AAAI 2024 Highlights
I attended the 2024 edition of the AAAI conference in Vancouver last week to present our paper, “Maximizing the Success Probability of Policy Allocations in Online Systems”. Here is my recap.
AAAI is one of the leading conferences in Artificial Intelligence. Criteo AI Lab is focused on applying AI to problems in computational advertising. I attended the 2024 edition to present a paper that came out of a nice collaboration between multiple teams at Criteo AI Lab.
The AAAI 2024 conference continues the trend of AI conferences growing larger and bigger. According to some numbers discussed during the opening session, the 2024 edition counted 5193 attendees and 2340 main track papers.
It appears that this was the first time that not all accepted papers did an oral presentation, which avoided the multiplication of parallel tracks. Instead, three enormous poster sessions were held, each dealing with one third of the accepted papers. This enabled the organizers to propose a more human schedule, with only a few parallel tracks for the papers selected for an oral presentation.
One concern with big poster sessions is finding what could be interesting for your research. The organization was very receptive in this regard, as they changed the poster map for the final session to group posters according to topics (instead of using the submission numbers, that were used for the first two sessions).
The event was truly international. Sadly, many presentations couldn’t be done by the authors of the papers because of visa issues. I find it notable that China contributed more than 40% of the 12100 paper submissions.
It appears that assisting participants and organizers in sorting out the information and matching authors with readers could be a highly commendable RecSys initiative…
Criteo participation
Our paper: Maximizing the Success Probability of Policy Allocations in Online Systems
I wrote this paper with several researchers and engineers from Criteo. In this paper, we consider a decision maker who gets an incoming traffic of users. For example, in the context of display advertising, the decision maker can be the agent who bids in the real-time auctions.
The situation is as follows. There is a set of users, we can group them by buckets based on what we know about them, and a set of policies, and we need to assign the groups of users to those policies. We have past data that allows us to estimate the impact of allocating a policy to a user. It is natural given the data and the objective to formulate the problem as a value maximization problem under cost constraints.
This suffers important limitations.
- How do you account for stochasticity?
- What if the tradeoffs cannot be explained as a problem of maximizing value subject to some cost constraints?
To address this limitation, we define a success region, and we optimize for the chance that our policy allocation will lead to this success region. An important observation is that if you use a multivariate Gaussian to approximate how a policy affects a group of users — — which is very natural — — then the problem is tractable.
If I managed to arouse your curiosity, I added the poster below (and there is also the paper).
My main takeaways
A significant portion of the opening session was dedicated to some cases of fraud, especially collusion rings for the review cycle, duplicate submissions, and plagiarism. I find it quite positive that the organizers decided to talk about it. You can see the presentation here (access required).
Although language and vision dominated the conference, I was happy to see a lot of game theoretical papers on computational social choice. For instance, this one, this one, and this one.
There was an excellent plenary on LLMs where Chris Manning addressed this (humorous, I believe) warning to those who are indifferent to the LLMs’ progress: “If you are not just blown away by the fantastic progress (…) I honestly think you have no soul.” (if you have access, you can watch the full discussion here )
I found the presentation by Michael M. Bronstein on geometric deep learning very inspiring. The material of this presentation can be found in a monograph shared on arxiv.
Check all the stories from our research team 👉 https://medium.com/criteo-engineering/tagged/research
