[{"data":1,"prerenderedAt":3150},["ShallowReactive",2],{"blog-taxonomies":3,"blog-stats":507,"post-2026\u002F2026-01-03-论文实训草稿":1083},[4,11,16,21,25,29,36,39,44,47,52,58,63,66,71,76,80,84,87,93,96,101,105,110,113,117,122,127,132,136,141,143,146,151,155,157,162,165,169,173,176,179,183,186,190,193,197,200,205,208,211,214,219,222,228,234,237,239,245,249,253,256,260,266,271,276,281,283,287,290,293,298,301,304,306,308,311,314,316,318,321,327,332,337,342,347,352,354,356,358,361,365,367,369,371,374,376,378,381,386,390,393,396,400,404,407,411,413,415,417,421,423,425,428,431,433,435,437,439,441,443,446,448,450,454,456,459,461,463,465,467,469,471,473,477,483,486,488,491,493,495,497,499,501,503,505],{"category":5,"tags":6},"笔记",[7,8,9,10],"QQ","插件","LiteLoader","美化",{"category":12,"tags":13},"其他",[14,15,10],"博客","IP签名",{"category":17,"tags":18},"杂谈",[19,20],"互联网","观点",{"category":5,"tags":22},[14,23,24],"GitHub","数据可视化",{"category":5,"tags":26},[14,27,28],"访问统计","不蒜子",{"category":30,"tags":31},"工具",[32,33,34,35],"Python","爬虫","选课脚本","广州大学",{"category":5,"tags":37},[38,5],"数学分析",{"category":40,"tags":41},"算法",[42,40,43],"数据结构","线性表",{"category":40,"tags":45},[42,46],"复习",{"category":5,"tags":48},[49,50,51],"高等代数","线性代数","知识点",{"category":5,"tags":53},[14,54,55,56,57],"Typora","PicGo","图床","Markdown",{"category":30,"tags":59},[60,61,62],"Jupyter","工具配置","中文",{"category":30,"tags":64},[65,61,10],"VSCode",{"category":30,"tags":67},[68,69,70,35],"路由器","OpenWrt","校园网",{"category":12,"tags":72},[73,74,75],"网络协议","Wireshark","DEVP2P",{"category":5,"tags":77},[73,74,78,79],"Lua","解码器",{"category":5,"tags":81},[82,5,83],"常微分方程","数学",{"category":5,"tags":85},[86,5,83],"概率论",{"category":30,"tags":88},[89,90,91,92],"Docker","WSL","Windows","磁盘清理",{"category":30,"tags":94},[32,33,95,35],"课程表",{"category":5,"tags":97},[14,98,99,100],"Hexo","Gitalk","评论系统",{"category":5,"tags":102},[14,103,98,104],"Obsidian","写作工具",{"category":12,"tags":106},[107,108,109],"课题研究","泊松分布","参数估计",{"category":12,"tags":111},[107,108,109,112],"组会",{"category":12,"tags":114},[109,115,108,116],"IM模型","统计学",{"category":17,"tags":118},[119,120,121],"软考","网络规划","信息系统项目管理",{"category":30,"tags":123},[124,125,126],"Arch Linux","AppImage","Linux",{"category":5,"tags":128},[129,130,131],"编程语言","计算机科学","内存管理",{"category":5,"tags":133},[134,135],"Canvas","前端",{"category":5,"tags":137},[14,138,139,140],"GitHub Actions","Vite","部署",{"category":5,"tags":142},[135],{"category":5,"tags":144},[135,145],"系统设计",{"category":5,"tags":147},[135,148,149,150],"安全","XSS","CSRF",{"category":152,"tags":153},"项目",[134,154],"PixiJS",{"category":152,"tags":156},null,{"category":40,"tags":158},[42,159,160,161],"栈","队列","JavaScript",{"category":40,"tags":163},[42,164,161],"哈希表",{"category":5,"tags":166},[167,161,168],"ES6","Set",{"category":40,"tags":170},[40,171,172],"动态规划","背包问题",{"category":40,"tags":174},[40,175],"滑动窗口",{"category":40,"tags":177},[42,178,161],"链表",{"category":40,"tags":180},[40,181,161,182],"ACM","面试",{"category":40,"tags":184},[40,161,185],"数据类型转换",{"category":40,"tags":187},[42,188,178,189],"顺序表","TypeScript",{"category":5,"tags":191},[135,154,134,192],"渲染",{"category":5,"tags":194},[135,195,134,196],"富文本编辑","wangEditor",{"category":40,"tags":198},[199,40,178],"力扣",{"category":5,"tags":201},[202,203,204,91],"DNS","加密","网络安全",{"category":5,"tags":206},[135,207],"Axios",{"category":5,"tags":209},[135,207,189,210],"接口封装",{"category":5,"tags":212},[135,213,145],"JWT",{"category":152,"tags":215},[134,216,217,154,218],"画布","React","项目开发",{"category":17,"tags":220},[221],"日志",{"category":5,"tags":223},[224,225,226,227],"Neo4j","图数据库","中心度分析","PageRank",{"category":12,"tags":229},[230,231,232,233],"论文","机器学习","随机森林","保险欺诈",{"category":5,"tags":235},[236],"人工智能",{"category":5,"tags":238},[231,40,32],{"category":12,"tags":240},[241,242,243,244],"共识算法","最长链","区块链","算法设计",{"category":5,"tags":246},[161,247,248,135],"闭包","作用域",{"category":5,"tags":250},[14,138,251,252],"RSS","自动化",{"category":30,"tags":254},[89,90,91,255],"安装",{"category":30,"tags":257},[258,91,259],"MSYS2","开发环境",{"category":5,"tags":261},[262,263,264,265],"团队协作","Git","Commit规范","代码审查",{"category":30,"tags":267},[268,269,270],"oh-my-posh","终端美化","CLI",{"category":5,"tags":272},[273,274,275],"上下文工程","Agent","RAG",{"category":5,"tags":277},[278,279,280],"前端工程化","技术选型","工程基建",{"category":5,"tags":282},[275],{"category":5,"tags":284},[285,286],"SDD","规格驱动开发",{"category":5,"tags":288},[289],"neovim",{"category":5,"tags":291},[292],"技术写作",{"category":5,"tags":294},[295,32,296,89,297],"FastAPI","后端","工程实践",{"category":152,"tags":299},[274,300],"Harness Engineering",{"category":5,"tags":302},[303],"RFC",{"category":152,"tags":305},[274,300,275],{"category":152,"tags":307},[274],{"category":40,"tags":309},[199,40,161,310],"两数之和",{"category":40,"tags":312},[199,40,161,313],"双指针",{"category":40,"tags":315},[199,40,161,178],{"category":40,"tags":317},[199,40,161,175],{"category":40,"tags":319},[199,40,161,320],"回溯",{"category":5,"tags":322},[323,324,325,326],"分布式处理与计算","Spark","Scala","RDD",{"category":5,"tags":328},[323,329,330,331],"线性回归","逻辑回归","监督学习",{"category":5,"tags":333},[323,334,335,336],"数值优化","牛顿法","收敛性",{"category":5,"tags":338},[323,339,340,341],"PCA","数据降维","正则化",{"category":5,"tags":343},[323,344,345,346],"聚类","层次聚类","K-Means",{"category":5,"tags":348},[323,349,350,351],"随机模拟","统计推断","EM算法",{"category":5,"tags":353},[24],{"category":5,"tags":355},[24],{"category":5,"tags":357},[24],{"category":40,"tags":359},[42,360],"图论",{"category":40,"tags":362},[42,363,364],"树","二叉树",{"category":40,"tags":366},[42,160],{"category":40,"tags":368},[42],{"category":40,"tags":370},[42,159],{"category":40,"tags":372},[42,40,373],"复杂度分析",{"category":40,"tags":375},[42,43,178],{"category":40,"tags":377},[42,43,188],{"category":40,"tags":379},[42,380],"基础概念",{"category":17,"tags":382},[383,384,385],"随记","照片","高中",{"category":17,"tags":387},[383,388,389],"团建","战地",{"category":17,"tags":391},[383,392],"生活",{"category":17,"tags":394},[383,395],"文学",{"category":17,"tags":397},[383,398,399],"SCP","模因",{"category":17,"tags":401},[383,402,403],"画集","艺术",{"category":17,"tags":405},[14,406,383],"写作",{"category":17,"tags":408},[383,409,410],"音乐","摇滚",{"category":17,"tags":412},[383,392],{"category":17,"tags":414},[383,409],{"category":17,"tags":416},[383,35,392],{"category":17,"tags":418},[383,419,420],"同人","手书",{"category":17,"tags":422},[383,392],{"category":17,"tags":424},[383,392],{"category":17,"tags":426},[383,427],"植物",{"category":17,"tags":429},[383,430],"读书",{"category":17,"tags":432},[383,392],{"category":17,"tags":434},[383,392],{"category":17,"tags":436},[383,392],{"category":17,"tags":438},[383,392],{"category":17,"tags":440},[383,392],{"category":17,"tags":442},[383,392],{"category":17,"tags":444},[14,445,383],"一周年",{"category":17,"tags":447},[383,392],{"category":17,"tags":449},[383,392],{"category":17,"tags":451},[35,452,453],"开源组织","SITE-193",{"category":17,"tags":455},[383,392],{"category":17,"tags":457},[383,458],"年末",{"category":17,"tags":460},[383,392],{"category":17,"tags":462},[383,392],{"category":17,"tags":464},[14,17],{"category":17,"tags":466},[383,392],{"category":17,"tags":468},[383,392],{"category":17,"tags":470},[383,392],{"category":17,"tags":472},[383,392],{"category":17,"tags":474},[383,475,476],"游戏","Control",{"category":30,"tags":478},[479,480,481,482],"Quadim","图像处理","四叉树","Rust",{"category":17,"tags":484},[14,485],"上线纪念",{"category":17,"tags":487},[383,395],{"category":30,"tags":489},[203,490],"测试",{"category":17,"tags":492},[383],{"category":17,"tags":494},[383,392],{"category":17,"tags":496},[383],{"category":5,"tags":498},[40],{"category":5,"tags":500},[40],{"category":5,"tags":502},[40],{"category":5,"tags":504},[40],{"category":5,"tags":506},[40],[508,512,516,520,524,528,532,536,540,544,548,552,556,560,564,568,572,576,579,583,587,590,594,598,602,606,610,614,618,622,626,630,634,638,642,646,650,654,658,662,666,670,674,678,682,686,690,694,698,702,706,710,714,718,722,726,730,734,738,742,746,750,754,758,762,766,770,774,778,782,786,790,794,798,801,805,809,813,817,821,825,829,833,837,841,845,849,853,857,861,865,869,873,877,881,885,889,893,897,901,904,907,911,915,919,923,927,930,934,938,942,946,950,954,957,961,965,969,973,977,981,984,988,992,996,1000,1004,1007,1011,1015,1019,1023,1027,1031,1035,1039,1043,1047,1051,1055,1059,1063,1067,1071,1075,1079],{"path":509,"words":510,"published":511,"date":156},"\u002Fposts\u002F2024\u002F2024-06-25-liteloaderqqnt",1296,"2024-06-25 11:03:08",{"path":513,"words":514,"published":515,"date":156},"\u002Fposts\u002F2024\u002F2024-06-26-ip签名",0,"2024-06-26 22:53:53",{"path":517,"words":518,"published":519,"date":156},"\u002Fposts\u002F2024\u002F2024-06-29-何加盐中文互联网正在加速崩塌",4853,"2024-06-29 00:47:13",{"path":521,"words":522,"published":523,"date":156},"\u002Fposts\u002F2024\u002F2024-07-12-github-chart",270,"2024-07-12 