Temporal Prediction Algorithm For Social Information Propagation

The increasingly popular social network provides a broad data base and application scenario for information propagation prediction research. Information propagation prediction research combines with a known information propagation pattern and use some methods to predict propagation trend in the next period of time to pre-understand the entire process of information propagation. With information propagation prediction methods, Network companies can better provide users with personalized recommendation services and government departments can effectively control and guide public opinions.Information propagation prediction research involves multiple research areas like large-scale data parallel processing,analysis of toplogy structure of social network and analysis of text content, which attracts attention of researchers from multiple research areas like big data and cloud computation, complex network and natural language processing.Information propagation prediction is an important direction of social network research. The recent researches can be divided into graph and non-graph. Most nongraph based methods use the epidemic model and the classification model but rarely consider clustering characteristics of social time series. In Clustering based Temporal Prediction Algorithm CTP,each cluster center is used as a kind of propagation pattern and therefore prediction can be realized by finding out the nearest-neighbor pattern of the prediction object, that is, CTP treats the nearest-neighbor pattern as the prediction result. The prediction performance of CTP depends on fitting degree between the prediction object and its nearest-neighbor cluster center, and the higher the fitting degree, the better the prediction performance of CTP. The paper analyzes the physical meaning of the scaling distance and observes that the scaling distance can better measure similarity between time series. In the paper, we believe that the nearest-neighbor cluster center of the prediction object based on scaling distance may better match the predition object and obtain higher prediction performance, but the related literatures about CTP lack research that prediction performance is influenced by scaling distance. The paper combines CTP with scaling distance to propose Scaling Clustering based Temporal Prediction Algorithm S-CTP. The modified S-CTP takes scaled nearest-neighbor cluster center of the prediction object as prediction result to improve fitting degree between prediction result and the prediction object and therefore improve prediction performance.Experiment results on twitter and phrase datasets show that S-CTP obtains better generalization performance than CTP.In CTP, the similarity between some nearest-neighbor clustering members of the prediction object and the prediction object is higher while the similarity between the others and the prediction object is lower, which lowers prediction performance of CTP. The paper combines CTP with time series division to propose a Division Clustering based Temporal Prediction Algorithm D-CTP to solve the problem of lower prediction performance of CTP.D-CTP modified always takes the prediction object as a cluster center and conducts clustering on time series segments of known length of the prediction object and then refines cluster centers on time series segments of known length and prediction length. Similar to S-CTP, the paper combines D-CTP and the scaling distance to propose Scaling Division Clustering based Temporal Prediction Algorithm. Experiment results on twitter and phrase datasets show that the algorithm based on scaling distance and division clustering can further improve generalization performance of CTP on the basis of S-CTP.