Gabriele Paganelli

My master's thesis at the University of Padua studies deep clustering that can be read, not just run. VaDE (Variational Deep Embedding) combines a variational autoencoder with a Gaussian mixture to learn latent representations and clusters jointly, but its dense neural networks make both the latent dimensions and the clusters hard to interpret. The thesis adds a structured sparsity penalty to the ELBO, so that each latent dimension depends only on a small subset of the observed variables, and studies when such a sparse model remains identifiable.

We modeled the vertical distribution of soil organic carbon, nitrogen, and phosphorus across 40 Tanzanian agricultural fields using a Bayesian hierarchical model with field-level random intercepts and slopes, modeled jointly as multivariate random effects in Stan. Fields richer in organic carbon show significantly flatter depth profiles, and projpred variable selection identifies fine soil texture as the dominant within-field predictor, with management practices not significant once soil physical structure is accounted for.

This study analyzed the dynamic relationships between natural gas prices, renewable energy production, and the Italian national electricity price (PUN) over 2021–2025. Using a combination of univariate SARIMA, VAR, transfer function, and cointegration models, we identified short- and long-term interactions, highlighting the dominant role of gas in price formation, and provided predictive evidence that TFM_Gas is superior for short- to medium-term forecasting.

I developed and evaluated a sequence of Bayesian mixture models to classify high-energy neutrino events as astrophysical or atmospheric, combining spatial and energetic information. Models included joint spatial-energetic frameworks, spatial-weighted implementations, and energy-only variants with posterior spatial reweighting. Severe spectral degeneracy limited discrimination, but the workflow demonstrated a coherent Bayesian approach for large-scale event classification.

Our experiments demonstrate that explicitly modeling temporal structure in band-limited LFP activity with a GRU-Transformer yields significantly better decoding of visual stimuli than a linear classifier trained on flattened power features. Improvements were consistent across multiple sessions, conditions, and class-imbalanced datasets, highlighting the importance of temporal dynamics in neural decoding.

This project conducted a comparative analysis of The Simpsons and Family Guy dialogues using text mining, topic modeling (LDA and STM), and sentiment analysis. Results confirm many perceived differences between the two series and reveal evolving narrative and linguistic patterns over time.

My bachelor's thesis analyzed the Pantheon dataset, modeling over 11,000 historical figures as a directed network of interactions via Wikipedia hyperlinks, by integrating exploratory analysis (spatio-temporal distributions, centrality metrics) and inferential modeling with ERGMs on subnetworks. It identified data biases and model limitations (ERGMs degeneracy, computational complexity for large networks).