20:02:50",{"path":525,"words":526,"published":527,"date":156},"\u002Fposts\u002F2024\u002F2024-07-14-访客数统计",192,"2024-07-24 01:34:28",{"path":529,"words":530,"published":531,"date":156},"\u002Fposts\u002F2024\u002F2024-07-26-广大选课脚本",110,"2024-07-26 01:34:28",{"path":533,"words":534,"published":535,"date":156},"\u002Fposts\u002F2024\u002F2024-07-26-数学分析笔记其二",14,"2024-07-26 23:54:02",{"path":537,"words":538,"published":539,"date":156},"\u002Fposts\u002F2024\u002F2024-07-26-数据结构复习其二",12856,"2024-07-26 02:48:08",{"path":541,"words":542,"published":543,"date":156},"\u002Fposts\u002F2024\u002F2024-07-26-数据结构复习相关",32856,"2024-07-26 02:24:14",{"path":545,"words":546,"published":547,"date":156},"\u002Fposts\u002F2024\u002F2024-07-26-线性代数与空间解析几何知识点全汇总",6110,"2024-07-26 03:27:55",{"path":549,"words":550,"published":551,"date":156},"\u002Fposts\u002F2024\u002F2024-07-31-typora-picgo-兰空图床打造markdown写作环境",311,"2024-07-31 12:56:23",{"path":553,"words":554,"published":555,"date":156},"\u002Fposts\u002F2024\u002F2024-09-02-jupyter配置中文",441,"2024-09-02 23:49:05",{"path":557,"words":558,"published":559,"date":156},"\u002Fposts\u002F2024\u002F2024-09-02-自定义vscode背景图片",100,"2024-09-02 23:39:00",{"path":561,"words":562,"published":563,"date":156},"\u002Fposts\u002F2024\u002F2024-09-03-制作gzhu校园网路由器",1423,"2024-09-03 23:43:18",{"path":565,"words":566,"published":567,"date":156},"\u002Fposts\u002F2024\u002F2024-09-26-09-25-2024",136,"2024-09-26 00:49:26",{"path":569,"words":570,"published":571,"date":156},"\u002Fposts\u002F2024\u002F2024-11-04-创建一个网络包解码器分析devp2p协议",1943,"2024-11-04 16:45:19",{"path":573,"words":574,"published":575,"date":156},"\u002Fposts\u002F2024\u002F2025-03-09-2024一些笔记常微分",16,"2024-12-09 22:51:00",{"path":577,"words":574,"published":578,"date":156},"\u002Fposts\u002F2024\u002F2025-03-09-2024一些笔记概率论","2025-03-09 22:51:00",{"path":580,"words":581,"published":582,"date":156},"\u002Fposts\u002F2025\u002F2025-03-01-windows下释放docker占用的wsl空间",164,"2025-03-01 13:54:15",{"path":584,"words":585,"published":586,"date":156},"\u002Fposts\u002F2025\u002F2025-03-01-爬虫实战-爬取广州大学课程表",774,"2025-03-01 14:55:48",{"path":588,"words":530,"published":589,"date":156},"\u002Fposts\u002F2025\u002F2025-04-03-hexo集成gitalk的问题","2025-04-03 14:26:34",{"path":591,"words":592,"published":593,"date":156},"\u002Fposts\u002F2025\u002F2025-09-22-记一次配置obsidian配合hexo写博客",92,"2025-09-22 16:47:52",{"path":595,"words":596,"published":597,"date":156},"\u002Fposts\u002F2025\u002F2025-10-24-10-20课题",204,"2025-10-24 08:15:01",{"path":599,"words":600,"published":601,"date":156},"\u002Fposts\u002F2025\u002F2025-11-02-组会朝花夕拾",501,"2025-11-02 18:56:59",{"path":603,"words":604,"published":605,"date":156},"\u002Fposts\u002F2025\u002F2025-11-05-基于随机加权推断模型im的参数估计算法",481,"2025-11-05 20:51:41",{"path":607,"words":608,"published":609,"date":156},"\u002Fposts\u002F2025\u002F2025-11-09-关于软考高项网规",884,"2025-11-09 19:54:52",{"path":611,"words":612,"published":613,"date":156},"\u002Fposts\u002F2025\u002F2025-11-17-arch-linux运行appimage相关",365,"2025-11-17 09:59:59",{"path":615,"words":616,"published":617,"date":156},"\u002Fposts\u002F2025\u002F2025-11-19-相关概念",2058,"2025-11-19 11:06:20",{"path":619,"words":620,"published":621,"date":156},"\u002Fposts\u002F2025\u002F2025-11-21-web可视化实践canvas",6393,"2025-11-21 14:50:48",{"path":623,"words":624,"published":625,"date":156},"\u002Fposts\u002F2025\u002F2025-11-22-使用-github-actions-自动部署基于vite的项目到-github-pages",272,"2025-11-23 01:59:24",{"path":627,"words":628,"published":629,"date":156},"\u002Fposts\u002F2025\u002F2025-11-22-关于前端包管理器npmpnpmyarn和bun",4890,"2025-11-22 21:21:15",{"path":631,"words":632,"published":633,"date":156},"\u002Fposts\u002F2025\u002F2025-11-23-undo-redo-机制具体实现",7839,"2025-11-23 19:15:00",{"path":635,"words":636,"published":637,"date":156},"\u002Fposts\u002F2025\u002F2025-11-27-前端安全-关于xss与crsf",1866,"2025-11-28 02:05:25",{"path":639,"words":640,"published":641,"date":156},"\u002Fposts\u002F2025\u002F2025-12-05-现代协同-2d-画布编辑器pixijs-v8",4285,"2025-12-05 02:00:16",{"path":643,"words":644,"published":645,"date":156},"\u002Fposts\u002F2025\u002F2025-12-07-杂谈-课题组系统设计",3710,"2025-12-07 21:09:33",{"path":647,"words":648,"published":649,"date":156},"\u002Fposts\u002F2025\u002F2025-12-25-javascript中的数组方法与栈stack和队列queue的实现",474,"2025-11-29 16:58:15",{"path":651,"words":652,"published":653,"date":156},"\u002Fposts\u002F2025\u002F2025-12-25-关于javascript-实现哈希表",189,"2025-11-27 11:31:39",{"path":655,"words":656,"published":657,"date":156},"\u002Fposts\u002F2025\u002F2025-12-25-关于javascript的set方法",332,"2025-12-25 15:26:18",{"path":659,"words":660,"published":661,"date":156},"\u002Fposts\u002F2025\u002F2025-12-25-关于动态规划(背包问题为例)",1238,"2025-11-17 18:24:04",{"path":663,"words":664,"published":665,"date":156},"\u002Fposts\u002F2025\u002F2025-12-25-关于滑动窗口",712,"2025-11-20 15:04:40",{"path":667,"words":668,"published":669,"date":156},"\u002Fposts\u002F2025\u002F2025-12-25-关于链表(javascript)",500,"2025-11-25 13:00:48",{"path":671,"words":672,"published":673,"date":156},"\u002Fposts\u002F2025\u002F2025-12-25-面试算法acm模式构建构建输入输出模板",2051,"2025-12-25 23:09:45",{"path":675,"words":676,"published":677,"date":156},"\u002Fposts\u002F2025\u002F2025-12-26-javascript-数字数组字符串的处理",483,"2025-12-26 11:20:25",{"path":679,"words":680,"published":681,"date":156},"\u002Fposts\u002F2025\u002F2025-12-26-javascripttypescript-的顺序表链表实现",446,"2025-12-26 12:16:12",{"path":683,"words":684,"published":685,"date":156},"\u002Fposts\u002F2025\u002F2025-12-27-前端画布设计vol-1-实现基础yuan素渲染和状态控制",2136,"2025-12-28 01:23:08",{"path":687,"words":688,"published":689,"date":156},"\u002Fposts\u002F2025\u002F2025-12-27-前端画布设计vol-2-实现富文本编辑",1838,"2025-12-28 02:58:47",{"path":691,"words":692,"published":693,"date":156},"\u002Fposts\u002F2025\u002F2025-12-27-算法刷题-关于链表操作",2135,"2025-12-27 16:12:42",{"path":695,"words":696,"published":697,"date":156},"\u002Fposts\u002F2025\u002F2025-12-27-配置dnscrypt-proxy实现加密dns服务windows",1899,"2025-11-15 01:06:37",{"path":699,"words":700,"published":701,"date":156},"\u002Fposts\u002F2025\u002F2025-12-28-前端-关于网络请求xhrajaxfetchaxios",7788,"2025-10-28 15:48:46",{"path":703,"words":704,"published":705,"date":156},"\u002Fposts\u002F2025\u002F2025-12-28-前端-接口封装与请求规范axios为例",2373,"2025-12-29 01:01:18",{"path":707,"words":708,"published":709,"date":156},"\u002Fposts\u002F2025\u002F2025-12-28-前端-身份验证管理-基于-jwt-token-的实现",2709,"2025-11-23 22:04:41",{"path":711,"words":712,"published":713,"date":156},"\u002Fposts\u002F2025\u002F2025-12-29-记canvas画布项目开发",6254,"2025-12-29 04:52:15",{"path":715,"words":716,"published":717,"date":156},"\u002Fposts\u002F2025\u002Fabout-site",116,"2025-03-31 16:44:24",{"path":719,"words":720,"published":721,"date":156},"\u002Fposts\u002F2025\u002F使用neo4j图数据科学库gds进行中心度分析",143,"2025-04-02 23:37:24",{"path":723,"words":724,"published":725,"date":156},"\u002Fposts\u002F2026\u002F2026-01-03-论文实训草稿",10445,"2026-01-03 14:10:49",{"path":727,"words":728,"published":729,"date":156},"\u002Fposts\u002F2026\u002F2026-01-06-人工智能导论",11461,"2026-01-07 03:25:49",{"path":731,"words":732,"published":733,"date":156},"\u002Fposts\u002F2026\u002F2026-01-07-机器学习相关算法",82,"2026-01-08 02:28:13",{"path":735,"words":736,"published":737,"date":156},"\u002Fposts\u002F2026\u002F2026-01-16-关于低复杂度最长链共识算法设计",4466,"2026-01-16 10:33:23",{"path":739,"words":740,"published":741,"date":156},"\u002Fposts\u002F2026\u002F2026-01-19-杂记2026-01-19",5754,"2026-01-20 01:03:26",{"path":743,"words":744,"published":745,"date":156},"\u002Fposts\u002F2026\u002F2026-01-22-github-action-自动同步博客到-github-主页",1069,"2026-01-22 16:31:47",{"path":747,"words":748,"published":749,"date":156},"\u002Fposts\u002F2026\u002F2026-03-01-windows-wsl安装docker",257,"2026-03-01 20:55:31",{"path":751,"words":752,"published":753,"date":156},"\u002Fposts\u002F2026\u002F2026-03-14-关于-msys2",701,"2026-03-15 03:05:04",{"path":755,"words":756,"published":757,"date":156},"\u002Fposts\u002F2026\u002F2026-03-28-团队项目协作规范随记",2419,"2026-03-29 05:33:50",{"path":759,"words":760,"published":761,"date":156},"\u002Fposts\u002F2026\u002F2026-03-29-oh-my-posh配置分享",104,"2026-03-29 16:34:43",{"path":763,"words":764,"published":765,"date":156},"\u002Fposts\u002F2026\u002F2026-06-13-从上下文工程到-agent-harness-engineering",8840,"2026-06-13 21:06:20",{"path":767,"words":768,"published":769,"date":156},"\u002Fposts\u002F2026\u002F2026-07-13-简谈前端基建质量标准与ai友好建设",3814,"2026-07-13 21:06:20",{"path":771,"words":772,"published":773,"date":156},"\u002Fposts\u002F2026\u002F2026-07-15-agentic-rag-系统实践",13357,"2026-07-15 21:06:20",{"path":775,"words":776,"published":777,"date":156},"\u002Fposts\u002F2026\u002F2026-07-18-简谈sdd规范驱动开发-copy",4476,"2026-07-18 11:06:20",{"path":779,"words":780,"published":781,"date":156},"\u002Fposts\u002F2026\u002F2026-07-20-个人neovim配置分享",366,"2026-07-22 13:46:50",{"path":783,"words":784,"published":785,"date":156},"\u002Fposts\u002F2026\u002F2026-07-20-关于技术写作-设计文档",6763,"2026-07-20 