在技术变革加速与政策导向复杂交织的背景下, 地方政府亟需智能化、可解释的未来产业布局决策支持工具, 以应对区域间资源禀赋的差异与产业演化路径的不确定性。传统产业布局方法在处理多源异构特征、高阶结构交互与动态政策约束等方面存在明显不足。为此, 该文提出了一种融合特征重标定、多头注意力机制与神经矩阵分解的自适应深度推荐算法。该算法通过压缩与激励网络模块实现关键因素的动态加权, 通过多头注意力机制捕捉高阶非线性交互关系, 通过神经矩阵分解模块捕捉线性与非线性关联, 三者协同实现对地区-产业适配关系的精准建模与对产业布局的高质量推荐。实证结果表明: 该算法在多个性能指标上显著优于主流基线算法, 具有良好的泛化能力、结构可解释性与政策可采纳性。该研究不仅拓展了深度推荐算法在结构化决策支持场景中的理论应用边界, 也为地方政府制定高质量未来产业布局策略提供了数据驱动的系统性工具。

Objective: With the accelerating pace of technological innovation and the growing complexity of policy environments, local governments urgently require intelligent and interpretable decision support systems to guide the strategic layout of future-oriented industries. However, traditional industrial layout approaches rely heavily on expert judgment and static data analysis, making them ineffective at handling dynamic market demands, integrating heterogeneous features, and capturing the complex nonlinear interactions between regional resources and emerging industries. This study aims to address these critical limitations by developing an adaptive deep recommendation algorithm. The proposed algorithm provides data-driven, actionable insights to local policymakers, facilitating the accurate and strategic allocation of regional resources for emerging and strategic industries. Methods: This paper proposes a deep recommendation framework that integrates three key components: feature recalibration, multihead attention mechanisms, and neural matrix factorization. First, diverse regional attributes (such as gross domestic product (GDP), research and development (R&D) expenditure, and infrastructure indicators), industrial characteristics (such as growth rate and technological maturity), and policy orientations (extracted from over 280 local policy documents via semantic embedding) are transformed into dense vector representations through embedding layers. Next, a feature recalibration module inspired by the squeeze-and-excitation network is employed to dynamically reweight critical regional and industrial features, thereby underscoring influential factors and suppressing noise. Subsequently, multihead attention mechanisms are introduced to capture high-order nonlinear interactions across recalibrated features and to model complex interdependencies between regions, industries, and policy orientations effectively. Finally, neural matrix factorization techniques combine collaborative signals with the nonlinear embeddings to score and rank the suitability of each region-industry pairing quantitatively. The proposed model is trained on a comprehensive dataset comprising over 3 000 real-world region-industry samples with a margin-based loss function optimized through supervised learning. Extensive tuning of hyperparameters, including embedding dimensions, dropout rates, and attention head counts, ensures robust model performance. Results: Empirical validation demonstrates that the proposed adaptive deep recommendation algorithm substantially outperforms mainstream baseline models such as DeepFM, AutoInt, and FiBiNet across multiple performance metrics, including logarithmic loss (logloss) and area under the curve (AUC). Specifically, this algorithm achieves the highest AUC score of 0.714 6 and the lowest logloss of 0.485 8 among all tested models, which confirms its superior predictive accuracy and classification capability. Ablation experiments further reveal that each module contributes distinctively to overall performance: Removing the feature recalibration module, multihead attention mechanism, or neural matrix factorization component results in a noticeable degradation of predictive accuracy. Additionally, comparative analysis with a prominent large language model (LLM), such as GPT-4o, highlights the advantage of this structured algorithm in handling numerical and structured data, in contrast to semantic reasoning that limit structural modeling capacities of the LLM. Visualization of attention weights confirms the algorithm's interpretability and explicitly demonstrates its sensitivity to key factors such as regional R&D intensity, infrastructure readiness, and industry technological maturity. Conclusions: This study successfully establishes an adaptive, interpretable, and highly effective deep recommendation algorithm tailored explicitly to future-oriented industrial layout planning. By integrating dynamic feature recalibration, high-order feature interaction modeling, and robust collaborative filtering mechanisms, the proposed algorithm remarkably enhances the accuracy, interpretability, and practical applicability of the recommendations. The proposed algorithm not only provides local policymakers with transparent, data-driven decision support but also sets a theoretical foundation for further exploration into advanced recommendation frameworks. Future research will aim to incorporate temporal dynamics through real-time data streams; expand the framework to multiobjective scenarios covering economic, social, and ecological benefits; integrate structured recommendation outputs with semantic insights from expert knowledge; and ultimately realize a comprehensive, adaptive industrial policy decision support platform.