Hello! My name is Akos and I lead the News Feed Relevance data science team (which includes News Feed, Stories and Notifications ranking) at Facebook across our Menlo Park and New York offices.
Our team's mission is to improve the way posts are ranked in News Feed through hands-on quantitative research and insights, applied science, machine learning and innovation. We are part of the larger Analytics organization at Facebook, and work in close partnership with the product and engineering teams, as well as academic teams at Facebook. We are always looking for bright, applied-minded candidates. If interested, feel free to reach out (https://www.linkedin.com/in/akos-lada-737aba31/). Below you can find an overview of some recent work by me and my colleagues in case you are interested, and also check out an overview of the ranking algorithm here: https://newsroom.fb.com/news/2018/05/inside-feed-news-feed-ranking/.
I joined Facebook in July 2015, and first worked on the Core Data Science group's Economics team on questions ranging from quality to personalization in both News Feed ranking and the advertisements space. A few months later we created a new News Feed Science team to focus fully on News Feed. In 2017 I transferred to the larger Facebook Analytics team and have built up the News Feed Relevance Data Science team in the following years.
Our team's work tends to fall into the intersection of social sciences and machine learning / statistics methodology. It's hard to cleanly separate out the two parts, but here are examples of projects that are heavier on the former vs the latter side:
The first type is the more social science -leaning work:
For instance, when we rank and make improvements to News Feed, we rely on a set of core values to guide our thinking and help us keep the central experience of News Feed intact as it evolves. One of our values is that the stories in your feed should be informative.
What is the optimal way to create a ranking signal to predict what is most informative, so that these stories can appear higher in your feed? We have researched multiple approaches and we have found that the best approach is to directly talk to people. We now regularly ask members of our Feed Quality Program to look at each story in their feed and rank it on a scale of one to five - one being “really not informative” and five being “really informative.” This program includes global crowd-sourced surveys of tens of thousands of people per day, as well as people who answer more detailed questions about what they like seeing in their feeds.
We then use machine learning to combine this signal with how relevant the story might be to you personally - taking into account things like your relationship with the person or publisher that posted, or what you choose to click on, comment on or share - to best predict stories that you might personally find informative.
We launched this change in August 2016. You can read the public communication about it here and read the popular coverage in Techcrunch or CNN here and here.
Another question we worked on: how can we show people better, more authentic public content?
We have heard from our community that authentic stories are the ones that resonate most — those that people consider genuine and not misleading, sensational or spammy. So we decided to use machine learning models to rank this content higher.
We launched this change in January 2017, you can read the public communication about it here and read coverage of it in Techcrunch or the BBC here and here.
The second type of work leans heavier on methodology. We create novel solutions in causal inference and machine learning to solve social science style problems we face at Facebook, such as how can we model meaningful social interactions, network effects or what kind of content users are interested in, all of which can build into the News Feed ranking algorithm. We combine the latest statistical/econometric (eg causal inference, heterogeneous treatment effect estimation) and machine learning methodologies (eg gradient boosted regression trees, neural networks, embeddings, lasso, ridge regression, elastic nets) to create new methodologies to solve complex problems. Some publicly available examples from this body of work:
One example of this type of work is a method that Alex Peysakhovich and I recently developed for personalization. In a nutshell about this work: imagine you work on a new page recommendation unit and want to find out which users would like to see more page content on Facebook because then we could show more page recommendations to these users. We came up with the following simple idea: we can use observational data (de-identified and aggregated) to capture correlations and under certain statistical conditions, these correlations help us find out which types of users benefit the most from a certain policy. For page recommendation, for instance, we can see what characterizes people who usually spend more time on Facebook on days when they see more page posts and show more recommendations to these types of people.
You can find more details about our method here: https://dl.acm.org/doi/abs/10.1145/3328526.3329558 (best paper award at the ACM Conference on Economics and Computation conference in 2019) and https://arxiv.org/abs/1611.02385 (presented at the Code @ MIT Conference on Digital Experimentation in 2016).