13:46:50",{"path":787,"words":788,"published":789,"date":156},"\u002Fposts\u002F2026\u002F2026-07-29-fastapi-工程实践随记",3417,"2026-07-29 04:58:34",{"path":791,"words":792,"published":793,"date":156},"\u002Fposts\u002Frfc_agentic_project\u002F2026-06-19-rfc-liskin_code_agent",3563,"2026-06-19 15:18:20",{"path":795,"words":796,"published":797,"date":156},"\u002Fposts\u002Frfc_agentic_project\u002F2026-07-25-rfc-glushkov_ogas",13461,"2026-07-28 23:46:50",{"path":799,"words":800,"published":793,"date":156},"\u002Fposts\u002Frfc_agentic_project\u002F2026-08-03-adr-ogas-arkhiv",2698,{"path":802,"words":803,"published":804,"date":156},"\u002Fposts\u002Frfc_agentic_project\u002Frag\u002F2026-08-12-arkhiv_rag-vol-1",4366,"2026-08-12 15:18:20",{"path":806,"words":807,"published":808,"date":156},"\u002Fposts\u002Fleetcode\u002F2025-12-25-力扣百题速练javascripttypescriptvol-1",3108,"2025-12-25 13:57:08",{"path":810,"words":811,"published":812,"date":156},"\u002Fposts\u002Fleetcode\u002F2025-12-26-力扣百题速练javascripttypescriptvol-2",3185,"2025-12-26 13:57:08",{"path":814,"words":815,"published":816,"date":156},"\u002Fposts\u002Fleetcode\u002F2025-12-28-力扣百题速练javascripttypescriptvol-3",2533,"2025-12-29 02:15:44",{"path":818,"words":819,"published":820,"date":156},"\u002Fposts\u002Fleetcode\u002F2026-01-17-力扣百题速练javascripttypescriptvol-4",3024,"2026-01-17 22:53:01",{"path":822,"words":823,"published":824,"date":156},"\u002Fposts\u002Fleetcode\u002F2026-01-19-力扣百题速练javascripttypescriptvol-5",293,"2026-01-19 15:15:12",{"path":826,"words":827,"published":828,"date":156},"\u002Fposts\u002F分布式处理与计算\u002F分布式处理与计算spark系统与scala语言",4320,"2026-07-06 14:10:49",{"path":830,"words":831,"published":832,"date":156},"\u002Fposts\u002F分布式处理与计算\u002F分布式处理与计算分类与回归分析",2081,"2026-07-07 12:40:29",{"path":834,"words":835,"published":836,"date":156},"\u002Fposts\u002F分布式处理与计算\u002F分布式处理与计算数值优化方法",2182,"2026-07-07 13:50:29",{"path":838,"words":839,"published":840,"date":156},"\u002Fposts\u002F分布式处理与计算\u002F分布式处理与计算数据降维",1583,"2026-07-07 21:50:29",{"path":842,"words":843,"published":844,"date":156},"\u002Fposts\u002F分布式处理与计算\u002F分布式处理与计算聚类分析",1882,"2026-07-07 16:50:29",{"path":846,"words":847,"published":848,"date":156},"\u002Fposts\u002F分布式处理与计算\u002F分布式处理与计算随机模拟与统计推断",2433,"2026-07-06 22:40:29",{"path":850,"words":851,"published":852,"date":156},"\u002Fposts\u002F数据可视化\u002F2026-06-14-数据可视化-基本图像",10011,"2026-06-14 20:18:20",{"path":854,"words":855,"published":856,"date":156},"\u002Fposts\u002F数据可视化\u002F2026-06-14-数据可视化-概论",10030,"2026-06-14 15:18:20",{"path":858,"words":859,"published":860,"date":156},"\u002Fposts\u002F数据可视化\u002F2026-06-15-数据可视化-可视化库",2032,"2026-06-15 20:18:20",{"path":862,"words":863,"published":864,"date":156},"\u002Fposts\u002F数据结构\u002F2025-09-29-数据结构-图",5451,"2025-04-15 20:48:06",{"path":866,"words":867,"published":868,"date":156},"\u002Fposts\u002F数据结构\u002F2025-09-29-数据结构-树",4211,"2025-04-30 02:19:06",{"path":870,"words":871,"published":872,"date":156},"\u002Fposts\u002F数据结构\u002F2025-09-29-数据结构-队列",921,"2025-04-11 15:30:08",{"path":874,"words":875,"published":876,"date":156},"\u002Fposts\u002F数据结构\u002F关于数据结构的一些想法",456,"2025-04-04 19:13:49",{"path":878,"words":879,"published":880,"date":156},"\u002Fposts\u002F数据结构\u002F数据结构-栈",920,"2025-04-18 00:17:37",{"path":882,"words":883,"published":884,"date":156},"\u002Fposts\u002F数据结构\u002F数据结构-算法复杂度",2153,"2025-04-05 16:09:55",{"path":886,"words":887,"published":888,"date":156},"\u002Fposts\u002F数据结构\u002F数据结构-线性表链表",4572,"2025-04-06 00:24:40",{"path":890,"words":891,"published":892,"date":156},"\u002Fposts\u002F数据结构\u002F数据结构-线性表顺序表",4191,"2025-04-05 22:09:59",{"path":894,"words":895,"published":896,"date":156},"\u002Fposts\u002F数据结构\u002F数据结构-绪论",2881,"2025-04-05 11:31:22",{"path":898,"words":899,"published":156,"date":900},"\u002Fposts\u002F杂记\u002F2023-12-26-2023年12月26日",7,"2023-12-26 23:51:33",{"path":902,"words":899,"published":156,"date":903},"\u002Fposts\u002F杂记\u002F2024-01-01-2024团建","2024-01-01 01:51:27",{"path":905,"words":581,"published":156,"date":906},"\u002Fposts\u002F杂记\u002F2024-01-09-随记","2024-01-09 20:50:35",{"path":908,"words":909,"published":156,"date":910},"\u002Fposts\u002F杂记\u002F2024-03-12-花名",5275,"2024-03-12 16:04:23",{"path":912,"words":913,"published":156,"date":914},"\u002Fposts\u002F杂记\u002F2024-04-15-逆模因",1353,"2024-04-15 00:26:16",{"path":916,"words":917,"published":156,"date":918},"\u002Fposts\u002F杂记\u002F2024-05-03-艺作",27,"2024-05-03 00:01:00",{"path":920,"words":921,"published":156,"date":922},"\u002Fposts\u002F杂记\u002F2024-05-21-关于博客",1720,"2024-02-21 23:21:11",{"path":924,"words":925,"published":156,"date":926},"\u002Fposts\u002F杂记\u002F2024-06-20-dnvrc维克多崔专访-2023",2036,"2024-06-20 00:28:39",{"path":928,"words":760,"published":156,"date":929},"\u002Fposts\u002F杂记\u002F2024-06-25-6月25日随记","2024-06-25 12:46:36",{"path":931,"words":932,"published":156,"date":933},"\u002Fposts\u002F杂记\u002F2024-06-25-歌名-想見星星",148,"2024-06-25 22:47:59",{"path":935,"words":936,"published":156,"date":937},"\u002Fposts\u002F杂记\u002F2024-07-11-桂花岗",6,"2024-07-11 00:28:17",{"path":939,"words":940,"published":156,"date":941},"\u002Fposts\u002F杂记\u002F2024-07-15-rolling-bocci",9,"2024-07-15 21:25:56",{"path":943,"words":944,"published":156,"date":945},"\u002Fposts\u002F杂记\u002F2024-07-22-七月九日",34,"2024-07-09 17:54:23",{"path":947,"words":948,"published":156,"date":949},"\u002Fposts\u002F杂记\u002F2024-07-22-六月随记",1052,"2024-07-22 01:03:19",{"path":951,"words":952,"published":156,"date":953},"\u002Fposts\u002F杂记\u002F2024-07-22-石菖蒲",41,"2024-07-10 01:26:05",{"path":955,"words":514,"published":156,"date":956},"\u002Fposts\u002F杂记\u002F2024-07-23-新书","2024-07-23 01:05:49",{"path":958,"words":959,"published":156,"date":960},"\u002Fposts\u002F杂记\u002F2024-07-29-年前",74,"2024-07-29 23:59:07",{"path":962,"words":963,"published":156,"date":964},"\u002Fposts\u002F杂记\u002F2024-07-30-周边",138,"2024-07-30 02:03:23",{"path":966,"words":967,"published":156,"date":968},"\u002Fposts\u002F杂记\u002F2024-07-31-2024-7-31日",185,"2024-07-31 01:21:52",{"path":970,"words":971,"published":156,"date":972},"\u002Fposts\u002F杂记\u002F2024-08-02-八月",1710,"2024-08-30 00:31:02",{"path":974,"words":975,"published":156,"date":976},"\u002Fposts\u002F杂记\u002F2024-09-01-九月",1484,"2024-09-26 16:30:15",{"path":978,"words":979,"published":156,"date":980},"\u002Fposts\u002F杂记\u002F2024-09-08-2024年9月8日",97,"2024-09-08 13:59:36",{"path":982,"words":514,"published":156,"date":983},"\u002Fposts\u002F杂记\u002F2024-09-19-博客站的一周年","2024-09-18 23:36:24",{"path":985,"words":986,"published":156,"date":987},"\u002Fposts\u002F杂记\u002F2024-10-03-十月",1028,"2024-10-31 23:54:27",{"path":989,"words":990,"published":156,"date":991},"\u002Fposts\u002F杂记\u002F2024-11-04-十一月",2134,"2024-11-30 00:04:39",{"path":993,"words":994,"published":156,"date":995},"\u002Fposts\u002F杂记\u002F2024-11-10-gzhu-site-193",80,"2024-11-10 19:19:12",{"path":997,"words":998,"published":156,"date":999},"\u002Fposts\u002F杂记\u002F2024-12-01-十二月",455,"2024-12-30 23:13:47",{"path":1001,"words":1002,"published":156,"date":1003},"\u002Fposts\u002F杂记\u002F2024-12-21-年末",23,"2024-12-21 23:21:49",{"path":1005,"words":522,"published":156,"date":1006},"\u002Fposts\u002F杂记\u002F2025-01-30-一月","2025-01-30 23:44:41",{"path":1008,"words":1009,"published":156,"date":1010},"\u002Fposts\u002F杂记\u002F2025-03-01-二月",224,"2025-02-28 00:11:08",{"path":1012,"words":1013,"published":156,"date":1014},"\u002Fposts\u002F杂记\u002F2025-03-01-博客杂谈",344,"2025-03-01 16:03:18",{"path":1016,"words":1017,"published":156,"date":1018},"\u002Fposts\u002F杂记\u002F2025-03-05-三月",2090,"2025-03-05 19:17:28",{"path":1020,"words":1021,"published":156,"date":1022},"\u002Fposts\u002F杂记\u002F2025-07-10-七月",361,"2025-07-10 22:04:32",{"path":1024,"words":1025,"published":156,"date":1026},"\u002Fposts\u002F杂记\u002F2025-09-21-九月",781,"2025-09-21 20:53:37",{"path":1028,"words":1029,"published":156,"date":1030},"\u002Fposts\u002F杂记\u002F2025-10-09-十月",591,"2025-10-09 00:35:07",{"path":1032,"words":1033,"published":156,"date":1034},"\u002Fposts\u002F杂记\u002Fcontrol",13,"2023-09-25 13:03:59",{"path":1036,"words":1037,"published":156,"date":1038},"\u002Fposts\u002F杂记\u002Fquadim四叉树图像处理",1094,"2023-10-21 11:27:06",{"path":1040,"words":1041,"published":156,"date":1042},"\u002Fposts\u002F杂记\u002Fsuccess",24,"2023-09-22 15:59:07",{"path":1044,"words":1045,"published":156,"date":1046},"\u002Fposts\u002F杂记\u002F三光年",295,"2023-11-18 04:18:40",{"path":1048,"words":1049,"published":156,"date":1050},"\u002Fposts\u002F杂记\u002F加密测试",961,"2023-11-21 14:24:01",{"path":1052,"words":1053,"published":156,"date":1054},"\u002Fposts\u002F杂记\u002F啊啊啊啊啊啊啊",54,"2023-09-20 17:29:33",{"path":1056,"words":1057,"published":156,"date":1058},"\u002Fposts\u002F杂记\u002F杂谈",124,"2023-09-25 23:11:32",{"path":1060,"words":1061,"published":156,"date":1062},"\u002Fposts\u002F杂记\u002F艹",25,"2023-10-21 