Andrea Alberti

This study presents two machine learning models (MLP_Ensemble5 and MLP_Ensemble2) aimed at enhancing heart disease detection from heart sound recordings using ensemble techniques. The project involved data preprocessing, feature extraction (MFCCs, Chroma STFT, etc.), and feature selection using data from the PASCAL CHSC2011 challenge.

This study investigated the complex relationships between symptoms and diseases using network analysis techniques on a large dataset. A bipartite graph was created and analyzed using metrics like degree distribution, Hidalgo's method of reflections, betweenness centrality, and community detection. Novel features derived from these network metrics were used to train predictive models (Logistic Regression, Random Forest, MLP).

This project involved a comprehensive analysis of an Amazon book review dataset using big data tools (Hadoop, Spark) and data science techniques. It investigated factors influencing review helpfulness, such as review length, sentiment, and user rating. Predictive models (Random Forest, SVR, MLP) were built using Word2Vec embeddings to estimate review helpfulness.

Implemented and compared Multinomial Naïve Bayes (MNB) and Logistic Regression (LR) classifiers for identifying clickbait headlines from a dataset of 32,000 examples. The project evaluated performance under two scenarios: maximizing overall accuracy and minimizing the False Positive Rate (FPR). Utilized Bag-of-Words feature representation with different vocabulary sizes.

This project experimentally assessed the impact of DNS reflection and amplification attacks within a controlled local network environment. It explored how different DNS request types affect amplification factors and analyzed the consequences on the target system's latency and the DNS server's resource utilization (CPU, RAM). Attacks were simulated using custom Scapy scripts with IP address spoofing.

This project developed models to classify images into 15 different cake categories. It compared two main approaches: using a Multi-Layer Perceptron (MLP) with low-level image features (Color Histogram, Edge Direction Histogram, Co-occurrence Matrix), and using an MLP with neural features extracted from a pre-trained CNN (PVMLNet). Transfer learning was also explored by adapting the pre-trained CNN.

Developed two image processing programs. The first focuses on image binarization using a custom histogram-based thresholding technique, offering automatic and manual tuning. The second program detects vanishing points and lines using techniques like the Canny edge detector, probabilistic Hough transform, and the RANSAC algorithm. Both programs include command-line interfaces.

This project developed and compared Multinomial Naive Bayes (MNB) and Logistic Regression (LR) classifiers for predicting sentiment (positive/negative) in movie reviews, using the dataset from Maas et al. The study explored the impact of vocabulary size and preprocessing techniques like stopword removal and stemming on classifier performance. Analysis included evaluating accuracy, overfitting trends, and identifying the most impactful words.