This repository contains a comprehensive collection of academic projects and research completed during my studies at NOVA Information Management School, as well as additional certificates obtained outside of university.
Note: Due to company arrangements, source code from academic projects is not included as it may contain confidential information. However, the reports and presentations provide a clear overview of the business context, methodology, and results.
Analyses and models were implemented in Python, combining Jupyter notebooks for EDA and preprocessing with structured scripts for pipelines and deployment.
- /certificates – Certificates from courses and trainings (e.g. Deep Learning, NLP, Celonis).
- /master-projects – Academic projects completed with industry partners and in coursework.
- /master-thesis – Master’s thesis and supporting material.
- Built AI-driven monthly sales forecasts for 14 product groups to enhance pricing strategies and reduce inventory costs for Siemens' Smart Infrastructure Division.
- Implemented multiple recommendation strategies including RF analysis, "Did You Forget?" reminders, and similar product suggestions to enhance customer engagement. Awarded as second best project in the class.
- Developed predictive models to forecast business process outcomes and optimize decision-making through machine learning classification techniques.
- Developed a Business Intelligence system for a retail chain. Designed a data warehouse, dimensional model, and interactive dashboards for sales analysis in Power BI.
Developed a full-stack language learning application as a web app and mobile app currently available for testing on Android and iOS. Built with React, including database integration for user data, external data sources, and API-based features.
Title: Emotion Recognition with EEG Data: Leveraging Pretrained Convolutional Neural Networks
Abstract:
Emotion recognition using electroencephalogram (EEG) signals is a key area in brain–
computer interfaces with valuable applications in domains such as mental health and clinical
diagnostics. To address the challenge of limited data availability in EEG-based emotion
recognition, convolutional neural networks (CNNs) pretrained on image data can be leveraged
by transforming EEG signals into two-dimensional spectrograms using the Short-Time Fourier
Transform (STFT). However, existing literature lacks a systematic comparison of preprocessing
methods, feature representations, and CNN adaptation strategies, leaving open questions
about best practices in this domain. This study addresses these gaps by systematically
evaluating multiple EEG preprocessing configurations and comparing two types of 2D
representations: spectrograms and stacked spectrograms, the latter of which has not yet been
explored in the context of emotion recognition. The evaluation framework includes four
pretrained CNN architectures (MobileNetV2, ResNet50, InceptionV3, and DenseNet121) and
compares two approaches to leveraging these models: transfer learning through fine-tuning
and feature extraction with an external SVM classifier. All experiments were conducted on the
SEED dataset, which contains EEG recordings labeled with positive, neutral, and negative
emotional states. The results show that the identified preprocessing configurations
substantially improved validation accuracy. While stacked spectrograms demonstrated
potential for capturing multichannel EEG dynamics, they were ultimately outperformed by
individual spectrograms due to reduced training data availability. Feature extraction
surpassed transfer learning in both validation accuracy and computational efficiency. The best
overall performance was achieved using InceptionV3 and DenseNet121 in combination with
an SVM classifier, reaching a competitive average validation accuracy of 86.67%, and
significantly outperforming a baseline SVM model trained on traditional PSD features. The
findings offer valuable guidance for future work on EEG-based emotion recognition,
highlighting the effectiveness of feature extraction with pretrained CNNs and demonstrating
the value of adapting visual deep learning models to neurophysiological data..
Programming Languages: Python, R, SQL, DAX
Statistical Analysis: Panel data analysis, econometrics (FE, RE, OLS), hypothesis testing
Machine Learning: Classification (LogReg, Random Forest, ..), clustering (K-means, DBSCAN, ..), time series forecasting (Prophet, AutoML, ..)
Deep Learning: CNNs (ResNet, Inception, DenseNet, ..), TensorFlow, PyTorch, transfer learning
Data Mining: Segmentation, pattern recognition, feature engineering, preprocessing
Business Intelligence: Data warehousing, dimensional modeling, ETL, interactive dashboards (Power BI)
Business Process Management: Value-added analysis, waste identification, optimization frameworks (Pareto, PICK)
Tools & Platforms: Jupyter Notebooks, Git, TensorFlow, scikit-learn, pandas, numpy, MongoDB, Neo4j, Microsoft Fabric, Azure
Software & App Development: React, API integrations, database-backed user data, Firebase, Capacitor, web and mobile app development
Email: [david.psiuk@web.de]