17:40:43",{"path":1064,"words":1065,"published":1066,"date":156},"\u002Fposts\u002F神经网络\u002F神经网络与深度学习卷积神经网络",3168,"2026-06-07 11:23:08",{"path":1068,"words":1069,"published":1070,"date":156},"\u002Fposts\u002F神经网络\u002F神经网络与深度学习循环神经网络",2348,"2026-06-07 11:33:08",{"path":1072,"words":1073,"published":1074,"date":156},"\u002Fposts\u002F神经网络\u002F神经网络与深度学习改进学习方法",3360,"2026-06-07 10:13:08",{"path":1076,"words":1077,"published":1078,"date":156},"\u002Fposts\u002F神经网络\u002F神经网络与深度学习深度生成模型",2727,"2026-06-07 11:43:08",{"path":1080,"words":1081,"published":1082,"date":156},"\u002Fposts\u002F神经网络\u002F神经网络与深度学习神经网络基础",3205,"2026-06-07 10:03:08",{"id":1084,"title":1085,"abbrlink":1086,"body":1087,"category":12,"cover":156,"date":156,"description":1093,"extension":3140,"mathjax":3141,"meta":3142,"navigation":3141,"path":723,"published":725,"readingMinutes":3145,"seo":3146,"stem":3147,"sticky":156,"swiper_index":156,"tags":3148,"updated":156,"words":724,"__hash__":3149},"posts\u002Fposts\u002F2026\u002F2026-01-03-论文实训草稿.md","2026-01-03-基于传统机器学习技术构建欺诈识别系统","38065",{"type":1088,"value":1089,"toc":3102},"minimark",[1090,1094,1097,1100,1105,1117,1136,1151,1161,1179,1182,1210,1213,1217,1220,1225,1232,1242,1246,1249,1256,1282,1286,1292,1303,1307,1317,1335,1339,1350,1353,1357,1360,1365,1375,1378,1381,1386,1389,1392,1395,1398,1401,1405,1408,1411,1415,1418,1422,1425,1429,1438,1453,1456,1460,1464,1467,1485,1489,1492,1500,1756,1759,1767,1923,1926,1934,2090,2093,2101,2291,2294,2308,2310,2312,2316,2319,2323,2326,2329,2332,2336,2362,2365,2529,2544,2547,2551,2554,2559,2562,2565,2568,2572,2575,2580,2594,2597,2600,2605,2637,2640,2644,2647,2713,2718,2721,2725,2728,2787,2792,2795,2799,2802,2805,2809,2813,2816,2819,2830,2833,2837,2840,2843,2847,2850,2861,2865,2868,2871,2945,2947,2949,3035],[1091,1092,1093],"p",{},"汽车保险领域面临的日益严峻的欺诈风险使行业经济损失巨大。为有效应对此类问题，本研究运用传统机器学习技术构建欺诈识别系统，提供了数据处理、特征提取以及模型构建的完整流程，采用汽车保险理赔数据集，进行相关实验，重点考察随机森林算法与其他分类器的性能对比，包括数据清洗、特征选择和不平衡处理。通过交叉验证和指标评估，结果显示优化后的随机森林模型在 AUC 和召回率方面表现出色。该研究验证了传统算法在实际场景中的可靠性，并为保险企业风险管理提供实用建议。",[1091,1095,1096],{},"\\ckeywords{汽车保险欺诈；随机森林算法；传统机器学习；数据不平衡；特征提取；性能评估}",[1098,1099],"hr",{},[1101,1102,1104],"h1",{"id":1103},"_1","1",[1091,1106,1107,1108,1112,1113,1116],{},"保险业通过风险聚合与转移机制为社会经济活动提供安全保障，在全球经济运行中扮演着至关重要的角色。其庞大的资金池更是资本市场长期资本的重要来源之一 ",[1109,1110,1111],"strong",{},"(Barry & Charpentier, 2020)","。根据经济合作与发展组织的报告，保险业的稳定性直接关系到全球金融系统的韧性 ",[1109,1114,1115],{},"(OECD, 2023)","。在财产保险领域，汽车保险覆盖面最为广泛并与日常生活紧密联系，构成了该领域的核心业务板块，同时其受到日益猖獗的保险欺诈行为的严峻挑战。",[1091,1118,1119,1120,1123,1124,1127,1128,1131,1132,1135],{},"汽车保险欺诈已成为一个全球性的顽疾，对行业的财务健康和社会的诚信体系造成持续性的损害。欺诈行为导致保险公司支付了本不应承担的赔款，这些巨大的“渗漏”最终会通过提高保费的形式转嫁给所有诚实投保人，破坏了保险的公平性原则 ",[1109,1121,1122],{},"(Viaene & Dedene, 2004)","。据美国反保险欺诈联盟（Coalition Against Insurance Fraud）的最新报告，保险欺诈每年给美国造成的损失已超过 3000 亿美元，其中车险领域是重灾区 ",[1109,1125,1126],{},"(CAIF, 2022)","。欺诈的成因复杂，一方面，信息不对称使得保险公司难以在承保和理赔环节完全掌握投保人的真实风险与行为 ",[1109,1129,1130],{},"(Cohen & Siegelman, 2010)","；另一方面，技术的进步，特别是数字化理赔流程的普及，在提升效率的同时，也为新型、更隐蔽的欺诈手段提供了可乘之机 ",[1109,1133,1134],{},"(Severino & Peng, 2021)","。这不仅侵蚀了保险公司的承保利润，还可能导致定价模型失真，扭曲风险信号，长期而言将削弱保险的风险分担功能和社会效益。",[1091,1137,1138,1139,1142,1143,1146,1147,1150],{},"从广义上讲，保险欺诈是指任何以非法获取保险金为目的的故意行为。根据欺诈主体的不同，可分为保单持有人欺诈、第三方欺诈以及内部人员欺诈等 ",[1109,1140,1141],{},"(Derrig, 2002)","。鉴于数据的可获得性与研究的可操作性，本文的研究焦点将集中于",[1109,1144,1145],{},"汽车理赔欺诈","，即保单持有人或相关方在理赔环节，通过故意制造事故、夸大损失、伪造单据等手段骗取保险赔偿金的行为。这类欺诈是车险欺诈中最常见的形式，拥有相对丰富的公开研究数据基础，是应用数据驱动方法进行自动化检测的主要战场 ",[1109,1148,1149],{},"(Bhattacharyya et al., 2011)","。",[1091,1152,1153,1154,1157,1158,1150],{},"为应对这类欺诈威胁，保险公司正从依赖专家规则和人工审核，转向基于数据挖掘与机器学习（ML）的自动化检测系统。传统的规则引擎虽然解释性强，但难以捕捉复杂的非线性关系和新型欺诈模式。机器学习，特别是监督学习算法，能够从历史理赔数据中自动学习欺诈模式，展现出巨大潜力。在众多机器学习方法中，集成学习因其卓越的预测性能和鲁棒性而备受关注。以随机森林（Random Forest）为代表的集成算法，通过构建多棵决策树并综合其结果，能有效缓解单棵树的过拟合问题，对高维特征和非线性关系有良好的处理能力 ",[1109,1155,1156],{},"(Breiman, 2001; Polikar, 2012)","。更重要的是，保险欺诈数据天然具有高度不平衡性（正常理赔远多于欺诈理赔），而随机森林通过自助采样（Bootstrap sampling）和随机特征子空间选择，能在不均衡数据上构建多样化的基分类器，从而在一定程度上提升对少数类（欺诈）样本的识别能力 ",[1109,1159,1160],{},"(Xia et al., 2023; Phua et al., 2010)",[1091,1162,1163,1164,1167,1168,1171,1172,1175,1176,1150],{},"尽管机器学习在欺诈检测中的应用已取得丰硕成果，但现有研究仍存在一些有待深化之处。许多研究侧重于单一高级分类器（如 XGBoost、深度神经网络）的性能比拼，而相对忽视了",[1109,1165,1166],{},"数据预处理阶段与分类模型的系统性整合与优化","。特征工程、处理类别不平衡的重采样技术（如 SMOTE、ADASYN）以及特征选择，对于最终模型性能的影响至关重要，有时甚至不亚于分类器本身的选择 ",[1109,1169,1170],{},"(Wang et al., 2021; Chawla et al., 2002)","。此外，在",[1109,1173,1174],{},"真实的汽车保险公开数据集","上，对包含预处理流程在内的多种传统机器学习算法进行端到端的、公平的对比实验研究相对有限，特别是深入探讨不同预处理技术如何与不同算法交互以提升欺诈检测性能的研究尚不充分 ",[1109,1177,1178],{},"(He & Garcia, 2009; Gomes et al., 2021)",[1091,1180,1181],{},"本文旨在探索并验证一套结合预处理技术与经典机器学习算法的汽车保险欺诈检测框架。具体而言，本研究将在公开的汽车保险理赔数据集上，以随机森林算法为核心检测模型，系统性地集成多种特征编码、不平衡数据处理（如过采样与欠采样）及特征选择方法，构建一个完整的分析管道。通过设计详尽的对比实验，本文将评估该集成框架相对于单一模型及其他主流机器学习算法（如逻辑回归、支持向量机、XGBoost）在欺诈检测准确率、召回率、F1 分数等关键指标上的性能表现，从而为构建高效、实用的车险欺诈检测系统提供实证依据。主要要点如下：",[1183,1184,1185,1192,1198,1204],"ol",{},[1186,1187,1188,1191],"li",{},[1109,1189,1190],{},"提出一个基于机器学习的欺诈检测分析框架","：将数据处理、特征工程、不平衡学习、特征选择与随机森林分类器进行有机整合，形成了一个可复现、可评估的完整机器学习工作流，强调了预处理环节在模型构建中的基础性地位。",[1186,1193,1194,1197],{},[1109,1195,1196],{},"进行了相关的算法对比实验","：在公开基准数据集上，对包括随机森林在内的多种传统机器学习算法，在统一的预处理标准和评估指标下进行了性能对比与分析，为算法选择提供了实证参考。",[1186,1199,1200,1203],{},[1109,1201,1202],{},"深入探讨了不平衡数据处理策略的有效性","：实证检验了多种重采样技术在缓解保险欺诈数据类不平衡问题上的作用，并分析了其与不同分类器结合时的性能变化规律。",[1186,1205,1206,1209],{},[1109,1207,1208],{},"提供了结构化的方法学实现","：研究过程注重方法论的清晰描述与代码的结构化，确保了实验的可复现性，为后续研究者提供了可直接借鉴的技术路径和比较基线。",[1091,1211,1212],{},"本文结构安排如下：第 2 章将对保险欺诈检测，特别是基于机器学习的检测方法的相关文献进行综述；第 3 章将详细阐述本文所采用的研究方法，包括数据集描述、预处理技术、特征工程、使用的机器学习算法以及实验设计；第 4 章将展示并分析实验结果，对不同模型和策略的性能进行对比与讨论；第 5 章将总结全文，概括主要研究发现，指出本研究的局限性，并对未来研究方向提出展望。",[1101,1214,1216],{"id":1215},"_2","2",[1091,1218,1219],{},"本章旨在回顾保险欺诈检测方法的技术演进，梳理传统机器学习算法在该领域的应用现状、优势与挑战，明确本研究的理论背景与创新点。",[1221,1222,1224],"h2",{"id":1223},"_21-欺诈检测方法的演进","2.1 欺诈检测方法的演进",[1091,1226,1227,1228,1231],{},"保险欺诈检测的演进与信息技术的发展脉络高度吻合。早期的检测完全依赖具有丰富经验的核保员，主观性极强。随后，基于规则的专家系统（Expert Systems）通过将反欺诈知识编码为“IF-THEN”逻辑，实现了初步自动化 ",[1109,1229,1230],{},"(Bentley, 2000)","。然而，该系统难以应对高度协同且隐蔽的新型欺诈模式。",[1091,1233,1234,1235,1238,1239,1150],{},"20 世纪 90 年代末，统计方法被引入该领域。逻辑回归（Logistic Regression）因其能够量化风险概率而成为早期的主流工具 ",[1109,1236,1237],{},"(Brockett et al., 2002)","。进入 21 世纪后，数据挖掘与机器学习范式确立，研究焦点从简单的线性模型转向能够处理高维、非线性关系的复杂模型。现代欺诈检测已形成以监督学习为主、异常检测与图神经网络为辅的多维检测体系 ",[1109,1240,1241],{},"(Phua et al., 2010; West & Bhattacharya, 2016)",[1221,1243,1245],{"id":1244},"_22-传统机器学习算法在保险欺诈检测中的应用","2.2 传统机器学习算法在保险欺诈检测中的应用",[1091,1247,1248],{},"在众多机器学习算法中，传统（或经典）算法因其计算效率高、在中小规模数据集上表现稳健且易于部署，至今仍是行业应用的核心。",[1091,1250,1251,1252,1255],{},"单一分类器如决策树（DT）、支持向量机（SVM）和朴素贝叶斯在早期研究中应用广泛。决策树因其天然的解释性被青睐，但单棵树极易产生过拟合 ",[1109,1253,1254],{},"(Baesens et al., 2015)","。SVM 通过核函数处理高维非线性空间，但在大规模数据集上的训练开销较大。",[1091,1257,1258,1259,1262,1263,1150,1266,1269,1270,1273,1274,1277,1278,1281],{},"为克服单一模型的局限，集成学习（Ensemble Learning）成为该领域的最优解。",[1109,1260,1261],{},"随机森林（Random Forest, RF）"," 通过自助采样（Bootstrap）和随机特征选择构建决策树森林，能显著提升模型的泛化能力和抗噪性能 ",[1109,1264,1265],{},"(Breiman, 2001; Sahin et al., 2013)",[1109,1267,1268],{},"Itri 等人 (2019)"," 在汽车保险数据集上的对比实验表明，随机森林在 AUC 和准确率上始终处于领先地位。此外，以 ",[1109,1271,1272],{},"XGBoost (Chen & Guestrin, 2016)"," 和 ",[1109,1275,1276],{},"LightGBM (Ke et al., 2017)"," 为代表的梯度提升框架，通过迭代修正残差，进一步刷新了欺诈检测的精度上限。",[1109,1279,1280],{},"Jovanovic 等人 (2022)"," 的研究证实，经过精细参数调优的提升树模型在处理车险理赔数据时，其召回率（Recall）显著优于传统线性模型。",[1221,1283,1285],{"id":1284},"_23-多模型比较研究与算法选择策略","2.3 多模型比较研究与算法选择策略",[1091,1287,1288,1289,1150],{},"由于保险欺诈模式在不同地区和产品线之间存在显著差异，不存在“一劳永逸”的通用算法，因此多模型对比（Benchmarking）成为该领域实证研究的标准范式 ",[1109,1290,1291],{},"(Lessmann et al., 2015)",[1091,1293,1294,1295,1298,1299,1302],{},"多数实证研究（如 ",[1109,1296,1297],{},"Nordin et al., 2024","）表明，基于树的集成方法（RF, XGBoost, CatBoost）在整体性能上优于逻辑回归和感知器。由于欺诈检测是一个典型的非平衡代价问题，评估指标已从单纯的准确率转向更具代表性的 AUC-ROC、F1 分数以及在特定误报率下的召回率 ",[1109,1300,1301],{},"(Dal Pozzolo et al., 2015)","。然而，目前的对比研究大多聚焦于算法本身，对于预处理技术（如不同编码方式）与分类器之间的交互影响缺乏系统性探讨。",[1221,1304,1306],{"id":1305},"_24-数据不平衡与特征工程","2.4 数据不平衡与特征工程",[1091,1308,1309,1310,1313,1314,1150],{},"保险欺诈检测本质上属于典型的类别不平衡分类问题，在本数据集中，欺诈样本仅占约 24.7%，这种分布会导致模型在训练过程中过度偏向多数类（非欺诈），从而牺牲了对少数类（欺诈）的捕获能力 ",[1109,1311,1312],{},"(He & Garcia, 2009)","。在保险理赔实务中，分类错误的成本是非对称的：漏检一个真实的欺诈案例（假阴性，False Negative）所导致的直接经济赔付损失，通常远高于对正常理赔进行额外审核所产生的行政成本（假阳性，False Positive） ",[1109,1315,1316],{},"(Dal Pozzolo et al., 2015; Severino & Peng, 2021)",[1091,1318,1319,1320,1323,1324,1327,1328,1331,1332,1150],{},"针对这一挑战，本研究引入了",[1109,1321,1322],{},"合成少数类过采样技术（Synthetic Minority Over-sampling Technique, SMOTE）"," ",[1109,1325,1326],{},"(Chawla et al., 2002)","。与简单的随机过采样（Random Over-sampling）不同，SMOTE 通过在少数类样本及其最近邻之间进行线性插值来生成“合成”样本，从而有效地扩展了少数类的决策区域，并缓解了简单复制样本导致的过拟合风险 ",[1109,1329,1330],{},"(Fernández et al., 2018)","。通过应用 SMOTE 算法，本研究将训练集的正负样本比例调整至 1:1，使欺诈样本的表征更加充分，从而显著增强了模型对欺诈模式的学习效率和泛化性能 ",[1109,1333,1334],{},"(Jovanovic et al., 2022)",[1221,1336,1338],{"id":1337},"_25-可进一步探索的空间","2.5 可进一步探索的空间",[1091,1340,1341,1342,1345,1346,1349],{},"综上所述，尽管基于传统机器学习的欺诈检测已相对成熟，但仍存在以下待解决的问题：",[1343,1344],"br",{},"\n现有的研究往往将预处理、不平衡处理与分类器训练割裂开来，缺乏一个",[1109,1347,1348],{},"高度整合、可复现且端到端的分析框架","。