
1 Chao Liu, Fan Guo, and Christos Faloutsos (2009). “BBM: Bayesian Browsing Model from Petabyte-Scale Data”. In Proc. KDD.
2 Maximilian Speicher, Sebastian Nuck, Andreas Both and Martin Gaedke (2014). “StreamMyRelevance! Prediction of Result Relevance from Real-Time Interactions and its Application to Hotel Search”. In Proc. ICWE.