Many real-world complex systems can be well-represented with graphs, where nodes represent objects or entities and links/relations represent interactions between pairs of nodes. Link prediction (LP) is one of the most interesting and long-standing problems in the field of graph mining; it predicts the probability of a link between two unconnected nodes based on available information in the current graph. This thesis studies the LP problem in graphs. It consists of two parts: LP in simple graphs and LP knowledge graphs (KGs). In the first part, the LP problem is defined as predicting the probability of a link between a pair of nodes in a simple graph. In the first study, a few similarity-based and embedding-based LP approaches are evaluated and compared on simple graphs from various domains. he study also criticizes the traditional way of computing the precision metric of similarity-based approaches as the computation faces the difficulty of tuning the threshold for deciding the link existence based on the similarity score. We proposed a new way of computing the precision metric. The results showed the expected superiority of embedding-based approaches. Still, each of the similarity-based approaches is competitive on graphs with specific properties. We could check experimentally that similarity-based approaches are fully explainable but lack generalization due to their heuristic nature, whereas embedding-based approaches are general but not explainable. The second study tries to alleviate the unexplainability limitation of embedding-based approaches by uncovering interesting connections between them and similarity-based approaches to get an idea of what is learned in embedding-based approaches. The third study demonstrates how the similarity-based approaches can be ensembled to design an explainable supervised LP approach. Interestingly, the study shows high LP performance for the supervised approach across various graphs, which is competitive with embedding-based approaches.The second part of the thesis focuses on LP in KGs. A KG is represented as a collection of RDF triples, (head,relation,tail) where the head and the tail are two entities which are connected by a specific relation. The LP problem in a KG is formulated as predicting missing head or tail entities in a triple. LP approaches based on the embeddings of entities and relations of a KG have become very popular in recent years, and generating negative triples is an important task in KG embedding methods. The first study in this part discusses a new method called SNS to generate high-quality negative triples during the training of embedding methods for learning embeddings of KGs. The results we produced show better LP performance when SNS is injected into an embedding approach than when injecting state-of-the-art negative triple sampling methods. The second study in the second part discusses a new neuro-symbolic method of mining rules and an abduction strategy to explain LP by an embedding-based approach utilizing the learned rules. The third study applies the explainable LP to a COVID-19 KG to develop a new drug repurposing approach for COVID-19. The approach learns ”ensemble embeddings” of entities and relations in a COVID-19 centric KG, in order to get a better latent representation of the graph elements. For the first time to our knowledge, molecular docking is then used to evaluate the predictions obtained from drug repurposing using KG embedding. Molecular evaluation and explanatory paths bring reliability to prediction results and constitute new complementary and reusable methods for assessing KG-based drug repurposing. The last study proposes a distributed architecture for learning KG embeddings in distributed and parallel settings. The results of the study that the computational time of embedding methods improves remarkably without affecting LP performance when they are trained in the proposed distributed settings than the traditional centralized settings.