此外，虽然深度学习（如 TabNet）开始崭露头角，但传统算法在解释性要求较高的保险合规场景下仍具有不可替代的价值。本文将基于公开的车险数据集，系统验证一套集成了先进重采样策略与随机森林分类器的优化框架，旨在为该领域的工程实践提供清晰的方法论基准。",[1091,1351,1352],{},"本文将以一个公开的汽车保险理赔数据集为基础，构建一个系统化的分析流程，核心内容包括：（1）实施一套完整的预处理与特征工程方案；（2）在公平的实验设置下，系统对比包括逻辑回归、支持向量机、K 近邻、决策树、随机森林、AdaBoost 和 XGBoost 在内的七种传统机器学习算法的性能；（3）深入探讨随机森林算法在该任务中的优势及其原因；（4）实证分析不同的类不平衡处理策略对关键分类器性能的影响。本研究期望通过这些工作，为汽车保险欺诈检测的模型选择与工程实践提供一份实证参考与方法论范例。",[1101,1354,1356],{"id":1355},"_3","3",[1091,1358,1359],{},"本章详细阐述本研究的实验框架，包括数据集来源、数据预处理、类不平衡处理、分类算法选择以及实验设计。",[1361,1362,1364],"h3",{"id":1363},"_31-数据集描述","3.1 数据集描述",[1091,1366,1367,1368,1374],{},"本研究采用公开的汽车保险理赔数据集，该数据集源于 Kaggle 平台上的“Vehicle Insurance Claim Fraud Detection” ",[1369,1370,1371],"a",{"href":1371,"rel":1372},"https:\u002F\u002Fwww.kaggle.com\u002Fdatasets\u002Fbuntyshah\u002Fauto-insurance-claims-data\u002Fcode",[1373],"nofollow"," ，该数据集包含 1000 行数据，每一行代表一个理赔案例。包含 40 个特征（如保单信息、事故地点、车辆型号等）",[1091,1376,1377],{},"其目标变量为“fraud_reported”，取值为“Y”（欺诈）或“N”（非欺诈），属于典型的二分类任务。其中，欺诈样本（Y）数量为 247 条，非欺诈样本（N）数量为 753 条，欺诈样本占比约为 24.7%。这一类别不平衡现象符合保险行业中欺诈案例的特征。",[1091,1379,1380],{},"数据集特征涵盖客户个人信息、保单细节、事故信息以及理赔金额等维度。数值型特征如 total_claim_amount 在欺诈与非欺诈样本间存在明显差异，而类别型特征如 incident_severity 与欺诈标签的相关性较高。这些初步观察为后续特征工程提供了重要的依据。",[1382,1383,1385],"h4",{"id":1384},"_32-数据预处理","3.2 数据预处理",[1091,1387,1388],{},"为确保数据质量并满足模型输入要求，本研究实施了一系列数据预处理：",[1091,1390,1391],{},"缺失值处理：数据集部分特征（如 collision_type、property_damage、police_report_available）以“？”标记缺失值。本研究将“？”替换为 NaN，并采用众数填充策略（Mode Imputation）进行处理。",[1091,1393,1394],{},"为了降低维度和减少噪声干扰，进行无用特征剔除：移除了唯一标识符、高基数或对分类任务贡献有限的列，包括 policy_number（保单编号）、policy_bind_date（保单绑定日期）、policy_state（投保州）、insured_zip（邮编）、incident_location（事故地点）、incident_date（事故日期）、incident_state（事故州）、incident_city（事故城市）、insured_hobbies（爱好）、auto_make（车辆品牌）、auto_model（车型）、auto_year（生产年份）以及_c39（空列）。",[1091,1396,1397],{},"机器学习模型要求输入数据为数值型。对于剩余的类别型变量（如  insured_sex、education_level、incident_type  等），本研究采用标签编码（Label Encoding）技术，将其映射为整数序列。同时，目标变量  fraud_reported  被转换为二元数值标签，其中“N”映射为 0，“Y”映射为 1。",[1091,1399,1400],{},"预处理后，数据集保留了一些判别特征，包括数值型特征（如 months_as_customer、age、policy_deductable、policy_annual_premium、umbrella_limit、capital-gains、capital-loss、incident_hour_of_the_day、number_of_vehicles_involved、bodily_injuries、witnesses、total_claim_amount、injury_claim、property_claim、vehicle_claim）和编码后的类别特征，构成模型训练的输入向量空间。",[1382,1402,1404],{"id":1403},"_33-handling-class-imbalance","3.3 Handling Class Imbalance",[1091,1406,1407],{},"保险欺诈检测本质上属于高度类别不平衡的分类问题，欺诈样本仅占约 24.7%，多数类（非欺诈）主导数据集。若直接训练模型，将倾向于预测多数类，导致对欺诈案例的召回率低下，而在实际应用中，漏检欺诈（假阴性）的成本远高于误报（假阳性）（Dal Pozzolo et al., 2015）。",[1091,1409,1410],{},"本研究引入 Synthetic Minority Over-sampling Technique（SMOTE）算法（Chawla et al., 2002）。SMOTE 通过在少数类样本间进行线性插值合成新样本，避免简单复制带来的过拟合风险。通过应用 SMOTE，使训练集正负样本比例调整至 1:1（欺诈样本比例从约 24.7%提升至 50%），来增强模型对少数类的学习能力。",[1221,1412,1414],{"id":1413},"_34-实验设置与评估指标","3.4 实验设置与评估指标",[1091,1416,1417],{},"为验证所提方法的有效性并确保实验的可复现性，本研究设计了如下实验方案：",[1361,1419,1421],{"id":1420},"_341-数据集划分","3.4.1 数据集划分",[1091,1423,1424],{},"在应用 SMOTE 算法后，将平衡后的数据集按 80% : 20% 的比例随机划分为训练集和测试集。训练集用于模型的构建与参数学习，测试集仅用于最终性能评估，以验证模型的泛化能力。划分过程设置随机种子（random_state=42）以确保结果的一致性。",[1361,1426,1428],{"id":1427},"_342-模型选择","3.4.2 模型选择",[1091,1430,1431,1432,1434,1435,1437],{},"本研究选取  ",[1109,1433,1261],{},"  作为核心分类模型。随机森林作为一种集成学习算法，通过构建多棵决策树并利用 Bagging 策略进行投票，能够有效处理高维特征并具有较强的抗过拟合能力。",[1343,1436],{},"\n为评估随机森林的性能优势，本研究选取了两个经典算法作为基线模型（Baseline）：",[1439,1440,1441,1447],"ul",{},[1186,1442,1443,1446],{},[1109,1444,1445],{},"逻辑回归（Logistic Regression, LR）","：作为线性模型的代表，用于衡量非线性特征交互的重要性。",[1186,1448,1449,1452],{},[1109,1450,1451],{},"支持向量机（Support Vector Machine, SVM）","：作为经典的核方法分类器，用于对比不同决策边界的划分效果。",[1091,1454,1455],{},"所有模型均在经过标准化（StandardScaler）处理的特征上进行训练，其中随机森林的基学习器数量设置为 100（n_estimators=100）。",[1221,1457,1459],{"id":1458},"_35-实验设计与评估指标-experimental-design-and-metrics","3.5 实验设计与评估指标 (Experimental Design and Metrics)",[1361,1461,1463],{"id":1462},"_351-实验设置","3.5.1 实验设置",[1091,1465,1466],{},"所有实验均在 Python 环境下基于 Scikit-Learn 库实现。",[1183,1468,1469,1479],{},[1186,1470,1471,1474,1475,1478],{},[1109,1472,1473],{},"数据划分","：采用  ",[1109,1476,1477],{},"80\u002F20 划分原则","。在应用 SMOTE 之前，将原始数据按 80% 划分为训练集，20% 划分为测试集。",[1186,1480,1481,1484],{},[1109,1482,1483],{},"标准化","：对所有特征进行  StandardScaler  标准化处理，消除量纲差异对 SVM 和 LR 等距离敏感模型的影响。",[1361,1486,1488],{"id":1487},"_352-评估指标","3.5.2 评估指标",[1091,1490,1491],{},"鉴于准确率（Accuracy）在类别不平衡数据集中的局限性（例如，若模型将所有样本预测为多数类非欺诈，可轻松获得约 75%的准确率，从而掩盖对少数类的识别能力），本研究采用多维度评估指标，以全面衡量模型在欺诈检测任务中的性能。这些指标特别关注对少数类（欺诈）的识别能力，同时考虑整体泛化性能。",[1439,1493,1494],{},[1186,1495,1496,1499],{},[1109,1497,1498],{},"准确率（Accuracy）","：模型整体分类正确的比例。计算公式为",[1091,1501,1502],{},[1503,1504,1507,1579],"span",{"className":1505},[1506],"katex",[1503,1508,1511],{"className":1509},[1510],"katex-mathml",[1512,1513,1515],"math",{"xmlns":1514},"http:\u002F\u002Fwww.w3.org\u002F1998\u002FMath\u002FMathML",[1516,1517,1518,1574],"semantics",{},[1519,1520,1521,1525,1529],"mrow",{},[1522,1523,1524],"mtext",{},"Accuracy",[1526,1527,1528],"mo",{},"=",[1530,1531,1532,1549],"mfrac",{},[1519,1533,1534,1538,1541,1544,1546],{},[1535,1536,1537],"mi",{},"T",[1535,1539,1540],{},"P",[1526,1542,1543],{},"+",[1535,1545,1537],{},[1535,1547,1548],{},"N",[1519,1550,1551,1553,1555,1557,1559,1561,1563,1566,1568,1570,1572],{},[1535,1552,1537],{},[1535,1554,1540],{},[1526,1556,1543],{},[1535,1558,1537],{},[1535,1560,1548],{},[1526,1562,1543],{},[1535,1564,1565],{},"F",[1535,1567,1540],{},[1526,1569,1543],{},[1535,1571,1565],{},[1535,1573,1548],{},[1575,1576,1578],"annotation",{"encoding":1577},"application\u002Fx-tex"," \\text{Accuracy} = \\frac{TP + TN}{TP + TN + FP + FN}",[1503,1580,1584,1613],{"className":1581,"ariaHidden":1583},[1582],"katex-html","true",[1503,1585,1588,1593,1601,1606,1610],{"className":1586},[1587],"base",[1503,1589],{"className":1590,"style":1592},[1591],"strut","height:0.8778em;vertical-align:-0.1944em;",[1503,1594,1598],{"className":1595},[1596,1597],"mord","text",[1503,1599,1524],{"className":1600},[1596],[1503,1602],{"className":1603,"style":1605},[1604],"mspace","margin-right:0.2778em;",[1503,1607,1528],{"className":1608},[1609],"mrel",[1503,1611],{"className":1612,"style":1605},[1604],[1503,1614,1616,1620],{"className":1615},[1587],[1503,1617],{"className":1618,"style":1619},[1591],"height:1.2757em;vertical-align:-0.4033em;",[1503,1621,1623,1628,1752],{"className":1622},[1596],[1503,1624],{"className":1625},[1626,1627],"mopen","nulldelimiter",[1503,1629,1631],{"className":1630},[1530],[1503,1632,1636,1743],{"className":1633},[1634,1635],"vlist-t","vlist-t2",[1503,1637,1640,1738],{"className":1638},[1639],"vlist-r",[1503,1641,1645,1700,1711],{"className":1642,"style":1644},[1643],"vlist","height:0.8723em;",[1503,1646,1648,1653],{"style":1647},"top:-2.655em;",[1503,1649],{"className":1650,"style":1652},[1651],"pstrut","height:3em;",[1503,1654,1660],{"className":1655},[1656,1657,1658,1659],"sizing","reset-size6","size3","mtight",[1503,1661,1663,1668,1671,1675,1678,1682,1685,1688,1691,1694,1697],{"className":1662},[1596,1659],[1503,1664,1537],{"className":1665,"style":1667},[1596,1666,1659],"mathnormal","margin-right:0.1389em;",[1503,1669,1540],{"className":1670,"style":1667},[1596,1666,1659],[1503,1672,1543],{"className":1673},[1674,1659],"mbin",[1503,1676,1537],{"className":1677,"style":1667},[1596,1666,1659],[1503,1679,1548],{"className":1680,"style":1681},[1596,1666,1659],"margin-right:0.109em;",[1503,1683,1543],{"className":1684},[1674,1659],[1503,1686,1565],{"className":1687,"style":1667},[1596,1666,1659],[1503,1689,1540],{"className":1690,"style":1667},[1596,1666,1659],[1503,1692,1543],{"className":1693},[1674,1659],[1503,1695,1565],{"className":1696,"style":1667},[1596,1666,1659],[1503,1698,1548],{"className":1699,"style":1681},[1596,1666,1659],[1503,1701,1703,1706],{"style":1702},"top:-3.23em;",[1503,1704],{"className":1705,"style":1652},[1651],[1503,1707],{"className":1708,"style":1710},[1709],"frac-line","border-bottom-width:0.04em;",[1503,1712,1714,1717],{"style":1713},"top:-3.394em;",[1503,1715],{"className":1716,"style":1652},[1651],[1503,1718,1720],{"className":1719},[1656,1657,1658,1659],[1503,1721,1723,1726,1729,1732,1735],{"className":1722},[1596,1659],[1503,1724,1537],{"className":1725,"style":1667},[1596,1666,1659],[1503,1727,1540],{"className":1728,"style":1667},[1596,1666,1659],[1503,1730,1543],{"className":1731},[1674,1659],[1503,1733,1537],{"className":1734,"style":1667},[1596,1666,1659],[1503,1736,1548],{"className":1737,"style":1681},[1596,1666,1659],[1503,1739,1742],{"className":1740},[1741],"vlist-s","​",[1503,1744,1746],{"className":1745},[1639],[1503,1747,1750],{"className":1748,"style":1749},[1643],"height:0.4033em;",[1503,1751],{},[1503,1753],{"className":1754},[1755,1627],"mclose",[1091,1757,1758],{},"其中，TP（True Positive）为正确预测的欺诈样本，TN（True Negative）为正确预测的非欺诈样本，FP（False Positive）为误报的欺诈样本，FN（False Negative）为漏报的欺诈样本。尽管该指标直观，但在本研究中仅作为辅助参考。",[1439,1760,1761],{},[1186,1762,1763,1766],{},[1109,1764,1765],{},"精确率（Precision）","：预测为欺诈的样本中实际为欺诈的比例。