Real-world infection data are often highly fragmented, with many orphan cases. Link prediction algorithms can be used to reconstruct infection trees by predicting missing links between infected individuals. Identifying these missing links can help epidemiologists obtain a more complete view of the epidemiological situation, rather than a partial one, and support better decision-making. Moreover, the reconstructed networks can be used in epidemic modeling to simulate scenarios without relying on synthetic networks (e.g. scale-free, random, small-world), leading to more accurate and reliable predictions. In this work, machine learning classifiers were trained to predict whether an edge exists between a pair of nodes. Feature vectors were derived from node-level epidemiological and structural attributes, including outdegree, indegree, day of infection, NACE code, gender, and postcode. For each node pair, these features were combined using different operators to construct an edge-level feature vector. This representation, derived from information from existing links in the infection network, was used to train machine learning algorithms for edge prediction. Random Forest and Gradient Boosting achieved the best test-set performance with an AUC of ~ 0.83 – 0.92 for different waves. After the best-performing model was identified, it was applied to orphan cases, and the infection networks were reconstructed to include those. Across all waves, the inferred networks contained fewer isolated components and exhibited larger component sizes.