PyData TLV - Lightning Talks #3

We would like to thank Salesforce for sponsoring this meetup.

The next meetup will be Lightning talks meetup, 10 speakers talking 10 minute talks (sharp), on a python library, a project, or a cool DS application.

Lightning talk format:

10 minute talk (No questions, sorry)

2 minutes - Q&A while the next speaker gets ready

Agenda:

18:00-18:30 - Gathering

18:30-18:45 -A word from Salesforce, Our sponsor

18:45-18:50 - Guidelines and starting the stopwatch

18:50-19:05 - Clustering before labeling (Uri Goren)

19:05-19:20 - Functional data manipulation with pandas (Omri Mendels)

19:20-19:35 - 8 Ways to become a more impactful data scientist (Yair Mazor)

19:35-19:50 - Creating Beautiful Interactive Reports Automatically in HTML (Omri Allouche)

19:50-20:05 - Explainable Machine Learning - How to explain model predictions to customers (Erez Agami)

20:05-20:15 - Oversampling Techniques and Applications in Python (Liron Hayman)

20:05-20:25 - BREAK

20:30-20:45 - What metrics should be used for evaluating a model on an imbalanced data set? (Shir Meir Lador)

20:45-21:00 - Deriving Confidence in Regression Models (Shahar Harel)

21:00-21:15 - Semantic Segmentation for background removal (Gidi Shperber)

21:15-21:30 Deploying Keras models with Gitlab CI and Heroku (Alon Burg)

21:30-21:45 - Presenting ways for cooporation of seperate entities to create a win-win AI solution (Uri Yerushalmi)

Abstracts:

Clustering before labeling / Uri Goren, Yahoo - Document classification algorithms requires labels, and tagging may be costly. We will demonstrate a simple method for saving money by applying agglomerative clustering techniques on the raw data.

Creating Beautiful Interactive Reports Automatically in HTML / Omri Allouche, Gong.io -Presenting analysis results in a clear and inviting way is often as important as the analysis details itself. In the talk I'll show how to create beautiful interactive reports that load on any device (including mobile devices) without the need for software installation, by using the Jinja2

Explainable Machine Learning - How to explain model predictions to customers / Erez Agami, Salesforce - One of the challenges encountered in many machine learning projects is how to make the customers trust the predictions. Providing an explanation why the model gave certain predictions can help build trust. We will talk about how we dealt with the problem in some of our projects in Salesforce and what lessons we learned.

Oversampling Techniques and Applications in Python / Liron Hayman, Intuit - Classification models typically see improvements when classes are artificially made balanced. We will cover the SMOTE and ADASYN methods and their variations. Specifically, we will show the performance of these methods through examples using the imbalanced-learn Python library developed by Lemaitre, Nogueira and Aridas.

What metrics should be used for evaluating a model on an imbalanced data set? / Shir Meir Lador, BlueVine - The subject of model performance metrics can be somehow confusing, specifically when the data set is imbalanced (as happens so often in our usual problems). In order to clarify things we will review a few simple examples of an imbalanced data sets with the different type of metrics and see which reflects more correctly the model performance — ROC curve metrics — TPR and FPR or precision or recall and which metric should we use for the different possible scenarios for an imbalanced data set.

Deriving Confidence in Regression Models
/ Shahar Harel, SparkBeyond - Predicting confidence intervals for regression is of major importance to many applications. In this talk I will describe the main approaches for this topic, present our own insights on improving bagging based quantile regressors, show comparative results on a large set of benchmark obtained on an automatic machine learning environment without any manual tuning and discuss some open issues and conjectures.

Semantic Segmentation for background removal / Gidi Shperber,FastScience - Training a semantic segmentation deep learning model to remove background from images.

Deploying Keras models with Gitlab CI and Heroku / Alon Burg , FastScience - Deploying Keras models with Gitlab CI and Heroku or Client side KerasJS

Presenting ways for cooporation of seperate entities to create a win-win AI solution / Uri Yerushalmi, Dopamine.ai - Frequently, different entities/groups have the knowledge and data to create an AI based solution that can can be very helpful to one of the group. we will present a use case from an algo trading groups joining forces, using a scoring algo, to create a new AI solution.