Simone Piaggesi

Predicting cancer drug response from transcriptomic profiles is a cornerstone of precision oncology, yet the scientific value of machine learning models hinges not solely on predictive accuracy, but also on their capacity to generate reliable biological insights. Current explainability approaches in this setting are computationally costly, lack robustness, and reduce complex drug response to univariate gene importance scores, overlooking the coordinated gene activity that drives sensitivity and resistance. In this work, we present ILLUME+, a scalable post-hoc explainability framework that moves beyond single-gene assessments to capture multiple, complementary forms of explanation. Integrated into our end-to-end pipeline, ILLUME+ produces more stable gene importance scores than existing baselines, recovers established drug-gene associations and mechanisms of action, and enables AI-assisted hypothesis generation to uncover novel interaction-driven molecular signals in cancer biology.

Post-hoc explainability is essential for understanding black-box machine learning models. Surrogate-based techniques are widely used for local and global model-agnostic explanations but have significant limitations. Local surrogates capture non-linearities but are computationally expensive and sensitive to parameters, while global surrogates are more efficient but struggle with complex local behaviors. In this paper, we present ILLUME, a flexible and interpretable framework grounded in representation learning, that can be integrated with various surrogate models to provide explanations for any black-box classifier. Specifically, our approach combines a globally trained surrogate with instance-specific linear transformations learned with a meta-encoder to generate both local and global explanations. Through extensive empirical evaluations, we demonstrate the effectiveness of ILLUME in producing feature attributions and decision rules that are not only accurate but also robust and computationally efficient, thus providing a unified explanation framework that effectively addresses the limitations of traditional surrogate methods.

Node embeddings are low-dimensional vectors that capture node properties, typically learned through unsupervised structural similarity objectives or supervised tasks. While recent efforts have focused on post-hoc explanations for graph models, intrinsic interpretability in unsupervised node embeddings remains largely underexplored. To bridge this gap, we introduce DiSeNE (Disentangled and Self-Explainable Node Embedding), a framework that learns self-explainable node representations in an unsupervised fashion. By leveraging disentangled representation learning, DiSeNE ensures that each embedding dimension corresponds to a distinct topological substructure of the graph, thus offering clear, dimension-wise interpretability. We introduce new objective functions grounded in principled desiderata, jointly optimizing for structural fidelity, disentanglement, and human interpretability. Additionally, we propose several new metrics to evaluate representation quality and human interpretability. Extensive experiments on multiple benchmark datasets demonstrate that DiSeNE not only preserves the underlying graph structure but also provides transparent, human-understandable explanations for each embedding dimension.

Graphs are ubiquitous due to their flexibility in representing social and technological systems as networks of interacting elements. Graph representation learning methods, such as node embeddings, are powerful approaches to map nodes into a latent vector space, allowing their use for various graph tasks. Despite their success, only few studies have focused on explaining node embeddings locally. Moreover, global explanations of node embeddings remain unexplored, limiting interpretability and debugging potentials. We address this gap by developing human-understandable explanations for dimensions in node embeddings. Towards that, we first develop new metrics that measure the global interpretability of embedding vectors based on the marginal contribution of the embedding dimensions to predicting graph structure. We say that an embedding dimension is more interpretable if it can faithfully map to an understandable sub-structure in the input graph - like community structure. Having observed that standard node embeddings have low interpretability, we then introduce DINE (Dimension-based Interpretable Node Embedding), a novel approach that can retrofit existing node embeddings by making them more interpretable without sacrificing their task performance. We conduct extensive experiments on synthetic and real-world graphs and show that we can simultaneously learn highly interpretable node embeddings with effective performance in link prediction.

Methods that learn graph topological representations are becoming the usual choice to extract features to help solve machine learning tasks on graphs. In particular, low-dimensional encoding of graph nodes can be exploited in tasks such as link prediction and network reconstruction, where pairwise node embedding similarity is interpreted as the likelihood of an edge incidence. The presence of polyadic interactions in many real-world complex systems is leading to the emergence of representation learning techniques able to describe systems that include such polyadic relations. Despite this, their application on estimating the likelihood of tuple-wise edges is still underexplored. Here we focus on the reconstruction and prediction of simplices (higher-order links) in the form of classification tasks, where the likelihood of interacting groups is computed from the embedding features of a simplicial complex. Using similarity scores based on geometric properties of the learned metric space, we show how the resulting node-level and group-level feature embeddings are beneficial to predict unseen simplices, as well as to reconstruct the topology of the original simplicial structure, even when training data contain only records of lower-order simplices.

Mobile phone data have been extensively used to study urban mobility. However, studies based on gender-disaggregated large-scale data are still lacking, limiting our understanding of gendered aspects of urban mobility and our ability to design policies for gender equality. Here we study urban mobility from a gendered perspective, combining commercial and open datasets for the city of Santiago, Chile. We analyze call detail records for a large cohort of anonymized mobile phone users and reveal a gender gap in mobility: women visit fewer unique locations than men, and distribute their time less equally among such locations. Mapping this mobility gap over administrative divisions, we observe that a wider gap is associated with lower income and lack of public and private transportation options. Our results uncover a complex interplay between gendered mobility patterns, socio-economic factors and urban affordances, calling for further research and providing insights for policymakers and urban planners.

Reliable data about socio-economic conditions of individuals, such as health indexes, consumption expenditures and wealth assets, remain scarce for most countries. Traditional methods to collect such data include on site surveys that can be expensive and labour intensive. On the other hand, remote sensing data, such as high-resolution satellite imagery, are becoming largely available. To circumvent the lack of socio-economic data at high granularity, computer vision has already been applied successfully to raw satellite imagery sampled from resource poor countries. In this work we apply a similar approach to the metropolitan areas of five different cities in North and South America, starting from pre-trained convolutional models used for poverty mapping in developing regions. Applying a transfer learning process we estimate household income from visual satellite features. The urban environment we consider is characterized by different features with respect to the resource-poor training environment, such as the high heterogeneity in population density. By leveraging both official and crowd-sourced data at city scale, we show the feasibility of estimating the socio-economic conditions of different neighborhoods from satellite data.