CTR Prediction for DSP with Improved Cube Factorization Model from Historical Bidding Log

Abstract

In the real-time bidding (RTB) display advertising ecosystem, demand-side-platforms (DSPs) buy ad impressions through real-time auction or bidding from ad exchanges for advertisers. Receiving a bid request, DSP needs predict the click through rate (CTR) for ads and determine whether to bid and calculates the bid price according to the CTR estimated. In this paper, we address CTR estimation in DSP as a recommendation issue. Due to the complicated trilateral interactions among users, ads and publishers (web pages), conventional matrix factorization does not perform well. Adopting ideas from high-order singular value decomposition (HOSVD), we extend two dimensional matrix factorization model to three dimensional cube factorization containing users, ads and publishers, and propose an improved cube factorization model to address it. We evaluate its performance over a real-world advertising dataset and the results demonstrate that the improved cube factorization model outperforms the matrix factorization.