Projects
Data Science Portfolio
Stars
70
Data Science Portfolio
This portfolio consists of several projects illustrating the work I have done in order to further develop my data science skills.
Table of Contents
Projects
BERTopic: Topic Modeling with BERT
BERTopic is a novel topic modeling technique that leverages BERT embeddings and c-TF-IDF to create dense clusters allowing for easily interpretable topics whilst keeping important words in the topic descriptions.
KeyBERT: Keyword Extraction with BERT
KeyBERT is a minimal and easy-to-use keyword extraction technique that leverages BERT embeddings to create keywords and keyphrases that are most similar to a document.
PolyFuzz: A framework for Fuzzy String Matching
PolyFuzz performs fuzzy string matching, string grouping, and contains extensive evaluation functions. PolyFuzz is meant to bring fuzzy string matching techniques together within a single framework.
SoAn: Analyzing WhatsApp Messages
Repository | Notebook | nbviewer
- Created a package that allows in-depth analyses on whatsapp conversations
- Analyses were initially done on whatsapp messages between me and my fianciee to surprise her with on our wedding
- Visualizations were done in such a way that it would make sense for someone not familiar with data science
- Methods: Sentiment Analysis, TF-IDF, Topic Modeling, Wordclouds, etc.
ReinLife: Artificial Life with Reinforcement Learning
- Using Reinforcement Learning, entities learn to survive, reproduce, and make sure to maximize the fitness of their kin.
- Implemented algorithms: DQN, PER-DQN, D3QN, PER-D3QN, and PPO
VLAC: Vectors of Locally Aggregated Concepts
- It leverages clusters of word embeddings (i.e., concepts) to create features from a collection of documents
allowing for classification of documents - Inspiration was drawn from VLAD, which is a feature generation method for image classification
- The article was published in ECML-PKDD 2019
c-TF-IDF: Class-based TF-IDF
- Created an algorithm allowing for TF-IDF to work on a class-level
- Thereby creating categorical vectors that can be used for:
- Informative Words per Class: Which words make a class stand-out compared to all others?
- Class Reduction: Using c-TF-IDF to reduce the number of classes
- Semi-supervised Modeling: Predicting the class of unseen documents using only cosine similarity and c-TF-IDF
Reviewer: Character Popularity
Repository | Notebook | nbviewer
Reviewer can be used to scrape user reviews from IMDB, generate word clouds based on a custom class-based TF-IDF, and extract popular characters/actors from reviews. Methods:
- Named Entity Recognition
- Sentiment Analysis
- Scraper
- BERT
Board Game Exploration
- Created an application for exploring board game matches that I tracked over the last year
- The application was created for two reasons:
- First, I wanted to surprise my wife with this application as we played mostly together
- Second, the data is relatively simple (5 columns) and small (~300 rows) and I wanted
to demonstrate the possibilities of analyses with simple data
- Dashboard was created with streamlit and the deployment of the application was through Heroku
Statistically Generated Disney Tournament
- Created a tournament game for a friend of mine
- Tournament brackets are generated based on a seed score calculated through scraping data from IMDB and RottenTomatoes
- Methods: BeautifulSoup
Optimizing Emt Routes
- Project for the course Business Analytics in the master
- Optimization of managers visiting a set of cities
- Total of 133 cities, max distance 400km with time and capacity constraints
- Thus, a vehicle routing problem
- Methods: Integer Linear Programming, Tabu Search, Simmulated Annealing, Ant Colony Optimization, Python
Pothole Detection
Repository | Notebook| nbviewer
- Image classification of potholes in roads
- Explored different activation functions (Relu, Leaky Relu, Swish, Mish)
- Used EfficientNetB0 and applied transfer learning to quickly get a high accuracy (> 95%)
- Unfreezing certain layers helps with fine tuning the model
- Methods: Deep Learning, TF/Keras, EfficientNetB0
Exploring Explainable ML
Repository | Notebook | nbviewer | Medium
- Explored several methods for opening the black boxes that are tree-based prediction models
- Models included PDP, LIME, and SHAP
Deploying a Machine Learning Model
- Developed a set of steps necessary to quickly develop your machine learning model
- Used a combination of FastAPI, Uvicorn and Gunicorn to lay the foundation of the API
- The repository contains all code necessary (including dockerfile)
Retro Games Reinforcement Learning
- An overview of methods for creating state-of-the-art RL-algorithms
