Sequential Decision Making in Real-time Digital Advertising

点击翻译按钮获取中文摘要

Digital marketing has been transformed by the development of real-time bidding platforms. These platforms allow advertisers to decide sequentially whether to place a bid on a particular advertising location (e.g., a website), how much to bid on and who to bid on. The speed of the markets means that these decisions need to be taken almost instantaneously, which has led to many algorithmic approaches being used. Another key aspect of decision-making in the digital advertising world involves the balance between privacy and utility. The ideal situation for the marketer is to have access to the user's entire online browsing history whereas the ideal set up for an internet user is for the marketers to have minimal access to their information. This leads to a trade-off between what is the most affordable and possibly relevant browsing experience a user can have with the information they are willing to share with the marketers. The nuances of the real-time bidding problem mean that directly applying existing algorithms for sequential decision making may lead to sub-optimal decisions being taken. Therefore, the project will involve designing and analysing new algorithms tailored to specific settings in digital advertising, including privacy preserving learning. This has become extremely important and almost an existential threat for many advertising technology companies in recent time due to the deprecation of "cookies" (most popularly used digital identifier used to track user's online activity) by Google Inc. Without access to an identifier, how does one deliver a personalized privacy preserving browsing experience for an internet user? In this project, we aim to design new algorithms that draw ideas from reinforcement learning algorithms that deal with sequential decision-making with uncertainty, semi supervised learning algorithms that use aggregate information or partially labelled data to learn and optimize policies and active learning methodologies that focus on how to learn efficiently with limited samples, particularly when labelling data is expensive. This project falls within the EPSRC research themes of "Operational Research", "Statistics and Applied Probability" and "Digital economy" and aligns with EPSRC's strategic priority related to "artificial intelligence, digitisation and data: driving value and security". This research is being done in collaboration with "Imperial College, London" and "The Trade Desk", which is a multinational company that specializes in real-time programmatic marketing automation technologies. Towards this, our goals are two-fold: a) Design rigorous algorithms with provable guarantees b) Demonstrate the efficacy of our algorithms on real data and drive towards real world impact.