Supervised Random Walks: Predicting and Recommending Links in Social Networks Lars Backstrom and Jure Leskovec propose a method for predicting and recommending links in social networks by combining network structure with node and edge attributes. The algorithm uses supervised learning to assign edge strengths that guide random walks, making them more likely to visit nodes where new links will be created. This approach outperforms existing unsupervised and feature-based methods on Facebook and collaboration networks. The method involves learning edge strength functions that model random walk transition probabilities. By using node and edge features, the algorithm learns to bias random walks towards nodes where new links are likely to form. This is achieved through a supervised learning task where the goal is to learn a function that assigns strengths to edges such that random walks are more likely to visit nodes to which new links will be created. The algorithm is evaluated on synthetic and real-world data, showing that it performs well in both scenarios. On synthetic data, it successfully recovers true edge strength parameters even in the presence of noise. On real data, it outperforms existing methods in link prediction and recommendation tasks. The method is general and can be applied to various types of interactions, not just links in social networks. It combines node attributes and network structure information, making it effective for predicting unobserved links in protein-protein interaction networks or suggesting relevant web pages for bloggers. The algorithm uses a supervised random walk approach, where edge strengths are learned to bias the random walk towards target nodes. This is done by minimizing a loss function that ensures the random walk scores of target nodes are higher than those of non-target nodes. The optimization problem is solved using gradient-based methods, and the algorithm is shown to converge to a local optimum. The method is tested on real-world data, including four physics co-authorship networks and a Facebook network in Iceland. The results show that the algorithm performs well in predicting links to nodes that are 2-hops away from the seed node. The algorithm is also tested with different loss functions, edge strength functions, and random walk restart parameters, showing that the Wilcoxon-Mann-Whitney loss function performs best in terms of AUC and Precision at Top 20. The algorithm is extended to handle different edge types and social capital, showing that it can improve performance by considering additional features such as the number of common friends between nodes. The method is shown to be effective in both social networks and other domains, demonstrating its versatility and effectiveness in link prediction and recommendation tasks.Supervised Random Walks: Predicting and Recommending Links in Social Networks Lars Backstrom and Jure Leskovec propose a method for predicting and recommending links in social networks by combining network structure with node and edge attributes. The algorithm uses supervised learning to assign edge strengths that guide random walks, making them more likely to visit nodes where new links will be created. This approach outperforms existing unsupervised and feature-based methods on Facebook and collaboration networks. The method involves learning edge strength functions that model random walk transition probabilities. By using node and edge features, the algorithm learns to bias random walks towards nodes where new links are likely to form. This is achieved through a supervised learning task where the goal is to learn a function that assigns strengths to edges such that random walks are more likely to visit nodes to which new links will be created. The algorithm is evaluated on synthetic and real-world data, showing that it performs well in both scenarios. On synthetic data, it successfully recovers true edge strength parameters even in the presence of noise. On real data, it outperforms existing methods in link prediction and recommendation tasks. The method is general and can be applied to various types of interactions, not just links in social networks. It combines node attributes and network structure information, making it effective for predicting unobserved links in protein-protein interaction networks or suggesting relevant web pages for bloggers. The algorithm uses a supervised random walk approach, where edge strengths are learned to bias the random walk towards target nodes. This is done by minimizing a loss function that ensures the random walk scores of target nodes are higher than those of non-target nodes. The optimization problem is solved using gradient-based methods, and the algorithm is shown to converge to a local optimum. The method is tested on real-world data, including four physics co-authorship networks and a Facebook network in Iceland. The results show that the algorithm performs well in predicting links to nodes that are 2-hops away from the seed node. The algorithm is also tested with different loss functions, edge strength functions, and random walk restart parameters, showing that the Wilcoxon-Mann-Whitney loss function performs best in terms of AUC and Precision at Top 20. The algorithm is extended to handle different edge types and social capital, showing that it can improve performance by considering additional features such as the number of common friends between nodes. The method is shown to be effective in both social networks and other domains, demonstrating its versatility and effectiveness in link prediction and recommendation tasks.