Real-Time Optimization of Web Publisher RTB Revenues

Abstract

This paper describes an engine to optimize web publishers revenue from second-price auctions, which are widely used to sell online ad spaces in a mechanism called real-time bidding. This problem is crucial for web publishers, because setting appropriate reserve prices can increase significantly their revenue.We consider a practical setting where the only available information before an auction occurs consists of a user identifier and an ad placement identifier. Once the auction has happened, we observe censored outcomes : if the auction has been won (i.e the reserve price is smaller than the first bid), we observe the first bid and the closing price of the auction, otherwise we do not observe any bid value.The engine predicts an optimal reserve price for each auction and is based on two key components: (i) a non-parametric regression model of auction revenue based on dynamic, timeweighted matrix factorization which implicitly builds adaptive users' and placements' profiles; (ii) a non-parametric model to estimate the revenue under censorship based on an on-line extension of the Aalen's Additive Model.An engine very similar to the one described in this paper is applied to hundreds of web publishers across the world and yields a very significant revenue increase. The experimental results on a few of these publishers detailed in this paper show that it outperforms state-of-the-art methods and that it tackles very efficiently the censorship issue.

Optimizing reserve prices for publishers in online ad auctions

“…The problem of maximizing revenues in online advertising has received increasing attention in the machine learning literature over the last decade (see e.g. [1], [2], [3], [4], [5]). Some works such as [6], [7], [8] try to estimate or predict the winning bid (possibly in the presence of censoring) by using a historical dataset containing the top two bids and additional features relating to the context of the user visiting a website.…”

A PSO-based Algorithm for Reserve Price Optimization in Online Ad Auctions

“…Most of the studies mentioned above do not set reserve prices in an adaptive way that adjusts to changing environments. Some studies such as [5], [7], [8] do study adaptive reserve prices, but they assume that the winning bid and/ or second highest bid are observed. In this paper we do not make this assumption.…”

“…That is, the sample used to determine contains an unsuccessful sale. If there is no unsuccessful sale while determining , then 7is used if ( − , ) > 0 and (8) is used if ( − , ) < 0 with + 2 > − 3 . If there is an unsuccessful sale while determining , then (9) is used if ( − , ) > 0 and 10is used if ( − , ) < 0 with + 4 < − 5 .…”

“…By specifying appropriate values (the exact values are discussed in Section V) for the acceleration coefficients, the dynamics in (7) -(10) ensure that the locations of the particles get updated in the most informative direction relative to the value of . In particular, (7) and (8) ensure that the locations , are updated to explore values above (if there are no unsuccessful sales), and (9) and (10) ensure that the locations , are updated to explore values below (if there are unsuccessful sales). The motivation for these dynamics comes from the definition of the critical boundary as specified in Algorithm 1.…”

Optimal Allocation of Real-Time-Bidding and Direct Campaigns

“…In [12], the reserve prices are set in a static way (i.e. an optimal reserve price is estimated on a large set of auctions), whereas [15] and [3] define methodologies to set dynamic reserve prices which are predicted in real time before an auction happens. [11] and [10] study the optimal delivery of a budget across time in the RTB market.…”