Dual-Channel Feature Fusion for Joint Prediction in Dynamic Signed Weighted Networks

TL;DR

This work tackles joint prediction in dynamic signed weighted networks, focusing on predicting future link existence, edge signs, and edge weights simultaneously. It introduces LSWJP, which combines sign-aware node embeddings, multi-hop structural balance features, and temporal difference cues within a dual-channel Transformer-based encoder to decouple semantic and structural signals. A two-task decoder performs link existence, sign classification, and weight regression in a shared framework, using dynamic negative sampling for balanced training. Empirical results on diverse real-world datasets show strong link and sign prediction performance and notably improved weight prediction, validating the benefits of decoupled representations and joint optimization for dynamic signed weighted networks. The approach offers a generalizable framework that adapts to unsigned or unweighted networks and provides a pathway for more holistic network evolution forecasting in complex systems.

Abstract

Link prediction is central to unraveling social network evolution and node relationships, as well as understanding the characteristic mechanisms of complex networks. Currently, research on link prediction for complex dynamic networks integrating temporal evolution, relational polarity and edge weight information remains significantly underexplored, failing to meet practical demands. For dynamic signed-weighted networks, this paper proposes a tripartite joint prediction framework for unified forecasting of links, signs and weights. First, the dynamic network is decomposed into temporal snapshots, and node semantic embeddings are generated via sign-aware weighted random walks. We then design multi-hop structural balance and temporal difference features to capture the structural characteristics and dynamic evolution laws of the network, respectively. The model adopts a dual-channel feature decoupling mechanism: node semantic embeddings are used for link existence prediction, while relational sign features are fed into a Transformer encoder to model temporal dependencies. Finally, prediction results are output synergistically through a multi-task unit. Simulation experiments demonstrate that, compared with baseline methods, the proposed framework achieves an average 2%-4% improvement in the performance of link existence and relational sign prediction, and a significant 40%-50% reduction in edge weight prediction error.

Figures (9)

• Figure 1: Overall architecture diagram of the proposed LSWJP
• Figure 2: Variation of link prediction AUC with hyperparameters
• Figure 3: Variation of sign prediction AUC with hyperparameters
• Figure 4: Variation of weight prediction MAE with hyperparameters
• Figure 5: Variation of weight prediction RMSE with hyperparameters