| \contentsline {section}{\numberline {1}如何阅读这份讲义}{3}{section.1}% |
| \contentsline {section}{\numberline {2}执行摘要：最终系统做了什么}{3}{section.2}% |
| \contentsline {section}{\numberline {3}问题背景：为什么这是异构图推荐问题}{4}{section.3}% |
| \contentsline {subsection}{\numberline {3.1}节点与边的语义}{4}{subsection.3.1}% |
| \contentsline {subsection}{\numberline {3.2}评价指标和建模目标}{5}{subsection.3.2}% |
| \contentsline {section}{\numberline {4}数据挑战：稀疏、长尾和冷启动}{5}{section.4}% |
| \contentsline {section}{\numberline {5}总体架构设计}{6}{section.5}% |
| \contentsline {subsection}{\numberline {5.1}为什么采用两阶段堆叠}{6}{subsection.5.1}% |
| \contentsline {subsection}{\numberline {5.2}系统信息流}{7}{subsection.5.2}% |
| \contentsline {section}{\numberline {6}性能提升路径：从 0.885 到 0.96626}{7}{section.6}% |
| \contentsline {section}{\numberline {7}早期路线和关键判断}{8}{section.7}% |
| \contentsline {subsection}{\numberline {7.1}更复杂的 GNN 不一定更好}{8}{subsection.7.1}% |
| \contentsline {subsection}{\numberline {7.2}排序目标比普通二分类更合适}{8}{subsection.7.2}% |
| \contentsline {subsection}{\numberline {7.3}验证协议必须可靠}{8}{subsection.7.3}% |
| \contentsline {section}{\numberline {8}LightGCN 主干}{8}{section.8}% |
| \contentsline {subsection}{\numberline {8.1}LightGCN 提供哪些特征}{9}{subsection.8.1}% |
| \contentsline {section}{\numberline {9}显式图结构和 Meta-path 特征}{9}{section.9}% |
| \contentsline {section}{\numberline {10}内容特征与 BPR-MF}{10}{section.10}% |
| \contentsline {subsection}{\numberline {10.1}内容画像如何构造}{10}{subsection.10.1}% |
| \contentsline {subsection}{\numberline {10.2}BPR-MF 的作用}{10}{subsection.10.2}% |
| \contentsline {section}{\numberline {11}DeepWalk 与 Node2Vec 随机游走特征}{11}{section.11}% |
| \contentsline {section}{\numberline {12}核心创新：引用感知高阶传播}{12}{section.12}% |
| \contentsline {subsection}{\numberline {12.1}矩阵定义}{12}{subsection.12.1}% |
| \contentsline {subsection}{\numberline {12.2}为什么它有效}{12}{subsection.12.2}% |
| \contentsline {subsection}{\numberline {12.3}伪代码}{13}{subsection.12.3}% |
| \contentsline {subsection}{\numberline {12.4}消融结果}{13}{subsection.12.4}% |
| \contentsline {section}{\numberline {13}LightGBM OOF 堆叠}{14}{section.13}% |
| \contentsline {subsection}{\numberline {13.1}为什么需要 OOF}{14}{subsection.13.1}% |
| \contentsline {subsection}{\numberline {13.2}LightGBM 学到什么}{14}{subsection.13.2}% |
| \contentsline {section}{\numberline {14}提交校准：为什么不用验证最优概率阈值}{14}{section.14}% |
| \contentsline {subsection}{\numberline {14.1}核心判断}{15}{subsection.14.1}% |
| \contentsline {section}{\numberline {15}错误分析：剩余问题在哪里}{15}{section.15}% |
| \contentsline {section}{\numberline {16}项目文件地图}{16}{section.16}% |
| \contentsline {section}{\numberline {17}复现和检查命令}{17}{section.17}% |
| \contentsline {section}{\numberline {18}汇报时可以怎么讲}{18}{section.18}% |
| \contentsline {subsection}{\numberline {18.1}开场 30 秒}{18}{subsection.18.1}% |
| \contentsline {subsection}{\numberline {18.2}解释为什么不是更深 GNN}{18}{subsection.18.2}% |
| \contentsline {subsection}{\numberline {18.3}解释高阶传播创新}{18}{subsection.18.3}% |
| \contentsline {subsection}{\numberline {18.4}解释 rank cutoff}{18}{subsection.18.4}% |
| \contentsline {section}{\numberline {19}局限与后续优化方向}{18}{section.19}% |
| \contentsline {section}{\numberline {20}结论}{19}{section.20}% |
| \contentsline {section}{\numberline {A}附录：关键图表索引}{19}{appendix.A}% |
| |
