Le Liu

Uneven regional socioeconomic growth presents a fundamental economic challenge, as traditional budget allocation models frequently fall short by neglecting spatial topology and spillover effects. By striking a statistically correct balance between social equality and economic efficiency, this study seeks to maximize regional public funds. We present a new hybrid approach that uses Graph Attention Networks (GAT) in conjunction with multi-criteria optimization to model administrative districts as undirected graphs. By combining regularizations for temporal stability and spatial smoothness, the composite loss function scalarizes a logarithmic utility function for efficiency and a differentiable Gini index for equity. The GNN-based model removed major historical funding imbalances, reducing the highest regional funding gap from an 80-fold difference to a 1.08 ratio and reducing per capita funding variance by 92%, according to empirical testing conducted in Kazakhstan’s Almaty region. The system automatically reorganized 208.5 billion KZT (~463 million USD) at the ideal Pareto frontier (α = 0.5), allocating 96.4% of investments to the digitalization sector, which was given priority. According to the results, incorporating spatial topology, which explains 58.7% of choice variability, offers a highly scalable, empirically supported method for improving strategic public administration and reducing institutional stagnation.

With the growing popularity of eSports, video highlight detection, which encapsulates the most informative parts in a few seconds, has become a critical part of live competition. However, learning the spatial–temporal dependency efficiently and discriminatively in video highlight detection for league of legends (LoL) is a critical problem. In this study, to address these existing problems, we propose a novel discriminative and efficient non-local attention network (DENAN) for LoL highlight detection. In particular, both spatial and temporal dependencies are learned using an end-to-end lightweight trainable framework. An auxiliary triplet loss is used in discriminative training to learn robust LoL video feature representations and improve DENAN’s performance. Our experimental results on the NLACS and LMS datasets show the effectiveness of our method in terms of performance and computation cost.