Bid Landscape Forecasting and Cold Start Problem With Transformers
In Real-Time Bidding, advertisers aim to optimally bid within a limited budget constraint. Effective bidding strategies require bid landscape forecasting to predict the probability distribution of market price for each advertisement auction. This distribution has a complicated form with many peaks. Moreover, all probabilities of bids depend on each other. Most existing solutions mainly focus on learning a parameterized model based on some heuristic assumptions of distribution forms. In this paper, we propose a Transformer model that takes into account dependencies between bids improving the bid landscape forecasting. We also increase the quality of model prediction on the advertisement cold start for the cases of insufficient data. Our experiments on two real-world industrial datasets prove that the proposed model statistically significantly outperforms the state-of-the-art solutions both in terms of ANLP metrics by 8.75% and ROC-AUC by 1.1%. In addition, we show the industrial applicability of our approach.