Sansan Tech Blog


We joined the ASSA 2023!

Hi, I'm Juan from Sansan R&D. This year started with a lot of action for us at the SocSci Group. As you may already know, we were invited to present at this year's ASSA Annual Meeting to make a presentation about the research project in which researchers Komatsu, Nishida and I, together with professor Angelo Mele, are working on.

Professor Mele and I had the pleasure of representing our team on-site, and it was a great experience. I will dedicate this post to talking about our participation and what we took home from it.

About the ASSA Annual Meeting 2023

The ASSA Annual Meeting is a yearly event organized by the American Economic Association (AEA) and the Allied Social Science Associations (ASSA), which gathers 64 scholarly associations. The AEA is one of the oldest academic associations in Economics. It is well known for being the association behind high-impact journals such as the American Economic Review.

So as you may imagine this is a big and very important event for researchers in Economics. It represents a very good opportunity for networking, learning and obtaining valuable feedback from top researchers in every possible field in Economics.

It was organized in New Orleans from January 6th to 8th, on several locations around the event's headquarters at the Hilton New Orleans Riverside. The hotel is right besides the Mississippi river, within a 15 minutes walk from representative locations such as Lafayette Square, the French Quarter and the New Orleans Ernest N. Morial Convention Center (where Miss Universe was held this year.)

The sunrise from the Mississippi River.
Great juggling performance in Bourbon St.

About Our Participation

The organization of the event was great. On registration, we were given several conference-related materials, a small notebook and even a BUSINESS CARD HOLDER. Since we like business cards a lot, the presents are very appreciated.

We were also given an N95 mask, which was necessary for joining all the events. There was a reception on the night of the 5th, which was a great opportunity to chat with other participants and to try some Louisiana cuisine. I honestly fell in love with grits.

Welcome reception

We were invited by professor Matthew Harding from University of California-Irvine. He attended our Mini Conference on Network Economics last year, where we presented some results from our project with professor Mele. Just in case you haven't heard of it, our research team is attempting to develop scalable methods and a software package that can deal with very large network datasets, for the estimation of structural models of network formation.

Back at the time of our conference professor Harding was thinking of organizing a session on Big Data and scalable methods at the ASSA 2023. He was very kind of inviting us to his session since our research topic was close to what he had in mind.

Our session was called Big Data: At Scale Methods and Applications (G3). You can find all the details of the session here, including our slides. It was scheduled at the first time slot of the first day of the conference.

The session began with our presentation. I was in charge of the initial part, where we talk about Eight, the data we use and the purpose of our research, and professor Mele was in charge of presenting the model and the results. Professor Vincent Boucher, who presented in our Mini Conference on May last year, came to see our session and give us very valuable feedback. I think that our presentation communicated very well the high level of quality of the data and the importance of our research.

Sleepy but ready to begin our presentation

The presentation by Breitmar, Harding and Lamarche was very interesting as well. In their research they study advertisement auctions from the point of view of the publisher, and explore methods that help deal with aggregated data. They employed data from a large social network, so in a sense their project has many similarities with ours. It was great to see that projects involving collaborations between the industry and academia are becoming more and more common in Economics.

Another topic we are always interested in is open source software development. The presentation by Hersch et. al. deals with the problem of optimally allocating public facilities in a populated territory, subject so several constraints defined by policy makers. One of the novelties of their study is the usage of large open spatial datasets, including Open Street Maps. They also plan to release their code as open source software packages for Python, R, Stata and more, and are open for collaborations. For anyone who is interested in applying their coding skills for the development of high-value public goods, I think this is a great opportunity.

Machine Learning Everywhere

Back in the days when I entered grad school Machine Learning was becoming a really hot topic in IT circles, but its application was barely existent in Economics. In fact, its usage in Economics was kind of frowned upon because of its lack of explainability. That's why it was a surprise for me to see so many sessions focused on Machine Learning in the ASSA 2023.

For example, there were at least eight posters related to Machine Learning. There were two sessions (this and this) related to Machine Learning for causal inference at the same time as our session on scalable methods (which was a shame, because I wanted to attend both!)

ML-based methods can uncover causal relationships that are undetected by traditional methods

One interesting topic worth highlighting is the usage of adversarial estimation for causal inference, which was covered in the Adversarial Methods session. Adversarial estimation is a method that is analogous to Generative Adversarial Networks (GANs). Similar to their deep learning counterparts, adversarial methods in causal inference consist of a generator, which proposes a model, and a discriminator, which performs some classification task on the output from the generator. Because both models compete, the estimation task is a minimax problem.

When applied to the method of moments, the discriminator gets rewarded by finding large violations to the moment conditions, while the generator tries to minimize that reward by proposing a model that satisfies those conditions. In the case of structural estimation, the method is comparable to the simulated method of moments: the generator produces simulated data from structural model based on its proposed parameters, and the discriminator is a classifier that tries to distinguish between the actual data and data simulated by the model.

Closing Words

Joining the conference was a great experience and a great opportunity to share our research with top researchers from all fields in Economics. If I could summarize my impressions, I would say that:

  • The nature of research may be changing, moving towards open source projects and more collaboration between industry and academics to solve real world problems.
  • Machine Learning has become one more tool in the econometric tool box of many economists. Maybe this is a good time for everyone in Economics to catch up with the literature, and to brush up our coding skills.
  • Work opportunities for economists in private companies are increasing, and large Economics conferences are becoming great places for networking with researchers in the IT industry.

So what's next for us? We have submitted our paper for the conference's proceedings and are working now on finishing our reproducibility package. The AEA has a very high standard for reproducibility, so we are learning a lot from this experience. We will keep you updated once our paper is out.

© Sansan, Inc.