Functional Bid Landscape Forecasting for Display Advertising

Abstract

Real-time auction has become an important online advertising trading mechanism. A crucial issue for advertisers is to model the market competition, i.e., bid landscape forecasting. It is formulated as predicting the market price distribution for each ad auction provided by its side information. Existing solutions mainly focus on parameterized heuristic forms of the market price distribution and learn the parameters to fit the data. In this paper, we present a functional bid landscape forecasting method to automatically learn the function mapping from each ad auction features to the market price distribution without any assumption about the functional form. Specifically, to deal with the categorical feature input, we propose a novel decision tree model with a node splitting scheme by attribute value clustering. Furthermore, to deal with the problem of right-censored market price observations, we propose to incorporate a survival model into tree learning and prediction, which largely reduces the model bias. The experiments on real-world data demonstrate that our models achieve substantial performance gains over previous work in various metrics. The software related to this paper is available at https://github.com/zeromike/bid-lands.