Advanced Recommender Systems: Insights, Techniques, and Applications

Y-DATA Meetup #22
Advanced Recommender Systems: Insights, Techniques, and Applications
Hosted by Taboola. Talks are in English. More info about Y-DATA.
Directions: Jabotinsky 2, Ramat Gan, Atrium Tower ( Floor 23)
Google meet link for those who gonna join remotely

Agenda:
18:00 - 18:30 Registration, Mingling, Snacks & Beer
18:30 - 18:45 Opening, Host and Organizer words.
18:45 - 19:15 "Waterfall Auction Recommendations with Dynamic Programming״
Dan Halbersberg, DS Team Lead at Playtika
19:15-19:45 "Intro to Vector Search"
Uri Goren, founder of Argmax
19:45-20:15 "Uncertainty estimation in large scale recommender systems"
Alexey Pechorin, Data Scientist at Taboola

Abstracts:

1. "Waterfall Auction Recommendations with Dynamic Programming"
Over the last two decades, an increase in advertisement slot sales from online auctions has led to significantly stronger revenues for publishers. One of the most relevant and widely used auction protocols is called waterfall strategy, through which the publisher makes sequential, pre-defined, ad-requests to ad-suppliers through third party ad networks. Each ad-request targets a specific ad network with a specific price. As different waterfall strategies potentially lead to diverse total revenue, it is extremely beneficial for the publisher to optimize the strategy they use. However, the number of possible waterfall strategies grows exponentially with the number of ad networks and prices in play, which makes waterfall optimization a formidable task. In this presentation, we talk about the dual nature of the waterfall strategy as both an auction protocol and a recommendation technique. We propose a dynamic-programming based algorithm that recommends modifications to the waterfall strategy, empowering media buyers to maximize their revenues. This method has been extensively tested using real-world data and industry A/B testing, outperforming state-of-the-art methods and even human experts

2. "A brief intro to vector search"
While recommendation systems can wear many forms, lookalike search became the most dominant approach.
At its core, similarity search depends on semantic representation and an efficient search.
In this talk, we would introduce vector similarity search and discuss how to control the trade off between accuracy and performance

3. "Uncertainty estimation in large scale recommender systems"
Taboola's recommender system utilizes a clickthrough rate (CTR) model to estimate ad revenue but aims to also output uncertainty estimations to manage risk-reward tradeoffs and drive feature exploration. Traditional uncertainty estimation involves training multiple models, which is costly. We examine two methods that approximate uncertainty with a single model: Variational AutoEncoder (VAE) inspired sampling layers and Mixture of Local Experts (MoLE) based approximations. Both methods, by introducing non-determinism and multiple model outputs respectively, showed high correlations, with MoLE providing superior results. This talk discusses these paradigms, their comparisons, and their pros and cons.

Speakers:

Dan Halbersberg received his Ph.D degree in Industrial Engineering and Management from Ben Gurion University of the Negev, Beer Sheva, Israel in 2020. Since then he was postdoctoral researcher at the same university, and senior data scientist at Playtika Al lab. Currently, Dan manages the AI-marketing group at Playtika’s AI lab. Dan has 12 years of experience, leading data science projects in the industry and academia and publishing papers in top journals and conferences. His main interest are graphical models, analysis of temporal longitudinal data and deep neural network.

Uri Goren is a machine learning expert, leading argmax.ml, a recommendation system consulting firm.
Uri has a rich experience, having worked at several fortune 500 companies such as Microsoft, Intel, AT&T as a data scientist, and founded 2 ML centered startups.
On his free time, Uri hosts Explainable, a ML podcast in Hebrew, and organizes the Tel Aviv chapter of the PyData community

Alexey Pechorin is a seasoned Machine Learning Engineer in Taboola’s Algorithms department, focusing on bidding algorithms to serve recommendations. Alexey possesses MSc in CS from Tel Aviv University.