Results in Engineering (Dec 2023)
Nifty method for prediction dynamic features of online social networks from users’ activity based on machine learning
Abstract
Nowadays with the development of m4obile personal devices, the interaction of most people takes place through online social network more than ever. They rely on online applications to communicate, express their opinions, or react to others expressions instead of waiting the time to do that directly in a real life. Computationally, such interaction is modeled as a virtual network (or formally as graph) that is described with a set of features e.g. graph diameter, average-clustering coefficient etc. To compute these features, it is required to count or inspect all nodes or/and edges properties of the graph. When the graph is dynamic, i.e., the structure is change over time with each interaction, the computation of these features is a challenge and complex for time and space. Instead, AI based approaches are suggested to predict such features based on only few information of an interaction. This work trains the machine to learn computing the global features of online social networks through noticing the effect of users’ interaction on these features. Three datasets of different of two real online social networks are used in experiments. The obtained result shows an approximate accuracy of predication by RF 99% for Email interaction dataset and 82% by TreeNet for College message interaction dataset. The results of paper refer to: (a) The paper demonstrates the successful application of machine learning techniques to predict and analyze dynamic graph features in online social networks. (b) The Random Forest model emerges as the most suitable approach, providing accurate predictions and insights into the dynamics of dynamic graphs. (c) The study highlights the importance of the Graph Average Clustering Coefficient (GACC) as a significant feature in predicting global graph dynamics. (d) The research showcases the limitations of the KNN-Regression model and emphasizes the need for models that can handle the complexity and nonlinearity of dynamic graphs effectively.
