Asterios Stergioudis

Athens University of Economics and Business

Postgraduate Specialization Program, Big Data & Business Analytics

N/A – Present

Aristotle University of Thessaloniki (AUTH)

Degree, Informatics (orientation on Information Systems)

N/A – Present

University of Surrey

MSc, Management Information Systems

N/A – Present

Kaizen Gaming

Data Science Team Lead

June 1, 2023 – Present

Thessaloniki, Central Macedonia, Greece · Hybrid

VieConnect SAS

Data Scientist - Consultant

November 1, 2017 – May 1, 2023

Remote (Thessaloniki, GR - Toulouse, FR)

Citrix

Staff Data Scientist

October 1, 2017 – March 1, 2023

Remote (Patras - Santa Clara, CA - Fort Lauderdale, FL)

baresquare

Data Scientist

April 1, 2014 – October 1, 2017

Thessaloniki

Kaggle

Data Science Competitor (Expert tier)

June 1, 2013 – December 1, 2018

PwC Greece

Risk Assurance Services Consultant

June 1, 2012 – April 1, 2014

Athens

Pap Corp.

Front Office Manager (06-10/11) & Receptionist

March 1, 2011 – March 1, 2012

Chalkidiki

EuroLeague 2017 final-4 chatbot

December 1, 2016 – June 1, 2017

Came up with and developed the official EuroLeague 2017 final-4 chatbot using python, wit.ai, Redis. Gave keynote speech on "Conversations 2017: an international workshop on chatbot research and design": https://conversations2017.wordpress.com/about/

Kaggle - Walmart Trip Type Classification

October 1, 2015 – December 1, 2015

For this competition, the task was to categorize shopping trip types based on the items that customers purchased. Improving the science behind trip type classification will help Walmart refine their segmentation process. After extensive feature engineering and selection, the final model was a 2nd level model built on Gradient Boosted Trees, Random Forests, K-Nearest Neighbors and Deep Neural Networks CV probabilities. Final probabilities were pushed to the extremes for certain cases in order to minimize logloss. Programming language used: Python (pandas, scikit, XGBoost, Lasagne). Finished 55th among 1047 teams (top 6%).

Kaggle - Otto Group Product Classification Challenge

March 1, 2015 – May 1, 2015

For this competition, the provided dataset included 93 features for more than 200,000 products. The objective was to build a predictive model which was able to distinguish between the main Otto product categories. After the necessary feature engineering, the final model was an ensemble of Deep Neural Networks and Tree-based models. Programming language used: Python (scikit, XGBoost, Lasagne). Finished 284th among 3514 teams (top 8%).

Kaggle - Microsoft Malware Classification Challenge

February 1, 2015 – April 1, 2015

The goal of the competition was to group variants of malware files (~400 GB) into their respective families. The first step was to create features from raw hex dumps and .asm files. I extracted more than 140K features that referred to counts of bytes and assembly instructions (specific 1, 2 and 3-grams). Then, a few features were kept based on their importance calculated from GBMs (multiclass model and separately per-class). Several GBMs were trained on the resulting datasets (scikit and XGBoost). The final solution was a weighted average of 6 models. Probabilities were pushed to the extremes so as to minimize logloss. Achieved cross-validated accuracy of 99.7%. Programming language used: Python (scikit, XGBoost). Finished 19th among 377 teams (top 5%).

Evaluation of Dimension Reduction Techniques for Text Classification and Sentiment Detection in Voice Recordings

July 1, 2014 – January 1, 2015

Thesis project for the fulfillment of the Specialization Program of Big Data & Business Analytics at Athens University of Economics and Business. Eleven dimensionality reduction algorithms were evaluated and compared based on the F1-score achieved in the classification of text and voice recordings datasets. The results indicate that Singular Value Decomposition was usually the best performing method while other methods such as Gaussian Random Projections could be a viable alternative that is easily scalable to very large datasets and leads to small or no sacrifices on classification performance.

Kaggle - Greek Media Monitoring Multilabel Classification

June 1, 2014 – July 1, 2014

The goal of the competition was to classify documents into on or more topics. The documents were given in tf-idf format. Data were amended with SVD components, LDA topics and the local CV predictions were also appended to the original dataset. The best model was a blend of various linear models (logistic regression, linear SVMs, Linear Discriminant Analysis) trained on different datasets using the Binary Relevance transformation. Moreover, separate models were trained on document clusters. Lastly, regression models trained after 'Principal Label Space Transformation' also boosted the final prediction. Programming language used: Python. Finished 16th among 121 teams.

Kaggle - See Click Predict Fix

November 1, 2013 – Present

The purpose of the competition was "to quantify and predict how people will react to a specific 311 issue. Being able to predict the most pressing 311 topics will allow governments to focus their efforts on fixing the most important problems." The issue descriptions and summaries were transformed with tf-idf and then with SVD. New features were created (e.g. distance of issue from city centre) and two types of models were trained, namely Random Forests and Stochastic Gradient Boosting. The models were combined which improved the prediction accuracy significantly. Programming languages/tools used: Python, R, ACL. Finished 83rd among 533 teams.

Kaggle - StumbleUpon Evergreen Classification Challenge

August 1, 2013 – October 1, 2013

The purpose of the project was to "to build a classifier which will evaluate a large set of URLs and label them as either evergreen or ephemeral" (i.e. whether a website maintains a timeless quality and can be recommended to users long after it is discovered). Website text was transformed with tf-idf, features-topics obtained from Latent Dirichlet Allocation were appended to the dataset and trained logistic regression models. Also, trained a Gradient Boosting Machine model on most important features of the text. Finally, the above models were combined using simple average. Programming languages used: R, Python. Finished 31st among 625 teams.

User acceptance of free and open source software: an empirical study comparing Linux and Windows in Greece

May 1, 2009 – August 1, 2009

Dissertation project for the fulfillment of the MSc in Management Information Systems at University of Surrey. After reviewing the existing user acceptance models, a new model was proposed based on UTAUT. Moreover, the sociocultural factors affecting the study were taken into consideration using Hofstede's cultural dimensions. The proposed model was then tested using questionnaires and the responses were analyzed with SPSS. The model proved to be more successful than its predecessors in predicting the acceptance of free and open source software by individuals. Based on the findings, proposals were made on the right way to introduce such software in a business setting while minimizing the resistance shown by the users. The project's grade was 83%.

Performance measurement of wireless local area networks

September 1, 2007 – January 1, 2008

Thesis project for the fulfillment of the degree in Informatics at Aristotle University of Thessaloniki. Working in a group of 3 persons, the performance of the protocols 802.11b/g was tested. Their performance was assessed on various scenarios with and without obstacles between the transmitter and the receiver, at various transmission speeds, with different modulations (CCK and OFDM) and various packet sizes using Commview for WiFi software application. Based on the findings, proposals were made on the best use of wireless LANs in standard business settings. The project's grade was 10/10.

SAP TERP1e - SAP Solution Architect

SAP S.A.

June 24, 2026 – Present