Ego-clustering based propagation prediction of time series in social networks

Online social networks are now recognized as an important platform for propagating information. Recently, a lot of efforts are made to predict the popularity of social information, to help the government and companies effectively control and guide public opinions. Information propagation in social networks exist many different temporal patterns, which are useful reference for predicting the future spreading of news. Considering temporal patterns are related to the network structure and information content, some researchers formulated a time series clustering problem to obtain tem­poral patterns in social networks. Previous clustering based algorithms take each cluster center as a typical propagation pattern, and then classify prediction object into its nearest-neighbor cluster. However, the nearest pattern can not fit the prediction object very precisely. This paper proposes a novel E-CTPM (Ego-Clustering based Temporal Prediction Model). E-CTPM utilizes the prediction object itself as a fixed cluster center, which can attract the most similar time series into one cluster and generate an ego-pattern for the prediction object. The tailored pattern fits the prediction object well, so it is suitable to indicate the future propagation. Experiments are carried on twitter and phrase datasets, with the results that E-CTPM outperforms the existing algorithms by achieving lower prediction bias and prediction variance. Meanwhile, E-CTPM has general applicability, which is able to work with multiple clustering methods and avoids the prediction difference resulting from classification methods.