A two-stage model leveraging friendship network for community evolution prediction in interactive networks
Yanmei Hu, Yihang Wu, Biao Cai
TL;DR
This work tackles the problem of predicting both the type and the extent of community evolution in interactive networks, addressing the gap of fine-grained evolution and cross-network information utilization. It introduces LTSModel, a two-stage framework that first classifies evolution type with a hybrid Random Forest–SVM classifier and then regresses the evolution extent using a gradient-boosted regressor, while incorporating features from both the interactive network and the more stable friendship network. Empirical results on Sina Weibo data and public datasets demonstrate substantial improvements over diverse baselines in type accuracy and extent accuracy (MAPE), with ablations confirming the value of cross-network features and the hybrid classifier. The approach advances practical forecasting of event-related community dynamics and suggests future work on joint optimization and deep embedding methods for further performance gains.
Abstract
Interactive networks representing user participation and interactions in specific "events" are highly dynamic, with communities reflecting collective behaviors that evolve over time. Predicting these community evolutions is crucial for forecasting the trajectory of the related "event". Some models for community evolution prediction have been witnessed, but they primarily focused on coarse-grained evolution types (e.g., expand, dissolve, merge, split), often neglecting fine-grained evolution extents (e.g., the extent of community expansion). Furthermore, these models typically utilize only one network data (here is interactive network data) for dynamic community featurization, overlooking the more stable friendship network that represents the friendships between people to enrich community representations. To address these limitations, we propose a two-stage model that predicts both the type and extent of community evolution. Our model unifies multi-class classification for evolution type and regression for evolution extent within a single framework and fuses data from both interactive and friendship networks for a comprehensive community featurization. We also introduce a hybrid strategy to differentiate between evolution types that are difficult to distinguish. Experimental results on three datasets show the significant superiority of the proposed model over other models, confirming its efficacy in predicting community evolution in interactive networks.