计算公式为",[1091,1768,1769],{},[1503,1770,1772,1809],{"className":1771},[1506],[1503,1773,1775],{"className":1774},[1510],[1512,1776,1777],{"xmlns":1514},[1516,1778,1779,1806],{},[1519,1780,1781,1784,1786],{},[1522,1782,1783],{},"Precision",[1526,1785,1528],{},[1530,1787,1788,1794],{},[1519,1789,1790,1792],{},[1535,1791,1537],{},[1535,1793,1540],{},[1519,1795,1796,1798,1800,1802,1804],{},[1535,1797,1537],{},[1535,1799,1540],{},[1526,1801,1543],{},[1535,1803,1565],{},[1535,1805,1540],{},[1575,1807,1808],{"encoding":1577}," \\text{Precision} = \\frac{TP}{TP + FP}",[1503,1810,1812,1834],{"className":1811,"ariaHidden":1583},[1582],[1503,1813,1815,1819,1825,1828,1831],{"className":1814},[1587],[1503,1816],{"className":1817,"style":1818},[1591],"height:0.6833em;",[1503,1820,1822],{"className":1821},[1596,1597],[1503,1823,1783],{"className":1824},[1596],[1503,1826],{"className":1827,"style":1605},[1604],[1503,1829,1528],{"className":1830},[1609],[1503,1832],{"className":1833,"style":1605},[1604],[1503,1835,1837,1840],{"className":1836},[1587],[1503,1838],{"className":1839,"style":1619},[1591],[1503,1841,1843,1846,1920],{"className":1842},[1596],[1503,1844],{"className":1845},[1626,1627],[1503,1847,1849],{"className":1848},[1530],[1503,1850,1852,1912],{"className":1851},[1634,1635],[1503,1853,1855,1909],{"className":1854},[1639],[1503,1856,1858,1884,1892],{"className":1857,"style":1644},[1643],[1503,1859,1860,1863],{"style":1647},[1503,1861],{"className":1862,"style":1652},[1651],[1503,1864,1866],{"className":1865},[1656,1657,1658,1659],[1503,1867,1869,1872,1875,1878,1881],{"className":1868},[1596,1659],[1503,1870,1537],{"className":1871,"style":1667},[1596,1666,1659],[1503,1873,1540],{"className":1874,"style":1667},[1596,1666,1659],[1503,1876,1543],{"className":1877},[1674,1659],[1503,1879,1565],{"className":1880,"style":1667},[1596,1666,1659],[1503,1882,1540],{"className":1883,"style":1667},[1596,1666,1659],[1503,1885,1886,1889],{"style":1702},[1503,1887],{"className":1888,"style":1652},[1651],[1503,1890],{"className":1891,"style":1710},[1709],[1503,1893,1894,1897],{"style":1713},[1503,1895],{"className":1896,"style":1652},[1651],[1503,1898,1900],{"className":1899},[1656,1657,1658,1659],[1503,1901,1903,1906],{"className":1902},[1596,1659],[1503,1904,1537],{"className":1905,"style":1667},[1596,1666,1659],[1503,1907,1540],{"className":1908,"style":1667},[1596,1666,1659],[1503,1910,1742],{"className":1911},[1741],[1503,1913,1915],{"className":1914},[1639],[1503,1916,1918],{"className":1917,"style":1749},[1643],[1503,1919],{},[1503,1921],{"className":1922},[1755,1627],[1091,1924,1925],{},"该指标衡量模型的误报控制能力，在保险场景中有助于评估额外审核成本。",[1439,1927,1928],{},[1186,1929,1930,1933],{},[1109,1931,1932],{},"召回率（Recall \u002F Sensitivity）","：实际欺诈样本中被正确识别的比例。计算公式为",[1091,1935,1936],{},[1503,1937,1939,1976],{"className":1938},[1506],[1503,1940,1942],{"className":1941},[1510],[1512,1943,1944],{"xmlns":1514},[1516,1945,1946,1973],{},[1519,1947,1948,1951,1953],{},[1522,1949,1950],{},"Recall",[1526,1952,1528],{},[1530,1954,1955,1961],{},[1519,1956,1957,1959],{},[1535,1958,1537],{},[1535,1960,1540],{},[1519,1962,1963,1965,1967,1969,1971],{},[1535,1964,1537],{},[1535,1966,1540],{},[1526,1968,1543],{},[1535,1970,1565],{},[1535,1972,1548],{},[1575,1974,1975],{"encoding":1577}," \\text{Recall} = \\frac{TP}{TP + FN}",[1503,1977,1979,2001],{"className":1978,"ariaHidden":1583},[1582],[1503,1980,1982,1986,1992,1995,1998],{"className":1981},[1587],[1503,1983],{"className":1984,"style":1985},[1591],"height:0.6944em;",[1503,1987,1989],{"className":1988},[1596,1597],[1503,1990,1950],{"className":1991},[1596],[1503,1993],{"className":1994,"style":1605},[1604],[1503,1996,1528],{"className":1997},[1609],[1503,1999],{"className":2000,"style":1605},[1604],[1503,2002,2004,2007],{"className":2003},[1587],[1503,2005],{"className":2006,"style":1619},[1591],[1503,2008,2010,2013,2087],{"className":2009},[1596],[1503,2011],{"className":2012},[1626,1627],[1503,2014,2016],{"className":2015},[1530],[1503,2017,2019,2079],{"className":2018},[1634,1635],[1503,2020,2022,2076],{"className":2021},[1639],[1503,2023,2025,2051,2059],{"className":2024,"style":1644},[1643],[1503,2026,2027,2030],{"style":1647},[1503,2028],{"className":2029,"style":1652},[1651],[1503,2031,2033],{"className":2032},[1656,1657,1658,1659],[1503,2034,2036,2039,2042,2045,2048],{"className":2035},[1596,1659],[1503,2037,1537],{"className":2038,"style":1667},[1596,1666,1659],[1503,2040,1540],{"className":2041,"style":1667},[1596,1666,1659],[1503,2043,1543],{"className":2044},[1674,1659],[1503,2046,1565],{"className":2047,"style":1667},[1596,1666,1659],[1503,2049,1548],{"className":2050,"style":1681},[1596,1666,1659],[1503,2052,2053,2056],{"style":1702},[1503,2054],{"className":2055,"style":1652},[1651],[1503,2057],{"className":2058,"style":1710},[1709],[1503,2060,2061,2064],{"style":1713},[1503,2062],{"className":2063,"style":1652},[1651],[1503,2065,2067],{"className":2066},[1656,1657,1658,1659],[1503,2068,2070,2073],{"className":2069},[1596,1659],[1503,2071,1537],{"className":2072,"style":1667},[1596,1666,1659],[1503,2074,1540],{"className":2075,"style":1667},[1596,1666,1659],[1503,2077,1742],{"className":2078},[1741],[1503,2080,2082],{"className":2081},[1639],[1503,2083,2085],{"className":2084,"style":1749},[1643],[1503,2086],{},[1503,2088],{"className":2089},[1755,1627],[1091,2091,2092],{},"在反欺诈应用中，召回率是首要优化目标，因为漏检欺诈（FN）的经济代价远高于误报（FP）。",[1439,2094,2095],{},[1186,2096,2097,2100],{},[1109,2098,2099],{},"F1 分数（F1-score）","：精确率与召回率的调和平均，适用于不平衡数据下的综合评估。计算公式为",[1091,2102,2103],{},[1503,2104,2106,2147],{"className":2105},[1506],[1503,2107,2109],{"className":2108},[1510],[1512,2110,2111],{"xmlns":1514},[1516,2112,2113,2144],{},[1519,2114,2115,2118,2120,2123,2126],{},[1522,2116,2117],{},"F1-score",[1526,2119,1528],{},[2121,2122,1216],"mn",{},[1526,2124,2125],{},"×",[1530,2127,2128,2136],{},[1519,2129,2130,2132,2134],{},[1522,2131,1783],{},[1526,2133,2125],{},[1522,2135,1950],{},[1519,2137,2138,2140,2142],{},[1522,2139,1783],{},[1526,2141,1543],{},[1522,2143,1950],{},[1575,2145,2146],{"encoding":1577}," \\text{F1-score} = 2 \\times \\frac{\\text{Precision} \\times \\text{Recall}}{\\text{Precision} + \\text{Recall}}",[1503,2148,2150,2171,2191],{"className":2149,"ariaHidden":1583},[1582],[1503,2151,2153,2156,2162,2165,2168],{"className":2152},[1587],[1503,2154],{"className":2155,"style":1818},[1591],[1503,2157,2159],{"className":2158},[1596,1597],[1503,2160,2117],{"className":2161},[1596],[1503,2163],{"className":2164,"style":1605},[1604],[1503,2166,1528],{"className":2167},[1609],[1503,2169],{"className":2170,"style":1605},[1604],[1503,2172,2174,2178,2181,2185,2188],{"className":2173},[1587],[1503,2175],{"className":2176,"style":2177},[1591],"height:0.7278em;vertical-align:-0.0833em;",[1503,2179,1216],{"className":2180},[1596],[1503,2182],{"className":2183,"style":2184},[1604],"margin-right:0.2222em;",[1503,2186,2125],{"className":2187},[1674],[1503,2189],{"className":2190,"style":2184},[1604],[1503,2192,2194,2198],{"className":2193},[1587],[1503,2195],{"className":2196,"style":2197},[1591],"height:1.2834em;vertical-align:-0.4033em;",[1503,2199,2201,2204,2288],{"className":2200},[1596],[1503,2202],{"className":2203},[1626,1627],[1503,2205,2207],{"className":2206},[1530],[1503,2208,2210,2280],{"className":2209},[1634,1635],[1503,2211,2213,2277],{"className":2212},[1639],[1503,2214,2217,2243,2251],{"className":2215,"style":2216},[1643],"height:0.8801em;",[1503,2218,2219,2222],{"style":1647},[1503,2220],{"className":2221,"style":1652},[1651],[1503,2223,2225],{"className":2224},[1656,1657,1658,1659],[1503,2226,2228,2234,2237],{"className":2227},[1596,1659],[1503,2229,2231],{"className":2230},[1596,1597,1659],[1503,2232,1783],{"className":2233},[1596,1659],[1503,2235,1543],{"className":2236},[1674,1659],[1503,2238,2240],{"className":2239},[1596,1597,1659],[1503,2241,1950],{"className":2242},[1596,1659],[1503,2244,2245,2248],{"style":1702},[1503,2246],{"className":2247,"style":1652},[1651],[1503,2249],{"className":2250,"style":1710},[1709],[1503,2252,2253,2256],{"style":1713},[1503,2254],{"className":2255,"style":1652},[1651],[1503,2257,2259],{"className":2258},[1656,1657,1658,1659],[1503,2260,2262,2268,2271],{"className":2261},[1596,1659],[1503,2263,2265],{"className":2264},[1596,1597,1659],[1503,2266,1783],{"className":2267},[1596,1659],[1503,2269,2125],{"className":2270},[1674,1659],[1503,2272,2274],{"className":2273},[1596,1597,1659],[1503,2275,1950],{"className":2276},[1596,1659],[1503,2278,1742],{"className":2279},[1741],[1503,2281,2283],{"className":2282},[1639],[1503,2284,2286],{"className":2285,"style":1749},[1643],[1503,2287],{},[1503,2289],{"className":2290},[1755,1627],[1091,2292,2293],{},"该指标在 Precision 与 Recall 间取得平衡，提供单一数值总结模型效能。",[1439,2295,2296,2302],{},[1186,2297,2298,2301],{},[1109,2299,2300],{},"AUC-ROC（Area Under the Receiver Operating Characteristic Curve）","：ROC 曲线下的面积，衡量模型在不同分类阈值下的区分能力。