Integration of Algorithmic and Local Approaches for Link Prediction: An Analysis on Protein-Protein Interactions and Social Networks

Abstract

Complex network analysis is applied in various fields such as network-based systems, social media recommendation systems, shopping platforms, and treatment methodologies. In this context, predicting the probability of connection between two nodes has become a focal point. Another significant aspect is the prediction of connections between proteins, especially with the increase in epidemic diseases. Link prediction methods, based on graph structures, aim to predict interactions between two nodes and measure the probability of connection between them. These methods proceed by relying on similarity values and can have multiple approaches, including local, global, and algorithmic. This study has emerged from a combination of algorithmic and local network approaches. Support Vector Machines are employed to predict connections in gene-protein networks and social network structures. Data sets from multiple social media platforms and human protein-protein interaction (PPI) data were utilized. Derived data were created by calculating local index values, including the number of neighbors, Adamic Adar index, Jaccard coefficient, and label values for each node. To enhance success rates, a model was developed that applied the discretization method as a preprocessing technique across all data sets. Machine learning algorithms such as Bayesian Networks, Multilayer Perceptron (MLP), Random Forest, and k-Nearest Neighborhood (kNN) were compared and evaluated. The results indicate that the Twitch dataset, which has the highest number of edges, produced successful outcomes. The contribution of edge numbers in the network structure to performance is highlighted, and it is observed that more successful metric values were obtained for the data with applied discretization.