Counterfactual Learning-to-Rank for Additive Metrics and Deep Models.

Implicit feedback (e.g., clicks, dwell times) is an attractive source of training data for Learning-to-Rank, but it inevitably suffers from biases such as position bias. It was recently shown how counterfactual inference techniques can provide a rigorous approach for handling these biases, but existing methods are restricted to the special case of optimizing average rank for linear ranking functions. In this work, we generalize the counterfactual learning-to-rank approach to a broad class of additive rank metrics -- like Discounted Cumulative Gain (DCG) and Precision@k -- as well as non-linear deep network models. Focusing on DCG, this conceptual generalization gives rise to two new learning methods that both directly optimize an unbiased estimate of DCG despite the bias in the implicit feedback data. The first, SVM PropDCG, generalizes the Propensity Ranking SVM (SVM PropRank), and we show how the resulting optimization problem can be addressed via the Convex Concave Procedure (CCP). The second, Deep PropDCG, further generalizes the counterfactual learning-to-rank approach to deep networks as non-linear ranking functions. In addition to the theoretical support, we empirically find that SVM PropDCG significantly outperforms SVM PropRank in terms of DCG, and that it is robust to varying severity of presentation bias, noise, and propensity-model misspecification. Moreover, the ability to train non-linear ranking functions via Deep PropDCG further improves DCG.