AlgoIL #7 - Clustering and Sampling

Details
AlgoIL is back after the holidays with two surprising talks about creative solution to seemingly simple tasks that turn out to be hard problems!

Agenda:
18:00 Gathering and snacks
18:30 Opening words
18:40 Dr. Inbal Budowski-Tal - Generality Speaking - an Algorithm for Clustering Short Sentences
19:25 Short break
19:35 Shaked Zychlinski - Designing Algorithms for Fast Weighted Sampling in Production
20:20 Time for more snacks and chats
_______________________________________________________________________________________
Both talks will be in Hebrew, will be filmed and uploaded to our YouTube channel.
Thanks EverCompliant for hosting this event!
_______________________________________________________________________________________

Dr. Inbal Budowski-Tal - Generality Speaking - an Algorithm for Clustering Short Sentences

Clustering short sentences - sounds like a rather trivial task, doesn’t it? Indeed, many tools are already out there, but the deep and fascinating world of NLP keeps surprising.
Join me on my journey in tackling this apparently-not-so-trivial task. We will understand the background and the complexity of the task, dust ourselves off after some trial and error, and finally come to the solution - a mixture of word embedding, TF-IDF, and a novel concept of Word Generality - which is a measure of a word's contribution to the meaning of the sentence.
In this talk, we will cover both basic and advanced topics in NLP, and most importantly - we will experience the joy of finding a hard problem, and then finding the right algorithm to solve it!

About the speaker:
Dr. Inbal Budowski-Tal
Director of AI, EverCompliant
A versatile researcher, heading the AI team at EverCompliant, a FinTech startup in the world of Risk Analysis. Currently focusing on NLP, formerly a bioinformatics researcher at a pharmaceutical startup, before that handled user signals at Microsoft, and even before that studied geometrical data in my Ph.D. My passion towards data is agnostic to its type, and I believe that the basic understanding of how to represent it, measure it and model it is applicable in various business domains.

_______________________________________________________________________________________

Shaked Zychlinski - Designing Algorithms for Fast Weighted Sampling in Production

Taboola's core business is to match personalized content to users of thousands of websites across the globe. We do so (quite successfully) by using state-of-the-art deep learning models, which learns from the feedback of millions of users. Unfortunately, DL-based recommendation systems can easily repeat the same action over and over again, and can become redundant eventually. We therefore must stray from our system's recommendations, and doing so intelligently isn't a simple task. In this talk we'll cover why naive methods of exploration fail, discuss different methods of smart exploration and demonstrate how we developed a weighted sampling method for exploration, including its necessary mathematical proof.

About the speaker:
Shaked Zychlinski is an algorithm engineer at Taboola, where he works on machine-learning applications for recommendation systems. He fell in love with this field when he trained a neural-network to play Tic-Tac-Toe, and it pulled a trick on him and won the game. He has a B.Sc and M.Sc in physics, and a constant crave for cookies.

_______________________________________________________________________________________