AUC 值介于 0.5（随机猜测）与 1.0（完美分类）之间，对类别不平衡具有较强鲁棒性，是本研究评估模型综合性能的核心指标。",[1186,2303,2304,2307],{},[1109,2305,2306],{},"混淆矩阵（Confusion Matrix）","：以矩阵形式直观展示 TP、TN、FP、FN 的分布，便于分析模型在具体类别上的错误类型。",[1098,2309],{},[1098,2311],{},[1101,2313,2315],{"id":2314},"_4-实验结果与分析","4 实验结果与分析",[1091,2317,2318],{},"本章基于公开的汽车保险理赔数据集（1000条记录，欺诈样本占比约24.7%），呈现扩展后的对比实验结果。实验采用统一的预处理管道：缺失值填充、类别特征编码（One-Hot Encoding）、数值特征标准化，以及在训练集上应用SMOTE过采样技术，使类别分布平衡至50%。所有模型通过网格搜索进行超参数优化，并使用分层抽样划分训练\u002F测试集（比例约7:3，测试集302条记录）。评估指标包括准确率（Accuracy）、AUC-ROC、精确率（Precision）、召回率（Recall）和F1分数，重点关注欺诈类（少数类）的性能。",[1361,2320,2322],{"id":2321},"_41-实验环境与设置","4.1 实验环境与设置",[1091,2324,2325],{},"实验在一台配备 12th Gen Intel(R) Core(TM) i7-12700H (20) @ 4.70 GHz 的电脑上进行，内存为 32 GB，操作系统是 64 位 Windows 11。",[1091,2327,2328],{},"主要使用 Python3.10 实现。Pandas 的数据帧负责加载数据集。Scikit Learn 库实现了机器学习和集成模型。作品的源代码、可视化和数据发布在作者的 GitHub 网站上。",[1091,2330,2331],{},"【链接！】",[1382,2333,2335],{"id":2334},"_41-模型性能对比","4.1 模型性能对比",[1091,2337,2338,2339,2342,2343,2342,2346,2342,2349,2342,2352,2342,2355,2358,2359,1150],{},"实验对比了七种经典机器学习算法：",[1109,2340,2341],{},"逻辑回归（Logistic Regression）","、",[1109,2344,2345],{},"支持向量机（SVM）",[1109,2347,2348],{},"决策树（Decision Tree）",[1109,2350,2351],{},"K-最近邻（KNN）",[1109,2353,2354],{},"AdaBoost",[1109,2356,2357],{},"随机森林（Random Forest）"," 和 ",[1109,2360,2361],{},"XGBoost",[1091,2363,2364],{},"关键性能指标汇总如下（表4-1）：",[2366,2367,2368,2393],"table",{},[2369,2370,2371],"thead",{},[2372,2373,2374,2378,2381,2384,2387,2390],"tr",{},[2375,2376,2377],"th",{},"模型",[2375,2379,2380],{},"准确率",[2375,2382,2383],{},"AUC",[2375,2385,2386],{},"欺诈类精确率",[2375,2388,2389],{},"欺诈类召回率",[2375,2391,2392],{},"欺诈类F1分数",[2394,2395,2396,2416,2436,2456,2476,2493,2511],"tbody",{},[2372,2397,2398,2402,2405,2408,2411,2414],{},[2399,2400,2401],"td",{},"Logistic Regression",[2399,2403,2404],{},"0.7119",[2399,2406,2407],{},"0.7749",[2399,2409,2410],{},"0.73",[2399,2412,2413],{},"0.72",[2399,2415,2410],{},[2372,2417,2418,2421,2424,2427,2430,2433],{},[2399,2419,2420],{},"SVM",[2399,2422,2423],{},"0.8642",[2399,2425,2426],{},"0.9255",[2399,2428,2429],{},"0.86",[2399,2431,2432],{},"0.89",[2399,2434,2435],{},"0.87",[2372,2437,2438,2441,2444,2447,2450,2453],{},[2399,2439,2440],{},"Decision Tree",[2399,2442,2443],{},"0.7417",[2399,2445,2446],{},"0.7416",[2399,2448,2449],{},"0.76",[2399,2451,2452],{},"0.74",[2399,2454,2455],{},"0.75",[2372,2457,2458,2461,2464,2467,2470,2473],{},[2399,2459,2460],{},"KNN",[2399,2462,2463],{},"0.7185",[2399,2465,2466],{},"0.8234",[2399,2468,2469],{},"0.67",[2399,2471,2472],{},"0.94",[2399,2474,2475],{},"0.78",[2372,2477,2478,2480,2483,2486,2489,2491],{},[2399,2479,2354],{},[2399,2481,2482],{},"0.8311",[2399,2484,2485],{},"0.8975",[2399,2487,2488],{},"0.84",[2399,2490,2488],{},[2399,2492,2488],{},[2372,2494,2495,2498,2501,2504,2507,2509],{},[2399,2496,2497],{},"Random Forest",[2399,2499,2500],{},"0.8675",[2399,2502,2503],{},"0.9382",[2399,2505,2506],{},"0.88",[2399,2508,2429],{},[2399,2510,2435],{},[2372,2512,2513,2515,2518,2521,2523,2526],{},[2399,2514,2361],{},[2399,2516,2517],{},"0.8444",[2399,2519,2520],{},"0.9232",[2399,2522,2435],{},[2399,2524,2525],{},"0.82",[2399,2527,2528],{},"0.85",[1091,2530,2531,2532,2534,2535,2537,2538,2540,2541,2543],{},"从以上结果可见，",[1109,2533,232],{},"在AUC（0.9382）和整体准确率（0.8675）上表现最佳，欺诈类F1分数达0.87，实现了精确率与召回率的良好平衡。",[1109,2536,2361],{},"和",[1109,2539,2420],{},"紧随其后，AUC分别达0.9232和0.9255，显示出梯度提升树和核方法在非线性模式捕获上的优势。",[1109,2542,2460],{},"虽在欺诈类召回率上最高（0.94），但精确率较低，导致整体F1分数不如集成模型。线性模型如逻辑回归和单一决策树性能相对较弱，AUC低于0.80，表明其难以有效处理高维非线性关系。",[1091,2545,2546],{},"ROC曲线比较进一步证实，随机森林曲线最接近左上角，主导其他模型，尤其在低假阳性率区间表现出色。",[1382,2548,2550],{"id":2549},"_42-roc曲线分析","4.2 ROC曲线分析",[1091,2552,2553],{},"ROC曲线用于评估模型在不同分类阈值下的真阳性率（True Positive Rate）与假阳性率（False Positive Rate）的权衡关系，曲线下面积（AUC）值越高，模型的整体区分能力越强。图4-1展示了七种模型的ROC曲线对比结果。",[1091,2555,2556],{},[1109,2557,2558],{},"图4-1 不同模型的ROC曲线对比",[1091,2560,2561],{},"（如图像ID:0所示，随机森林的ROC曲线最靠近左上角，AUC达到0.9382；XGBoost（AUC=0.9232）和SVM（AUC=0.9255）紧随其后；AdaBoost（AUC=0.8975）表现良好；KNN（AUC=0.8234）和逻辑回归（AUC=0.7749）相对较低；决策树（AUC=0.7416）表现最弱。随机分类器的对角线作为基准参考。）",[1091,2563,2564],{},"从图中可见，随机森林的ROC曲线在几乎整个假阳性率区间内均主导其他模型，尤其在低假阳性率（\u003C0.2）区域上升更快，表明其能够在保持较低误报率的同时实现较高的欺诈捕获率。XGBoost和SVM的曲线高度重叠，性能接近随机森林，体现了梯度提升树和核方法在非线性模式识别上的优势。相比之下，单一决策树和逻辑回归的曲线较低，反映了其对复杂交互特征和类别不平衡的处理能力不足。KNN虽在高假阳性率区间召回较高，但早期上升缓慢，导致整体AUC较低。",[1091,2566,2567],{},"该结果进一步证实，集成学习算法（随机森林、XGBoost、AdaBoost）在本数据集上具有显著优越性，与SMOTE过采样相结合后，其对少数类（欺诈）的敏感性得到有效提升。随机森林的最高AUC值验证了其作为本研究核心模型的合理性，为实际部署提供了更可靠的阈值选择依据。",[1382,2569,2571],{"id":2570},"_43-随机森林模型详细分析","4.3 随机森林模型详细分析",[1091,2573,2574],{},"作为核心模型，随机森林的混淆矩阵如下（图4-2）：",[1091,2576,2577],{},[1109,2578,2579],{},"图4-2 随机森林混淆矩阵",[1439,2581,2582,2585,2588,2591],{},[1186,2583,2584],{},"真阴性（TN）：124（正确识别非欺诈）",[1186,2586,2587],{},"假阳性（FP）：18（误报欺诈）",[1186,2589,2590],{},"假阴性（FN）：22（漏检欺诈）",[1186,2592,2593],{},"真阳性（TP）：138（正确识别欺诈）",[1091,2595,2596],{},"模型对欺诈类的召回率达0.86，误报率较低，适合实际业务中平衡审核成本与风险控制的需求。",[1091,2598,2599],{},"特征重要性分析（基于Gini不纯度下降）显示前10个关键特征（图4-3）：",[1091,2601,2602],{},[1109,2603,2604],{},"图4-3 随机森林Top 10特征重要性",[1183,2606,2607,2610,2613,2616,2619,2622,2625,2628,2631,2634],{},[1186,2608,2609],{},"incident_severity（事故严重程度）：0.2294",[1186,2611,2612],{},"total_claim_amount（总赔付金额）：0.0528",[1186,2614,2615],{},"vehicle_claim（车辆赔付金额）：0.0499",[1186,2617,2618],{},"witnesses（目击证人数）：0.0492",[1186,2620,2621],{},"authorities_contacted（当局联系情况）：0.0453",[1186,2623,2624],{},"insured_occupation（投保人职业）：0.0431",[1186,2626,2627],{},"policy_annual_premium（年保费）：0.0424",[1186,2629,2630],{},"injury_claim（伤残赔付金额）：0.0406",[1186,2632,2633],{},"property_claim（财产赔付金额）：0.0401",[1186,2635,2636],{},"incident_hour_of_the_day（事故发生小时）：0.0373",[1091,2638,2639],{},"这些特征与保险实务高度一致：轻微事故伴随高赔付金额、缺乏外部证据等往往为欺诈信号。该分析提升了模型的可解释性。",[1382,2641,2643],{"id":2642},"_44-不平衡数据处理策略的有效性分析","4.4 不平衡数据处理策略的有效性分析",[1091,2645,2646],{},"为系统评估不平衡处理策略的影响，本研究对比了三种场景：无重采样、SMOTE过采样和随机欠采样。表4-2展示了随机森林在不同策略下的性能（测试集保持原始分布）。",[2366,2648,2649,2664],{},[2369,2650,2651],{},[2372,2652,2653,2656,2658,2660,2662],{},[2375,2654,2655],{},"策略",[2375,2657,2380],{},[2375,2659,2383],{},[2375,2661,2389],{},[2375,2663,2392],{},[2394,2665,2666,2683,2696],{},[2372,2667,2668,2671,2674,2677,2680],{},[2399,2669,2670],{},"无重采样",[2399,2672,2673],{},"0.7800",[2399,2675,2676],{},"0.7348",[2399,2678,2679],{},"0.47",[2399,2681,2682],{},"0.54",[2372,2684,2685,2688,2690,2692,2694],{},[2399,2686,2687],{},"SMOTE过采样",[2399,2689,2500],{},[2399,2691,2503],{},[2399,2693,2429],{},[2399,2695,2435],{},[2372,2697,2698,2701,2704,2707,2710],{},[2399,2699,2700],{},"随机欠采样",[2399,2702,2703],{},"0.7172",[2399,2705,2706],{},"0.8225",[2399,2708,2709],{},"0.69",[2399,2711,2712],{},"0.70",[1091,2714,2715],{},[1109,2716,2717],{},"表4-2 不同不平衡处理策略对随机森林性能的影响",[1091,2719,2720],{},"SMOTE过采样显著提升了AUC和欺诈类召回率（从0.47提高至0.86），F1分数提升明显。随机欠采样虽提高了召回率，但整体准确率和AUC下降，表明信息损失较大。无重采样策略下模型对少数类敏感性不足。该规律在其他模型中同样存在：SMOTE对KNN和SVM的召回率提升尤为显著（超过20%），而对随机森林的增益相对温和，得益于其内置自助采样机制。综合而言，SMOTE在本数据集上是最稳健的不平衡处理策略。",[1382,2722,2724],{"id":2723},"_45-特征选择的影响分析","4.5 特征选择的影响分析",[1091,2726,2727],{},"本研究采用递归特征消除（RFE）结合随机森林进行特征选择，对比了使用全部特征与Top 20特征的模型性能。",[2366,2729,2730,2748],{},[2369,2731,2732],{},[2372,2733,2734,2737,2739,2741,2743,2745],{},[2375,2735,2736],{},"特征集",[2375,2738,2380],{},[2375,2740,2383],{},[2375,2742,2389],{},[2375,2744,2392],{},[2375,2746,2747],{},"训练时间减少",[2394,2749,2750,2768],{},[2372,2751,2752,2755,2758,2761,2763,2766],{},[2399,2753,2754],{},"全部特征",[2399,2756,2757],{},"0.9091",[2399,2759,2760],{},"0.9777",[2399,2762,2506],{},[2399,2764,2765],{},"0.90",[2399,2767],{},[2372,2769,2770,2773,2776,2779,2782,2784],{},[2399,2771,2772],{},"RFE选择20个特征",[2399,2774,2775],{},"0.7650",[2399,2777,2778],{},"0.7501",[2399,2780,2781],{},"0.58",[2399,2783,2781],{},[2399,2785,2786],{},"约30%",[1091,2788,2789],{},[1109,2790,2791],{},"表4-3 特征选择对随机森林性能的影响",[1091,2793,2794],{},"结果显示，使用全部特征的模型性能显著优于仅选20个特征的模型（AUC下降约0.2276）。这表明数据集特征整体信息丰富，强行降维会导致关键信号丢失。尽管特征选择可减少约30%训练时间，但在本任务中不推荐大幅降维，以优先保证检测精度。",[1382,2796,2798],{"id":2797},"_46-结果讨论","4.6 结果讨论",[1091,2800,2801],{},"扩展对比实验表明，集成学习算法（随机森林、XGBoost、AdaBoost）在本欺诈检测任务中显著优于传统单一分类器，其中随机森林综合性能最优。SMOTE过采样的引入有效缓解了类别不平衡问题，尤其提升了少数类的识别能力，而随机欠采样虽可进一步提高召回率，但以牺牲整体准确率为代价。特征选择实验显示，本数据集特征冗余度较低，保留全部特征更有利于模型性能。",[1091,2803,2804],{},"整体框架在中小规模结构化数据上表现出高效性和鲁棒性，随机森林结合SMOTE与特征重要性分析，不仅实现了高精度检测，还提供了业务可解释的洞察，为汽车保险欺诈风险管理提供了实用参考。尽管XGBoost性能接近，但随机森林在可解释性和计算稳定性上的优势使其更适合实际部署。",[1361,2806,2808],{"id":2807},"第5章-结论与展望","第5章 结论与展望",[1382,2810,2812],{"id":2811},"_51-研究结论","5.1 研究结论",[1091,2814,2815],{},"本研究针对汽车保险理赔欺诈检测问题，构建了一个系统化的机器学习分析框架。该框架涵盖数据预处理、类别不平衡处理、模型训练以及多维度性能评估等环节。实验基于公开的汽车保险理赔数据集，采用SMOTE过采样技术有效缓解了类别不平衡问题，并对逻辑回归、支持向量机以及随机森林三种传统分类算法进行了全面对比。",[1091,2817,2818],{},"实证结果表明：",[1183,2820,2821,2824,2827],{},[1186,2822,2823],{},"随机森林模型在综合性能上表现出色，其AUC值达到0.9382，整体准确率达0.8675，欺诈类召回率为0.86，F1分数为0.87。