Optimizing ranking for response prediction via triplet-wise learning from historical feedback

In the real-time bidding (RTB) display advertising ecosystem, when receiving a bid request, Demand-side platform (DSP) needs to predict user response on each ad impression and determines whether to bid and calculates the bid price according to its prediction. When given a fixed advertising budget, in order to maximize the return on investment (ROI), DSP aims to buy in more conversions and then more clicks than non-clicks. In this paper, we consider response prediction problem as a ranking problem for impression chances and propose a triplet-wise comparison based learning optimization which derived from Bayesian personalized ranking (BPR) based on pairwise learning to learn model parameters. Pairwise learning can only employ one type of historical click and conversion information through optimizing the correct order of random pair of a positive and a negative example for binary classification. While triplet-wise learning combines these two kinds of historical response information into the same model through taking into consideration the correct order of the pair of conversion and click-only as well as the pair of click-only and non-click. Since our method accomplishes the click and conversion prediction tasks in the same predicting procedure, our algorithm is good at ranking click impressions higher than non-click ones and conversion impressions higher than click-only ones. In this way, under a fixed budget, biding algorithm would preferentially buy in more conversions than others and then more clicks than non-clicks. Our experiments demonstrate that the improved method not only outperforms both pairwise and MSE schemes on three classes ranking in terms of multi-AUC, NDCG etc., but also, outperforms others on binary classification for click and non-click on the targeted real-world bidding log data owing to the introduction of historical conversion information.