I am a data scientist who focuses on solving complex business pain points. My technical background has prepared me to tackle these challenges using an analytical approach. This complements my adaptability and communication skills which have developed from living in and traveling to over 30 countries.
September 2020 - December 2020
March 2017 - June 2020
August 2012 - July 2014
September 2019 - December 2019
July 2019 - Current
pierre.m.winter@gmail.com
For the full story check out this blog post
Have you ever received a benefit package from an employer and felt that it was too complicated to keep track of your spending amounts? Or can you recall a time when the benefit reimbursement process was fast and free of hassle? To address these increasingly common problems, the startup Benepass is combining employer benefits and stipends on a single payment card in order to simplify and modernize the benefits process. Benepass has an important pain point: When an employee from Company X goes to a store and purchases something, how can Benepass use the transaction data to assign it to a benefit category that the employee is eligible for?
This project was performed as a 4-week independent consulting project, by which I was introduced to Benepass through the Insight Data Science Program. The data given to me by Benepass has a mixture of numerical, text, and categorical features (green), data which I use for post-processing (yellow), as well as multiple labels (blue).
In general, I overcame 3 significant challenges with this dataset: there were only 3 usable features, the dataset was very small (350 rows), and the labels were highly imbalanced. To manipulate the data and train a machine learning model I made use of the SMOTE, Pandas, and Scikit-learn libraries in Python.
I successfully created a multi-class classification model (called WhichBenefit) which produces probability scores for each class, where in most cases the predicted class can be taken as the class with the highest probability. The model probabilities are then filtered through a selection of benefits that a given employee is eligible for. I implemented this data product into a production-ready Python package which is now implemented in Benepass' live workflow.
WhichBenefit adds business value by automating the classification of transactions into benefit categories and by significantly increasing the performance over the current approach. It also allows for new companies and transactions, which will help Benepass scale up their production with minimal problems. As a Python package, it is also adaptable such that the machine learning model can be retrained and quickly replace the current model in the transaction pipeline.
Humanitarian aid organizations are unable to provide accurate cash-based assistance in conflict regions. This is often due to the lack of pricing information for basic necessities found in local markets. How can we estimate missing prices of market goods in Syria such that families in need are given an adequate amount of aid?
This project was performed as part of the Hack4Good Program hosted by the Analytics Club at ETH Zürich. My team consisted of 3 other university students of varying backgrounds, all of whom I had met for the first time. We were given 8 weeks to provide a data analytics solution for our partner NGO, IMPACT Initiatives, who provided us with a sparse and unstructured dataset containing monthly market prices for regions across Syria.
Of the methods tested, we found two which proved to be most effective and which showed the smallest nRMSE error after cross-validation: Adapted k-Nearest Neighbors (a-kNN) and Multiple Imputation by Chained Equations (MICE). These approaches group the market price data in a hierarchical way such that time-series analysis was performed before geographical grouping.
The solution we propose is simple and generalizable such that it can be implemented with minimal costs and in other regions of conflict. We provided IMPACT Initiatives with software written in Python and R as well as a more structured and fully imputed dataset. They have communicated that they want to use this in their production level operations in order to better inform the amount of assistance needed by these families.
Calculating the chemical energy of molecules in an accurate way is extremely time consuming, even when using state-of-the-art methods on powerful computers. This can be attributed to the poor scaling of computation time with respect to the number of atoms in the system. How can we obtain accurate chemical energies while reducing the number of expensive computations?
This project was completed as part of my independent doctoral research at ETH Zürich. I was fortunate to have access to the academic computing cluster which allowed me to perform these expensive chemistry calculations in a reasonable amount of time.
In order to reduce the number of computations of the chemical energy (as well as its first and second derivatives), my solution included only calculating the energy and its derivatives at a few points along a multidimensional reaction pathway and using a regression technique to interpolate new points. This was done with a kernel-based machine learning approach called Gaussian Process Regression, for which I developed a modular Python code that will soon be released as open-source.
Using Gaussian Process Regression, I was able to drastically reduce the number of expensive calculations needed and at the same time I obtained interesting scientific results for my research which would otherwise be unattainable. The solution is generalizable to most chemical systems and since it is a Bayesian method, it does not require the manual tuning of function parameters. I can also use it to obtain an explicit measure of the error when predicting new energies.
Whether competing with friends recreationally or placing bets professionally, the growth of fantasy sports has led to the use of data analytics to predict team and player performance. This can be challenging for complex team sports, especially in fast-paced competitive leagues. Using simple metrics, how can we predict a top-performing fantasy soccer team in the English Premier League (EPL)?
This is a personal project inspired by my desire to improve my own fantasy soccer scores when competing with friends. I made use of open-source software such as Python, Jupyter, and Matplotlib. The fantasy football data was obtained via requests from the official Fantasy Premier League API.
As part of my solution I created a return on investment (ROI) measure for all players and teams in the EPL. The ROI relates the total fantasy points earned by a player over the course of the 38-week season to their purchase value and their number of minutes played. From this analysis I was able to easily select a mixture of high-ROI and high point-producing players within my budget constraints.
This project was initially performed during the 2018/2019 EPL season, but the solution I propose can be easily implemented for future seasons. The Jupyter notebook I created also comes with visualizations which categorize the most valuable players by their positions. This functionality allows for more manager flexibility such that should a player become injured, a replacement player can easily be found for that corresponding position.
As a job-seeker, searching for open positions can be a daunting task given the plethora of career decisions that need to be made. Finding free time to manually scroll through job postings also presents constraints which could add to the stress of searching for the right position. Can publicly available job postings be aggregated and visualized in an automated way?
This is a personal project inspired by my desire to help myself and family members search for relevant and highly-rated positions and companies by geographic location. I made use of open-source software such as Python, Requests, BeautifulSoup, and Tableau.
My solution was to build an easy-to-use web scraper that performs HTTP requests to job posting websites and parses the publicly available job data. In this particular implementation I wanted to find up to 500 job listings on Indeed for full-time data scientist positions within 50 miles of both Los Angeles and San Francisco. The data containing position, company, city, job posting date, and job rating was parsed and aggregated into a .csv file. This data was then easily visualized using Tableau to get a good understanding of where the jobs are located with zip code granularity as well as the spread of job ratings.
As is, this project can currently be used to search for jobs in the U.S. on Indeed's website and visualize them in a matter of minutes. This helps accelerate the job search process by automating the extraction of job market data. Extension to other countries and other job posting websites is simple and a GUI to do this is in progress.
The average number of friends that a person can maintain strong relationships with has been determined to be 150 or less (Dunbar's number). The size of friend groups may also vary among different countries and cultures. Can geographical differences in the number of friends be quantified using social media data?
This is a personal project based on my interest in social data science and understanding how sociology impacts human relationships. I made use of open-source software such as R and rtweet to obtain public Twitter data and I performed data aggregation and visualization in RStudio.
My solution was to obtain geogrpahical Twitter timeline data using the rtweet package and to quantify differences in the number of friends between Western and Central Europe. Once the data was cleaned and aggregated, I defined a method for counting the number of friends a Twitter user has and applied variable transformation to deal with the highly skewed data. A permutation test was then performed to compare the two European populations and it was found that they exhibit a statistically significant difference in their average number of friends.
This project is useful for understanding differences in Twitter data between two geographically different populations. This can be helpful for understanding cultural differences and possibly to aid in the marketing and advertisement industries. Extension to other countries and a more robust method for counting number of friends is currently in progress.
