Abstract: Why and how I chose my research directions

[Disclaimer ‘This is not a tech blog, but a general one. I just want to write a summary of various stages of my exploration and share it here since some people told me they find it helpful.’] When I was at PKU, I wasn’t particularly interested in academia and research, and spent most of my time doing internships in consulting. This included business consulting, real estate consulting, and pro bono consulting for NGOs. In the summer of 2013, my senior year, I realized I wanted to pursue a direction more aligned with my ambition and skill set. Given that my major was in urban science and economics, and after attending a lecture by Yu Zheng, then at Microsoft Research Asia, on smart cities, I decided to focus on smart cities and multi-agent interaction for my graduate research. I also learned about big data and machine learning during that period.

In 2014, I began my journey studying smart cities at UC Berkeley and continued my PhD in Systems Engineering in 2015. I delved into various urban systems, encompassing transportation, societal, and natural systems, and chose social-enabled urban data analytics as my PhD topic. However, when my advisor, Alexei, decided to pursue opportunities at a startup, my PhD came to an abrupt end in the summer of 2018.

Subsequently, I transitioned to the industry, joining A9. There, I pivoted to a new domain: search and natural language processing. It was a fresh field for me, but I embraced the challenge and was fortunate to grow my career with the rise of Transformers and LLM.

In spring 2023, coinciding with ChatGPT’s rise to fame, I was drawn to a domain closely related to my PhD — the initial enthusiasm that had driven me into research. I grew interested in understanding individual decision-making, multi-agent interaction and the optimal decision-making frameworks for platforms. I was presented with an opportunity to work on ad auctions, covering topics such as autobidding, allocation and pricing. I’m truly excited about this and hope that, looking back, the journey will all make sense.

[Disclaimer ‘These Views Are My Own’] [02/19/2024 Updates: ]It’s been a year since I became interested in the auction aspect of ad serving and monetization. I’ve made every effort to grasp each component of auction allocation, including ranker function design (such as bid squashing), pricing, and advertiser-level bidding (auto-bidding and bid shading), as well as pacing. I also had the opportunity to delve deeply into some fascinating topics like auction health evaluation metrics, causal methods for displacement cost, auction simulators, multi-agent systems for bidding and pacing, and online algorithms for bidding. Understanding so many things in a short period of time (06/2023-12/2023) has been a wild ride for me. However, after understanding the whole picture, I realized that the baseline methods are quite straightforward to implement, while more complicated methods might be very hard to deploy into production systems for some early-stage ad platforms.

At the same time, it’s interesting to see how the ad tech community responds to the AIGC trend. In my opinion, AIGC is adopted as a tool for advertisers to automatically generate campaigns, or at least automate some steps, such as generating the text or image of the ads, generating the target keywords for the ads, or even free-form targeting, or generating original bids and bid optimization rules based on customer input. However, from an ad serving and monetization level, I don’t see a good use for LLM unless it’s for LLM-enhanced embedding for ad retrieval. It is unnecessary to introduce LLM to response modeling and relevance, apart from LLM-based data verification or AI-based data labeling. It is impossible to use the current generative AI-based AIGC framework to replace auction, bidding, and pacing, maybe multi AI agent but non-trivial to beat the baseline.