We would like to thank Taboola for hosting us after a long hiatus.
This time, we would meet (physically!) and talk about recommendation engines.

Agenda:
18:00-18:30 Gathering
18:30-18:45 A word from our sponsor (Taboola)
18:45-19:15 RMaking N Populations Work in 1 Model Architecture
(Marina Gandlin, Data Science Team Lead, Taboola Algo)
19:15-19:30 Break
19:30-20:00 Contextual Bandits for Pricing (Daniel Hen, Data scientist / Fyber)
20:30-21:00 Offline metrics for Recommendation systems (Benjamin Kempinski / Argmax)

========
Making N Populations Work in 1 Model Architecture (Marina Gandlin)

Homogenous labeled datasets and relatively easy to model with a single simple neural network. However, scarcity of labeled data sometimes forces us to combine different populations in order to have enough data for training, and thus creating a heterogeneous dataset. For example people from different countries might show very different behavior although they may share common characteristics. How do we make sure our model is both accurate and fair for all populations? In this talk we will shortly describe a nice architecture trick we use to allow the model to enjoy both generalization from multiple populations and specialization for each of them.

Contextual Bandit for Pricing (Daniel Hen)

Contextual bandits are commonly used for recommendation, in which each item is seen as a categorical variable.
We extended the use of bandits for Real-time-bidding auction, by modeling the target variable as ordinal.
In this talk we would cover experiments we took, and give a brief overview on Bandits

Offline Metrics for Recommendation (Benjamin Kempinski)

Recommendation systems are usually tested online. While this process is accurate, it is costly and not scalable. In this talk we would cover how similarity based recommendation can be modeled as a regression problem, and we will share a dashboard we designed for offline simiarity experiments.