How Algorithms Read Our Emotions on Facebook?

Social networks are a gold mine of data: posts, comments, photos and reactions reveal a lot about who we are and how we feel. A research team has built a machine‑learning system that can analyze these signals and reconstruct two key aspects: social bonds and users' emotional states.

The model starts with a graph of the network, where each node is a profile and each edge an interaction (like, comment, share). Using clustering techniques, it identifies "communities", i.e., groups of close friends or shared interests. Then, with sentiment analysis, it examines text and images to determine whether a post is positive, neutral or negative, providing an estimate of the average mood of each community.

To spot the most influential users, the system calculates centrality metrics: the most connected or frequently mentioned accounts appear as "key nodes". Finally, to protect privacy, data are processed with federated learning, a method that trains the model directly on devices without sending sensitive information to a central server.

The results show that useful insights can be extracted to promote mental well‑being and strengthen digital ties, always respecting user confidentiality.

The proposed framework combines three methodological pillars. 1) **Graph‑based clustering**: the Louvain algorithm is used to maximize modularity, allowing detection of dense communities even in networks with millions of nodes. 2) **Multimodal sentiment analysis**: for text, a BERT model fine‑tuned on Italian corpora is employed, while for images a ResNet‑50 CNN is trained to recognize facial expressions and emotional contexts. The outputs of both modules are fused through an attention layer, improving emotion classification accuracy. 3) **Centrality and influence**: beyond degree and betweenness, eigenvector centrality metrics are calculated to identify users with non‑linear diffusion power.

For privacy, federated learning is implemented with the FedAvg algorithm, securely aggregating local weights. Differential privacy mechanisms (ε‑DP) are added to limit the possibility of reconstructing individual data. The approach was tested on a dataset of 10 million Facebook interactions, achieving an F1‑score of 0.87 in sentiment classification and a precision of 0.81 in influential node detection. These results suggest that integrating graph theory, NLP and computer vision, with strong privacy safeguards, can provide a solid foundation for digital well‑being monitoring applications.