- Makes use of Gym, Retro-gym, Procgen, and Stable-baselines
Statistical Cross-Validation Techniques
Repository | Notebook | nbviewer | Medium
- Dived into several more in-depth techniques for validating a model
- Statistical methods were explored for comparing models including the Wilcoxon signed-rank test, McNemar's test, 5x2CV paired t-test and 5x2CV combined F test
Cluster Analysis: Create, Visualize and Interpret CustomerSegments
Repository | Notebook | nbviewer | Medium
- Explored several methods for creating customer segments; k-Means (Cosine & Euclidean) vs. DBSCAN
- Applied PCA and t-SNE for the 3 dimensional exploration of clusters
- Used variance between averages of clusters per variable to detect important differences between clusters
Exploring Advanced Feature Engineering Techniques
Repository | Notebook | nbviewer | Medium
- Several methods are described for advanced feature engineering including:
- Resampling Imbalanced Data using SMOTE
- Creating New Features with Deep Feature Synthesis
- Handling Missing Values with the Iterative Imputer and CatBoost
- Outlier Detection with IsolationForest
Predicting and Optimizing Auction Prices
- Data received from an auction house and therefore not made public
- Prediction of value at which an item will be sold to be used as an objective measure
- Optimize starting price such that predicted value will be as high as possible
- Methods: Classification (KNN, LightGBM, RF, XGBoost, etc.), LOO-CV, Genetic Algorithms, Python
Statistical Analysis using the Hurdle Model
- Used Apple Store data to analyze which business model aspects influence performance of mobile games
- Two groups were identified and compared, namley early and later entrants of the market
- The impact of entry timing and the use of technological innovation was analyzed on performance
- Methods: Zero-Inflated Negative Binomial Regression, Hurdle model, Python, R
Predict and optimize demand
- Part of the course Data-Driven SCM
- Optimizing order quantity based on predicted demand using machine learning methods
- Simulation model was created to check performance of method calculating expected demand
- Additional weather features were included
- Methods: Regression (XGBoost, LightGBM and CatBoost), Bayesian Optimization (Skopt)
Analyzing Google Takeout Data
- Analyzing my own data provided by Google Takeout
- Location data, browser search history, mail data, etc.
- Code to analyze browser history is included
- Code to create animation will follow
Cars Dashboard
- Created a dashboard for the cars dataset using Python, HTML and CSS
- It allows for several crossfilters (see below)
Qwixx Visualization
- Visualized 16 games I played with my wife
- The result is a heatmap in using the scorecard of Qwixx
- All rights belong to Qwixx (Nuernberger-Spielkarten-Verlag GmbH)
- The scorecard used for this visualization was retrieved from:
Academic Journey Visualization
- A visualization of my 9! year academic journey
- My grades and average across six educational programs versus the average of all other students in the same classes.
- Unfortunately, no data on the students average grade was available for classes in my bachelor.
- Grades range from 0 to 10 with 6 being a passing grade.
- Lastly, Cum Laude is the highest academic distinction for the masters program I followed.
Neural Style Transfer
- For the course deep learning I worked on a paper researching optimal selection of hidden layers to create the most appealing images in neural style transfer while speeding up the process of optimization
- Code is not yet provided as I used most of the following code from here and I merely explored different layers
Predicting Housing Prices
- Project for the course Machine Learning
- Participation of the kaggle competition House Prices: Advanced Regression Techniques
- We were graded on leaderboard scores and I scored 1st of the class with position 33th of the leaderboard
- I used a weighted average of XGBoost, Lasso, ElasticNet, Ridge and Gradient Boosting Regressor
- The focus of this project was mostly on feature engineering
Analyzing FitBit Data
Repository | Github | nbviewer
- My first data science project
- I looked at improving the quality of FitBit's heartrate measurement as I realized it wasn't all that accurate
- Used a simple 10-fold CV with regression techniques to fill missing heart rates
- There are many things that I would do differently now like using other validation techniques that are not as sensitive to overfitting on timeseries
Contact
If you are looking to contact me personally, please do so via E-mail or Linkedin:
- E-mail: [email protected]