Risk-Sensitive Deep Neural Learning to Rank

Learning to Rank (L2R) is the core task of many Information Retrieval systems. Recently, a great effort has been put on exploring Deep Neural Networks (DNNs) for L2R, with significant results. However, risk-sensitiveness, an important and recent advance in the L2R arena, that reduces variability and increases trust, has not been incorporated into Deep Neural L2R yet. Risk-sensitive measures are important to assess the risk of an IR system to perform worse than a set of baseline IR systems for several queries. However, the risk-sensitive measures described in the literature have a non-smooth behavior, making them difficult, if not impossible, to be optimized by DNNs. In this work we solve this difficult problem by proposing a family of new loss functions -- \riskloss\ -- that support a smooth risk-sensitive optimization. \riskloss\ introduces two important contributions: (i) the substitution of the traditional NDCG or MAP metrics in risk-sensitive measures with smooth loss functions that evaluate the correlation between the predicted and the true relevance order of documents for a given query and (ii) the use of distinct versions of the same DNN architecture as baselines by means of a multi-dropout technique during the smooth risk-sensitive optimization, avoiding the inconvenience of assessing multiple IR systems as part of DNN training. We empirically demonstrate significant achievements of the proposed \riskloss\ functions when used with recent DNN methods in the context of well-known web-search datasets such as WEB10K, YAHOO, and MQ2007. Our solutions reach improvements of 8% in effectiveness (NDCG) while improving in around 5% the risk-sensitiveness (\grisk\ measure) when applied together with a state-of-the-art Self-Attention DNN-L2R architecture. Furthermore, \riskloss\ is capable of reducing by 28% the losses over the best evaluated baselines and significantly improving over the risk-sensitive state-of-the-art non-DNN method (by up to 13.3%) while keeping (or even increasing) overall effectiveness. All these results ultimately establish a new level for the state-of-the-art on risk-sensitiveness and DNN-L2R research.