Andrea Alberti

Designed prevention and clinical support ensembles for early cardiac screening on the Dangerous Heartbeat Dataset (CHSC2011). Heart sounds were resampled at 4 kHz, segmented into 1-second windows, described with MFCC, chroma, spectral and temporal descriptors, and reduced from 338 to 41 features via Spearman-based filters. The prevention ensemble keeps false normals under control (ROC-AUC 0.96, TPR 43.4% at 1% FPR) while the five-class support ensemble delivers macro F1 81.6 with per-class risk analysis and SHAP explanations.

Mapped 773 diseases and 377 symptoms into a bipartite network to engineer graph-aware features for diagnosis. Method of Reflections, Disease/Symptom Influence indices, community detection and betweenness centrality drive new descriptors that complement one-hot symptoms. Logistic Regression, Random Forest and MLP models were benchmarked; the best logistic model matches the symptom-only baseline while using fewer inputs and exposes class-level accuracy insights.

Analyzed ~3M Amazon book reviews end-to-end with a big data stack (HDFS, Spark, MongoDB) to explain and predict perceived helpfulness. Hypothesis testing quantified the role of review length, sentiment and star rating, while Word2Vec embeddings fed Random Forest, SVR and MLP regressors for score prediction. The best Random Forest model achieved MSE 0.0259 (RMSE 0.1609, R² 0.253).

Benchmarked Multinomial Naive Bayes and Logistic Regression on 32k balanced news headlines to detect clickbait. Two deployment targets were explored: maximum accuracy (97.12% test accuracy with stopwords and 8k vocabulary) and zero false positives (0% FPR, 84% accuracy, TPR 68%). Detailed error analysis highlights impactful tokens and the trade-offs introduced by bias calibration.

Recreated DNS reflection and amplification attacks in a controlled LAN to measure amplification factors, target-side latency and server resource usage. Custom Scapy scripts spoofed victims while varying query types (A, MX, NS, ANY) at 10k–50k packets/s. The study documents latency spikes above 100 ms for ANY requests, CPU saturation during amplified attacks and evaluates mitigation strategies.

Compared handcrafted descriptors and CNN-derived features for classifying 15 cake categories (1,800 images). Low-level statistics (color histogram, edge direction, co-occurrence) fed an MLP but plateaued at 31% accuracy, while PVMLNet feature maps (layer −5) coupled with an MLP achieved 90% test accuracy. Transfer learning by fine-tuning PVMLNet reached 80%, highlighting the importance of deep representations.

Delivered two computer-vision utilities: (1) a histogram-driven binarisation tool with auto/manual tuning and GUI, and (2) a vanishing point detector that chains Canny, probabilistic Hough and RANSAC (500 iterations, 5 px tolerance). The pipeline adapts thresholds from image statistics, overlays the 15 most significant lines, and documents SSIM comparisons against Otsu.

Implemented sentiment classifiers on the IMDb dataset (50k reviews) comparing Multinomial Naive Bayes and Logistic Regression. Vocabulary size, stopword removal and stemming were studied to balance accuracy and overfitting. Naive Bayes with stopwords (vocab 1k) achieved 82.6% test accuracy, while Logistic Regression reached 85.4% with minimal tuning.