Year of Graduation
Advanced feature engineering and dimensionality reduction for structural connectome classification
In this work we investigated Autism Spectrum Disorder vs Typically Developing classification task based on structural connectomes. Using combination of different weighting schemes, topological normalizations and graph metrics we constructed about 500 feature sets and tested them using selected classifiers and cross-validation techniques. We found features obtained with combination of weighting by distance and topolgical normalization which achieved 0.8 ROC AUC score. It is comparable with results described in recent studies. We also tried dimensionality reduction on the best obtained features, but didn't find simple geometry in our data.