More generally, I like working on complex and interesting ecosystem-level questions using quantitative tools. Before coming to Facebook I finished a PhD in Political Economy at Harvard University between 2010 and 2015, under the guidance of James Robinson and Andrei Shleifer. Doing research at Harvard was not only intellectually stimulating but a lot of fun too! One of my main research papers focused on the causes of international conflict, where I showed (using game theory and econometric/statistical analysis) that cultural similarity often makes two countries more likely to fight, especially when the two countries have different political institutions because a dictator is keen to destroy a culturally similar democracy that could serve as an example for the dictator's citizens. You can read more about my paper here (https://papers.ssrn.com/sol3/papers.cfm?abstract_id=2452776) or in the popular media after the Russia-Ukraine conflict broke out (https://www.washingtonpost.com/news/monkey-cage/wp/2014/03/04/russia-vs-ukraine-a-clash-of-brothers-not-cultures). Before Harvard, I studied quantitative economics and had earned a degree in macroeconomics from Corvinus University of Budapest in my native Hungary.
If you are interested in joining our team, check out open positions on my Ranking Analytics team (https://www.facebook.com/careers/jobs/a0I1H00000K6qCxUAJ/). We are currently hiring for both Menlo Park and New York City.
Predicting Long-Term Outcomes in Experiments Through an Individual-Level Prediction Model (2020). Yuezhe Yao, Yuwen Zhang, Drew Dimmery, Daniel Day, Eytan Bakshy, Akos Lada. Code@MIT http://ide.mit.edu/events/2020-conference-digital-experimentation-code
Observational Data for Heterogeneous Treatment Effects with Application to Recommender Systems (2019). Akos Lada, Alexander Peysakhovich, Diego Aparicio, and Michael Bailey. EC '19 Proceedings of the 2019 ACM Conference on Economics and Computation. Pages 199-213. https://dl.acm.org/citation.cfm?id=3329558&dl=ACM&coll=DL#URLTOKEN#
Combining observational and experimental data to find heterogeneous treatment effects (2016). Alexander Peysakhovich, Akos Lada. https://arxiv.org/abs/1611.02385
The Power of Culture: Cultural Variables are the Best Country-level Predictors of How Individuals Interact on Social Media. Akos Lada (Facebook) and Alexander Peysakhovich (Facebook) poster. Code@MIT 2016. http://ide.mit.edu/sites/default/files/CODE%202016%20Program.pdf Email for availability
Clash of Brothers in a Contagious World: Wars to Avoid Diffusion (2015). IEHAS Discussion Papers 1333, Institute of Economics, Centre for Economic and Regional Studies, Hungarian Academy of Sciences. https://ideas.repec.org/p/has/discpr/1333.html
See more at: https://scholar.google.com/citations?user=BTl2mnsAAAAJ&hl=en
How Facebook's News Feed Ranking works to give you the most relevant content. (2019) Talk given at Crunch Data Engineering and Analytics Conference 2019: https://crunchconf.com/speaker/akoslada Email for availability
Designing Facebook's Algorithms (2018, 2019): talks at the Stanford Graduate School of Business and MIT Sloan Business School
Systems and methods for evaluating user activity (2019), US Patent 10,455,033, Akos Lada, Alexander Peysakhovich
Systems and methods for providing data analysis based on applying regression (2017), US Patent App. 15/147,805, Akos Lada, Alexander Peysakhovich
Systems and methods for providing data analytics based on geographical regions (2017), US Patent App. 15/144,617, Akos Lada, Alexander Peysakhovich
Facebook algorithm announcement (2017): New Signals to Show You More Authentic and Timely Stories. Akos Lada, Research Scientist, James Li, Research Scientist, and Shilin Ding, Engineering Manager https://newsroom.fb.com/news/2017/01/news-feed-fyi-new-signals-to-show-you-more-authentic-and-timely-stories/
Facebook algorithm announcement (2016): Showing You More Personally Informative Stories. Jie Xu, Research Scientist, Akos Lada, Data Scientist, and Vibhi Kant, Product Manager, News Feed https://newsroom.fb.com/news/2016/08/news-feed-fyi-showing-you-more-personally-informative-stories/
Washington Post article: Why China won’t let Hong Kong democratize (2014). https://www.washingtonpost.com/news/monkey-cage/wp/2014/10/02/why-china-wont-let-hong-kong-democratize/
Washington Post article: Russia vs. Ukraine: A clash of brothers, not cultures (2014). https://www.washingtonpost.com/news/monkey-cage/wp/2014/03/04/russia-vs-ukraine-a-clash-of-brothers-not-cultures/