A Practical Learning to Rank Approach for Smoothing DCG in Web Search Relevance

Опубликовано 17 августа 2016, 20:55

Discounted cumulative gain (DCG) is now widely used for measuring the performance of ranking functions especially in the context of Web search. It is therefore natural to learn a ranking function that directly optimizes DCG. However, DCG is non-smooth, rendering efficient gradient-based optimization algorithms inapplicable. To remedy this, smoothed versions of DCG have been proposed but with only partial success: they have yet to outperform other learning to rank algorithms using simple loss functions such as those based on pairwise preferences. In this talk, we first present analysis that shows it is ineffective using the gradient of the smoothed DCG to drive the optimization algorithm. We then propose a series of approaches that can significantly improve the optimization results of the smooth DCG cost function.