Data Science Lightning Talks 05 - Conference edition

The next Lightning talks will take place on October 11, 2022. This meetup will be hosted as a special section at the Slovenian Conference on AI 2022 at Jozef Stefan Institute.

Slovenian Conference on AI will start in the morning and will have 2 sections presenting academic papers. The third section starting at 17:00 is our meetup. As usual, the meetup will consists of several very short talks covering various data science and machine learning topics. As the idea is to spark conversation between AI researchers and AI practitioners, your talk can be on something you work on, something you stumbled upon, something you want to solve, or just something you want to share - as long as it is related to data science and machine learning. Submit your talk by September 30.

You're kindly invited to join the Slovenian Conference on AI as well, free of charge. The agenda is posted at the link below. If not, meetup starts at 17:00 and you'll have an opportunity to meet AI researchers after the lightning talks.

Conference agenda:
12:30 - 14:30 - Section A (6 talks)
14:30 - 15:00 - Coffee break
15:00 - 16:45 - Section B (5 talks) + best paper award
16:45 - 17:00 - Coffee break

Meetup agenda:
17:00 - 17:05 - Intro
17:05 - 18:05 - Lightning talks
18:05 - 19:00 - Discussion & refreshments

Talks:

Miha Jenko (Machine Learning Engineer @ Heureka Group)
Kubernetes is a hostile environment for ML pipelines
Kubernetes is a special computational environment that tries to address many needs of application developers and infrastructure teams. However, Kubernetes Pods are not the most suitable hosts for every machine learning pipeline due to their ephemeral nature.
ML pipelines frequently require stable computational environments because they might be running for days on end. Training jobs might be perfect and configured correctly, but due to specifics of environments like Google Kubernetes Engine we are not able to guarantee their completion. The talk will focus on ML Ops techniques and other considerations to increase resiliency of training jobs in Kubernetes in the public cloud.

Anže Alič (Intern @ Zemanta)
Deep & Cross models for CTR predictions
Click-through rate (CTR) prediction models present an extremely important part of online bidding. Over the last few years, researchers introduced many new deep learning architectures such as wide & deep, deep FM (factorization machine), etc. The main difference between these architectures is how they capture feature interactions. The Deep & Cross Network (DCN) does this with the Cross layer where we explicitly model feature interactions. In comparison to FM style models, DCN is more expressive while remaining cost-efficient.

Matjaž Žganec (Data scientist @ Genialis)
Feature reduction in development of robust clinical biomarkers
Humans have about 20,000 genes that control cellular structure and function. In disease models, we commonly identify subsets of genes related to particular biological functions. These feature sets, however, are often much larger than sample size. Furthermore, measurements of gene expression are susceptible to technical bias, measurement error,
biological variation, and other factors that affect some genes more than others. We will present methods of feature reduction that identify high quality genes, including gene transferability — a novel method pioneered by Genialis. Gene transferability ensures that the final signature gene set is transferable between different bias modalities such as tissue type, disease, assay platform, clinical lab, and testing location.

Roman Luštrik (Bioinformatician @ Genialis)
A rudimentary shiny app in hopefully under 5 minutes
For my next trick, I will try to convince audience that building an R shiny app is an easy process and there is nothing stopping you to deliver added value to your team in minutes. All this in under five minutes and live coding (but I practiced).

Miha Mlakar (Co-founder of Pareto d.o.o.)
Feature generalogic* – how to increase accuracy to your ML models
I will talk about how feature engineering and understanding the problem and using logic and common sense can help you increase the accuracy of your ML models.

Domen Krašovec & Črt Jarh (Data Scientists @ ZCAM)
Data Science at Zurich Insurance Group (ZCAM)
With insurance industry increasingly shifting in the more customer-centric direction, it's looking to adopt new analytical approaches that could help understand their customer needs and behavior. This led a historically very traditional industry to embrace opportunities offered by data science, machine learning and digital task automation. In our presentation we aim to present some of the opportunities and challenges posed by introduction of DS solutions to a global insurance corporation.