| Abstract: |  The potential value of predicting trends in social media rises with its growing dominance in our lives. Whereas many works focus on anomaly or trend detection, there is still little knowledge on the evolution of trend dynamics. Inspired by the studies on infection diffusion through a social network, we propose an approach to predict trends spread within a local subnetwork of Twitter, exploiting the network structure information beyond the scope of previous works. We base the anomaly pattern representation on graph features, reflecting various local relational topology options in the context of trend presence. Utilizing machine learning algorithm, the information extracted is used for prediction of future trends behavior and evaluated over several demarcated targets. The contribution of our graph approach is then measured against a baseline model, utilizing the same learning strategy, yet considering the trends as time series, absent any knowledge on the network topology. Moreover several other approaches are tested for comparison.
The results prove the network structure to play an important role in the trends spread dynamics, as the topology information extracted via graph features improves the accuracy of learner considerably, out of the reach of other methods tested. Further feature options and combinations can be considered for prospective improvements of the network related approach.
|
|---|