Data Science Challenges

Data science challenges are competitive events where individuals or teams work on real-world problems using data analysis and machine learning techniques. Below are challenges I participated in together with a team:


BIRDCLEF+ 2025

by Cornell Lab of Ornithology, USA
May 2025

Python Scikit-Learn NumPy PyTorch

Learn more about the challenge

Overview
The task was to develop a sound classification model that determines which of 206 different animal species are present in a soundclip, taken in a wild life reservoir. Such models are used to quantifiy the success of environmental restoration programs. The idea is that a high biodiversity and the presence of certain key species indicates a healthy environment and therefore the success of the restoration program.

Approach
I used the PANN audio model, which was pretrained on 5000 hours of sound and predicts 527 sound classes. With the help of this model, I created embeddings of the sound training data from the BIRDCLEF challenge. The idea is that this pretrained model is already capable of finding useful features in sound data, and can therefore be used as a feature extractor. To achieve this, I removed the output layer of the PANN, and used the output of the layer before as the embeddings. Then I trained another neural network model with far fewer parameters than PANN to map from those embeddings to the probabilities of the 206 animal classes of the BIRDCLEF challenge. Unfortunately, the challenge hosts required a minimal runtime of the approach in their cloud environment, which I could not meet with my approach, which is why I was unable to submit and achieve a ranking.


Richter’s Predictor: Modeling Earthquake Damage

by Driven Data, USA
March – May 2025

Python Scikit-Learn NumPy XGBoost SHAP

🏆 1007th Place / 2487 Submissions

Learn more about the challenge

Overview
The task was to develop a model that predicts the earthquake damage (low, medium, almost complete destruction) from features such as the geo location, and construction engineering features, such as the height of the building and the type of superstructure material. The data consisted of about 260,000 houses damaged in the Gorkha earthquake that hit Nepal in 2015. Note that this was a practice challenge with no price pool.

Approach
We used one-hot encoding and target encoding for the categorical features, and predicted the damage with an XGBoost model by threating this ordinal problem as a classification. As the objective function, we used multi-class classification – soft probability output. The final score was reported with the micro-averaged F1 score. Finally, we used SHAP to extract feature importances to find insights about how to build houses so that they get less earthquake damage. Our SHAP feature importance analysis showed that building age is the most important predictor of earthquake damage (besides proximity to the earthquake epicenter): newer buildings tend to suffer less, while older ones are more likely to be heavily damaged. Structural materials play a major role too. Buildings made of mud mortar stone, mud mortar brick, or adobe mud are associated with higher damage, which matches engineering knowledge—they are brittle, lack reinforcement, and perform poorly under lateral seismic forces. In contrast, cement mortar brick offers better protection due to its higher strength and cohesion.


Watt’s up? Synthetic Data for Buildings

by TU Vienna, Austria
22 – 23 February 2025

Python Scikit-Learn NumPy XGBoost

🏆 3rd Place / 8 Teams – 80€ Prize

Learn more about the challenge

Overview
The task was to develop a model that can input timecourse data of electricity consumption and output a synthetic version of this dataset that ensures the privacy of the individuals. This is important because electricity consumption data are sensitive and cannot be easily shared. However, widespread availability of such datasets could be very beneficial for model development, allowing insights that help save energy.

Approach
We chose an approach that samples pieces from the actual data, reassembles them, and adds Gaussian noise on top. This produced instances that closely resembled the real data in terms of statistical properties but were different enough to ensure they cannot be linked to individual examples of the real data.


Autoimmune Disease Machine Learning Challenge

by Eric and Wendy Schmidt Center at the Broad Institute, USA
Part 1: Oktober 2024 – February 2025

Part 2: November 2024 – March 2025

Part 3: December 2024 – April 2025

Python PyTorch Google Colab

Part 1: 🏆 2th Place / 75 submissions / 7800 participants – 2400$ Prize

Part 2: 🏆 6th Place / 11 submissions / 8200 participants – 700$ Prize

Part 3: 🏆 still ongoing

Learn more about the challenge

Overview
Autoimmune diseases occur when the immune system mistakenly attacks healthy cells, disrupting the body’s natural defense mechanisms. These diseases affect 50 million people in the U.S., with cases rising globally. One of the most prevalent autoimmune diseases is inflammatory bowel disease (IBD). To better understand and treat IBD, researchers sequence the RNA (transcriptome) of individual affected cells. However, this process is highly complex and expensive, making a predictive model for transcriptomics a valuable tool in advancing research. This challenge consisted of three independent sub-challenges, each addressing a distinct aspect of this broader research problem:

Approach
For task 1, we used the ResNet50 model implemented in PyTorch to directly
map from the input images to the 460 genes (end-to-end deep learning approach).

For task 2, we took our trained ResNet50 model and its predictions, which served as the input to a standard neural network, which mapped from the 460 predicted genes to all genes of the cell.

For task 3, we submitted many different approaches. Notable examples include (i) determining the AUC value for all features, and sorting them according to that, and (ii) predicting the cells with dysplasia with the help of a neural network model, for which we extracted feature importances with the Integrated Gradients method from the Captum library. Integrated Gradients is a feature importance algorithm for neural networks.


LEC Data Challenge 2023

by LEC Gmbh, Austria
July – August 2023

Python Perl XGBoost GMMs Scikit-Learn

🏆 1st Place / 46 Teams – 4,000€ Prize

Learn more about the challenge See the winners on LinkedIn

Overview
The task was to develop a machine learning solution to identify cylinder-specific events in time-series data from large combustion engines.

Approach
We trained an XGBoost algorithm to produce a model that predicted, for each
time point, which type of issue was present, if any. As input, we used slices
of the time course (nearest values in time). Because issues always covered
larger continuous time windows in the training data, we afterward
algorithmically identified regions in the XGBoost predictions where many issues were detected and unified them into a continuous stretch with one issue type.


Trigger Detection Challenge 2023

by Bauhaus-University Weimar, Germany
April – May 2023

Python Perl LLMs

Learn more about the challenge

Overview
A multi-label classification challenge to assign trigger warning labels (such as violence and blood) to fanfiction documents based on their content.
As I still was a beginner, I honestly was a bit overwhelmed and left the challenge before it ended. However, I learned a lot about the workflow on how to tackle such tasks.