Trend Analysis link prediction in complex networks and forecasting of international trade flows

Abstract

Trend analysis plays a central role in understanding and predicting the evolution of complex systems by extracting meaningful patterns from historical data. This thesis addresses two fundamental predictive tasks that arise in this context: link prediction in complex networks and time series forecasting, with a particular focus on social networks and international trade. These domains are characterized by large-scale, high-dimensional, and dynamically evolving data, which pose significant challenges to traditional heuristic, statistical, and econometric approaches. For link prediction in social networks, this thesis proposes a novel hybrid framework that integrates similarity measures with supervised machine learning and heuristicbased feature selection. The proposed approach learns predictive patterns directly from similarity-based features while employing a Binary Butterfly Optimization inspired strategy with a mutation-driven mechanism to select informative feature subsets, thereby reducing computational complexity while preserving an effective balance between exploration and exploitation. Model predictions are interpreted using SHAP-based explainability. Extensive experiments on a real-world social network datasets, including Facebook and Email-Eu-Core, demonstrate the effectiveness of the proposed framework, achieving prediction accuracies of up to 98% on Facebook and 85% on Email-Eu-Core. For time series forecasting in international trade, the thesis develops an interpretable forecasting framework based on the Temporal Fusion Transformer (TFT), capable of modeling long-range temporal dependencies and integrating heterogeneous inputs such as economic and geographic indicators. Using enriched datasets from the UN Comtrade database, the CEPII Gravity dataset, and the World Bank, the TFT-based model achieves a 17% improvement in the coefficient of determination (R2) compared to baseline models, including Random Forest and Graph Attention Networks. Building on this framework, a hybrid approach termed TransFusion-LM is introduced, which integrates large language models through prompt-based strategies without fine-tuning. By embedding TFT outputs and contextual features into structured prompts, TransFusion-LM further improves forecasting performance, with prompt-based GPT-4 variants achieving an R2 score of up to 94% while also reducing forecasting errors. In addition to improved accuracy, this approach provides intuitive, human-readable explanations of contextual influences on trade forecasts. Overall, this thesis advances trend-based predictive modeling by developing scalable, accurate, and interpretable frameworks for both network-based and time series-based prediction tasks. By combining optimization-inspired feature selection, transformer architectures and large language models, the proposed methods support more reliable and transparent decision-making in complex social and economic systems.