A Minimal Description Length Approach For Social Causal Discovery

With the rapid development of social network, the relative industries such as micro-business, microblog marketing and socialized e-business make great progresses since a growing amounts of people start mining its potential value. Among multitudinous correlative researches on social network, the user influence has realistic meaning to the opinion guidance and Microblog marketing, which is the obstruction and hot spot at the moment.The current research approaches of user influence focus on explicitly friend network of user. However, these methods are frequently replete with larger redundancy. Specifically, the massive explicitly friend networks have no actual function. Therefore, relying on the user’s behavior data to mining the causal network in user behaviors plays a key role in evaluating the user influence.Yet, the existing approaches on social network causal inference have two drawbacks:firstly, there are too many redundant edges in the causal network graph resulting from unable to recognize indirect casual influence. Next, without sufficient consideration of causality lag.Our mining model, MCRN, bases on the Minimal Description Length norm to uniformly modeling on the above two issues. On the respect of minimizing causal network redundancy, MCRN applies the Causal Transfer Entropy algorithm to discover the social network causality and expand the causal transfer entropy combing with causality lag length, which efficiently removes the redundant edges in the graph and improves the accuracy. With respect to the causality lag length, MCRN regards MDL as scoring system to weigh the uncertainty and complexity of model and effectually reduce the overfitting.Massive simulative datasets show that multiple evaluation indices of our model MCRN precedes similar algorithms such as TE, CSE, etc. And our experiment on real dataset from Sina blog found a phenomenon that many users explicitly declarative friend relations have no causal influence and interactivity in users with causality, both of which verified the validity of our model.Finally, based on MCRN theoretical model, this paper demonstrates a time series social networking discovery system scheme, the MCRN System, and envisages the system architecture and rudiment that makes it easy to accurately and intuitively analyze the causality between users and progressively apply it to other realms of real life.