Search and Score-Based Waterfall Auction Optimization

Abstract

Online advertising is a major source of income for many online companies. One common approach is to sell online advertisements via waterfall auctions, through which a publisher makes sequential price offers to ad networks. The publisher controls the order and prices of the waterfall in an attempt to maximize his revenue. In this work, we propose a methodology to learn a waterfall strategy from historical data by wisely searching in the space of possible waterfalls and selecting the one leading to the highest revenues. The contribution of this work is twofold; First, we propose a novel method to estimate the valuation distribution of each user, with respect to each ad network. Second, we utilize the valuation matrix to score our candidate waterfalls as part of a procedure that iteratively searches in local neighborhoods. Our framework guarantees that the waterfall revenue improves between iterations ultimately converging into a local optimum. Real-world demonstrations are provided to show that the proposed method improves the total revenue of real-world waterfalls, as compared to manual expert optimization. Finally, the code and the data are available here.

6 Appendix

Appendix A - The Pseudo-code for the Monte Carlo Tree Search Algorithm

In this Appendix, we present the pseudo-code for the MCTS proposed algorithm. As opposed to the S &S-based algorithm, here, the algorithm will adopt the neighbor waterfall with the greatest revenue potential at every iteration; not necessarily the one with the current highest revenue. From an algorithm perspective, the difference is that there are two for loops in lines 18 and 20.

Appendix B - An Example for the Data Processing Flow

Table 4 describes the data processing from raw-data (Table 4a) into a valuation matrix (Table 4c), using Algorithm 1. Table 4b is the output of row 5 in Algorithm 1. For example, rows 1 and 4 that are marked in red-bold in Table 4a are the raw data of user ‘4421AB3’ and ‘G’ with a single impression each. These two rows are converted to a vector with (at least) the two entries ‘[0.02,0.19]’ that are marked with a red-bold box in Table 4b, before the beta distribution parameters, $Beta(\alpha =0.93,\beta =10.99)$, are estimated as marked in red-bold in Table 4c.

Table 4. Raw-data samples from the auction dataset ($Waterfall_A$) and their processed output.