Scholarlink

<p>Ship material procurement suffers from persistent mismatches between buyer demand and supplier capacity, driven by customized components, volatile order cycles, and the reluctance of both sides to fully disclose planning information. Existing matching methods struggle under such semi-public conditions, where long forecasting horizons, heterogeneous data streams, and bilateral confidentiality must be reconciled within a single decision framework. We use the term semi-public to denote partial, selective, and conditional disclosure, in which information is visible to the matching platform and to a filtered set of counterparties but never made fully public. This paper proposes an MHA-TCN model that couples multi-head self-attention with temporal convolutional networks to address dynamic pre-matching under such disclosure. A tiered encoding scheme separates attributes by sensitivity level, applying bucketization and calibrated noise to balance information protection against predictive fidelity. Stacked dilated causal convolutions extract long-range temporal patterns from inventory, approval, and inquiry streams, while masked cross-attention fuses buyer and supplier representations through a learned gating mechanism that enforces bilateral shielding constraints during training. The model jointly optimizes compatibility scoring, time-banded quantity forecasting, and intent-handshake probability under a composite loss with explicit privacy penalty. Experiments on 3 years of procurement records from a large Chinese shipbuilder show that the proposed model improves top-5 matching F1 and lowers quantity-forecast error against the strongest baseline, and that it retains most of its matching quality under heavy attribute shielding; full numbers, with significance tests and confidence intervals, appear in Section 4. Ablation results confirm that the semi-public encoding and multi-head attention modules contribute most to performance, while case-level validation indicates hit rates above 87% against actual procurement decisions. The model offers a practical pathway toward proactive, privacy-respecting procurement planning in shipbuilding supply chains.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A semi-public dynamic pre-matching model based on multi-head self-attention and temporal convolutional network for ship material procurement

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年8月5日

查看全文
  • Chunlei Ding,
  • Hao Hu

摘要

Ship material procurement suffers from persistent mismatches between buyer demand and supplier capacity, driven by customized components, volatile order cycles, and the reluctance of both sides to fully disclose planning information. Existing matching methods struggle under such semi-public conditions, where long forecasting horizons, heterogeneous data streams, and bilateral confidentiality must be reconciled within a single decision framework. We use the term semi-public to denote partial, selective, and conditional disclosure, in which information is visible to the matching platform and to a filtered set of counterparties but never made fully public. This paper proposes an MHA-TCN model that couples multi-head self-attention with temporal convolutional networks to address dynamic pre-matching under such disclosure. A tiered encoding scheme separates attributes by sensitivity level, applying bucketization and calibrated noise to balance information protection against predictive fidelity. Stacked dilated causal convolutions extract long-range temporal patterns from inventory, approval, and inquiry streams, while masked cross-attention fuses buyer and supplier representations through a learned gating mechanism that enforces bilateral shielding constraints during training. The model jointly optimizes compatibility scoring, time-banded quantity forecasting, and intent-handshake probability under a composite loss with explicit privacy penalty. Experiments on 3 years of procurement records from a large Chinese shipbuilder show that the proposed model improves top-5 matching F1 and lowers quantity-forecast error against the strongest baseline, and that it retains most of its matching quality under heavy attribute shielding; full numbers, with significance tests and confidence intervals, appear in Section 4. Ablation results confirm that the semi-public encoding and multi-head attention modules contribute most to performance, while case-level validation indicates hit rates above 87% against actual procurement decisions. The model offers a practical pathway toward proactive, privacy-respecting procurement planning in shipbuilding supply chains.