该模型在精确率与召回率的平衡方面优于支持向量机（AUC=0.9255），并显著优于逻辑回归（AUC=0.7749）。",[1186,2825,2826],{},"SMOTE过采样策略显著提升了模型对少数类样本的识别能力，证实了数据层面不平衡处理在欺诈检测任务中的关键作用。",[1186,2828,2829],{},"特征重要性分析揭示了incident_severity（事故严重程度）、total_claim_amount（总赔付金额）以及vehicle_claim（车辆赔付金额）等特征对欺诈预测的主导贡献，这些发现与保险业务实践高度契合，增强了模型的可解释性。",[1091,2831,2832],{},"上述结论验证了传统集成学习算法，特别是随机森林，在处理高维、不平衡保险数据时的鲁棒性和有效性。该框架不仅实现了较高的预测精度，还保持了较强的解释能力，为数据驱动的欺诈检测提供了可靠的实证依据。",[1382,2834,2836],{"id":2835},"_52-理论与实践意义","5.2 理论与实践意义",[1091,2838,2839],{},"从理论层面，本研究丰富了保险欺诈检测领域的实证文献，强调了预处理阶段（尤其是类别不平衡处理）与分类模型的系统性整合。该框架回应了现有研究中对端到端流程优化的呼吁，并通过多模型公平对比为算法选择提供了参考基准。",[1091,2841,2842],{},"从实践层面，本研究为保险机构提供了可操作的自动化检测工具。该模型可集成至理赔审核系统，实现对潜在欺诈案件的优先筛选，从而降低经济损失并优化资源分配。同时，特征重要性分析为业务人员提供了明确的的风险信号指引，有助于构建分层审核机制，提升整体风控效率。",[1382,2844,2846],{"id":2845},"_53-研究局限性","5.3 研究局限性",[1091,2848,2849],{},"本研究仍存在若干局限性：",[1183,2851,2852,2855,2858],{},[1186,2853,2854],{},"数据集规模较小（仅1000条记录），且欺诈样本比例（约25%）高于实际行业水平，可能影响模型在极度不平衡场景下的泛化性能。",[1186,2856,2857],{},"特征集主要限于理赔环节结构化数据，未能融入多源异构信息（如事故照片、投保人信用记录），限制了模型的判别潜力。",[1186,2859,2860],{},"研究聚焦传统机器学习方法，未涉及深度学习或混合模型的对比。",[1382,2862,2864],{"id":2863},"_54-未来研究展望","5.4 未来研究展望",[1091,2866,2867],{},"基于本研究的基础，未来可从在更大规模的企业内部数据集上验证框架，评估其在真实分布下的稳健性与部署效果，同时也可以扩展至多模态数据融合（如结合图像与文本信息）或在线学习机制，以应对欺诈模式的动态演变。从而汽车保险欺诈检测技术有望向更高精度、更强解释性和更广适用性的方向演进，为行业数字化风险管理贡献更大价值。",[1091,2869,2870],{},"目前的工作展现了传统机器学习方法在保险欺诈检测中的应用潜力，期望为相关领域的理论发展与实践创新提供有益参考。",[1183,2872,2873,2879,2885,2891,2897,2903,2909,2915,2921,2927,2933,2939],{},[1186,2874,2875,2878],{},[1109,2876,2877],{},"Barry, S., & Charpentier, A. (2020)."," Machine Learning for Insurance. CRC Press. (经典教材，奠定背景)",[1186,2880,2881,2884],{},[1109,2882,2883],{},"Bhattacharyya, S., et al. (2011)."," Data mining for credit card fraud: A comparative study. Decision Support Systems. (关于数据驱动欺诈检测的经典对比研究)",[1186,2886,2887,2890],{},[1109,2888,2889],{},"Breiman, L. (2001)."," Random Forests. Machine Learning. (随机森林算法的开创性文献)",[1186,2892,2893,2896],{},[1109,2894,2895],{},"CAIF (2022)."," The Impact of Insurance Fraud on the U.S. Economy. Coalition Against Insurance Fraud. (权威行业损失数据)",[1186,2898,2899,2902],{},[1109,2900,2901],{},"Chawla, N. V., et al. (2002)."," SMOTE: Synthetic Minority Over-sampling Technique. JAIR. (处理不平衡数据的基石研究)",[1186,2904,2905,2908],{},[1109,2906,2907],{},"Cohen, A., & Siegelman, P. (2010)."," Testing for Adverse Selection in Insurance Markets. Journal of Risk and Insurance. (信息不对称理论应用)",[1186,2910,2911,2914],{},[1109,2912,2913],{},"Derrig, R. A. (2002)."," Insurance Fraud. Journal of Risk and Insurance. (保险欺诈定义的权威来源)",[1186,2916,2917,2920],{},[1109,2918,2919],{},"Gomes, H. M., et al. (2021)."," Machine learning for streaming data: state of the art, challenges, and opportunities. ACM SIGKDD. (讨论端到端流程的现代挑战)",[1186,2922,2923,2926],{},[1109,2924,2925],{},"He, H., & Garcia, E. A. (2009)."," Learning from Imbalanced Data. IEEE Transactions on Knowledge and Data Engineering. (不平衡学习的系统性评述)",[1186,2928,2929,2932],{},[1109,2930,2931],{},"OECD (2023)."," Global Insurance Market Trends 2023. (全球保险市场宏观背景)",[1186,2934,2935,2938],{},[1109,2936,2937],{},"Severino, M. K., & Peng, Y. (2021)."," Machine learning algorithms for fraud detection in property-casualty insurance: A review. Decision Support Systems. (最新的车险欺诈综述，非常契合本文)",[1186,2940,2941,2944],{},[1109,2942,2943],{},"Xia, X., et al. (2023)."," Random forest-based fraud detection in automobile insurance. Expert Systems with Applications. (2023年针对随机森林在车险应用的最新实证研究)",[1098,2946],{},[1098,2948],{},[1183,2950,2951,2956,2962,2968,2974,2980,2986,2992,2998,3003,3008,3014,3019,3024,3030],{},[1186,2952,2953,2955],{},[1109,2954,2889],{}," Random Forests. Machine Learning. (奠基性引用)",[1186,2957,2958,2961],{},[1109,2959,2960],{},"CAIF. (2022)."," The Impact of Insurance Fraud on the U.S. Economy. (最新行业数据)",[1186,2963,2964,2967],{},[1109,2965,2966],{},"Chen, T., & Guestrin, C. (2016)."," XGBoost: A Scalable Tree Boosting System. SIGKDD. (XGBoost 核心文献)",[1186,2969,2970,2973],{},[1109,2971,2972],{},"Han, W., et al. (2022)."," Learning from imbalanced data: A comparative study of SMOTE and its variations. Information Sciences. (重采样技术的最新对比)",[1186,2975,2976,2979],{},[1109,2977,2978],{},"Johnson, J. M., & Khoshgoftaar, T. M. (2019)."," Survey on deep learning with class imbalance. Journal of Big Data. (虽然讨论深度学习，但对不平衡问题的总结非常权威)",[1186,2981,2982,2985],{},[1109,2983,2984],{},"Jovanovic, M., et al. (2022)."," Building an efficient fraud detection system in the insurance industry. Decision Support Systems. (针对车险的最新应用研究)",[1186,2987,2988,2991],{},[1109,2989,2990],{},"Ke, G., et al. (2017)."," LightGBM: A Highly Efficient Gradient Boosting Decision Tree. NeurIPS. (LightGBM 核心文献)",[1186,2993,2994,2997],{},[1109,2995,2996],{},"Nordin, N., et al. (2024)."," Machine learning for insurance fraud detection: A performance-based comparative analysis. Financial Innovation. (2024 年最新的多算法对比实证研究)",[1186,2999,3000,3002],{},[1109,3001,2937],{}," Machine learning algorithms for fraud detection in property-casualty insurance: A review. Decision Support Systems. (该领域最权威的最新综述)",[1186,3004,3005,3007],{},[1109,3006,2925],{}," Learning from Imbalanced Data. (不平衡学习领域的引用率最高、最权威的综述，定义了该问题的基本挑战)。",[1186,3009,3010,3013],{},[1109,3011,3012],{},"Dal Pozzolo, A., et al. (2015)."," Calibrating Probability with Undersampling for Unbalanced Classification. (论证了非对称代价函数和不平衡分类在反欺诈中的关联)。",[1186,3015,3016,3018],{},[1109,3017,2937],{}," Machine learning algorithms for fraud detection in property-casualty insurance: A review. (2021年最新的车险欺诈综述，专门强调了漏检欺诈的昂贵代价)。",[1186,3020,3021,3023],{},[1109,3022,2901],{}," SMOTE: Synthetic Minority Over-sampling Technique. (SMOTE 算法的开创性文献，必须保留)。",[1186,3025,3026,3029],{},[1109,3027,3028],{},"Fernández, A., et al. (2018)."," SMOTE for Learning from Imbalanced Data: Progress and Challenges, Marking the 15-year Anniversary. (发表于《Nature》子刊或顶级期刊的综述，对 SMOTE 15年来的应用进行了权威总结，证明其在结构化数据上的稳健性)。",[1186,3031,3032,3034],{},[1109,3033,2984],{}," Building an efficient fraud detection system in the insurance industry. (针对保险行业最新的实证研究，证明了 SMOTE 在提升车险欺诈检测召回率方面的实效)。",[1439,3036,3037,3042,3047,3052,3057,3062,3067,3072,3077,3082,3087,3092,3097],{},[1186,3038,3039,3041],{},[1109,3040,2883],{}," Data mining for credit card fraud: A comparative study. Decision Support Systems.",[1186,3043,3044,3046],{},[1109,3045,2889],{}," Random Forests. Machine Learning.",[1186,3048,3049,3051],{},[1109,3050,2895],{}," The Impact of Insurance Fraud on the U.S. Economy. Coalition Against Insurance Fraud.",[1186,3053,3054,3056],{},[1109,3055,2901],{}," SMOTE: Synthetic Minority Over-sampling Technique. Journal of Artificial Intelligence Research (JAIR).",[1186,3058,3059,3061],{},[1109,3060,2907],{}," Testing for Adverse Selection in Insurance Markets. Journal of Risk and Insurance.",[1186,3063,3064,3066],{},[1109,3065,2913],{}," Insurance Fraud. Journal of Risk and Insurance.",[1186,3068,3069,3071],{},[1109,3070,2925],{}," Learning from Imbalanced Data. IEEE Transactions on Knowledge and Data Engineering.",[1186,3073,3074,3076],{},[1109,3075,2931],{}," Global Insurance Market Trends 2023. Organisation for Economic Co-operation and Development.",[1186,3078,3079,3081],{},[1109,3080,2966],{}," XGBoost: A Scalable Tree Boosting System. Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining.",[1186,3083,3084,3086],{},[1109,3085,2978],{}," Survey on deep learning with class imbalance. Journal of Big Data.",[1186,3088,3089,3091],{},[1109,3090,2990],{}," LightGBM: A Highly Efficient Gradient Boosting Decision Tree. Advances in Neural Information Processing Systems (NeurIPS).",[1186,3093,3094,3096],{},[1109,3095,3012],{}," Calibrating Probability with Undersampling for Unbalanced Classification. IEEE Symposium Series on Computational Intelligence.",[1186,3098,3099,3101],{},[1109,3100,3028],{}," SMOTE for Learning from Imbalanced Data: Progress and Challenges, Marking the 15-year Anniversary. Journal of Artificial Intelligence Research.",{"title":3103,"searchDepth":3104,"depth":3104,"links":3105},"",4,[3106,3108,3109,3110,3111,3118,3122],{"id":1223,"depth":3107,"text":1224},2,{"id":1244,"depth":3107,"text":1245},{"id":1284,"depth":3107,"text":1285},{"id":1305,"depth":3107,"text":1306},{"id":1337,"depth":3107,"text":1338,"children":3112},[3113],{"id":1363,"depth":3114,"text":1364,"children":3115},3,[3116,3117],{"id":1384,"depth":3104,"text":1385},{"id":1403,"depth":3104,"text":1404},{"id":1413,"depth":3107,"text":1414,"children":3119},[3120,3121],{"id":1420,"depth":3114,"text":1421},{"id":1427,"depth":3114,"text":1428},{"id":1458,"depth":3107,"text":1459,"children":3123},[3124,3125,3126,3134],{"id":1462,"depth":3114,"text":1463},{"id":1487,"depth":3114,"text":1488},{"id":2321,"depth":3114,"text":2322,"children":3127},[3128,3129,3130,3131,3132,3133],{"id":2334,"depth":3104,"text":2335},{"id":2549,"depth":3104,"text":2550},{"id":2570,"depth":3104,"text":2571},{"id":2642,"depth":3104,"text":2643},{"id":2723,"depth":3104,"text":2724},{"id":2797,"depth":3104,"text":2798},{"id":2807,"depth":3114,"text":2808,"children":3135},[3136,3137,3138,3139],{"id":2811,"depth":3104,"text":2812},{"id":2835,"depth":3104,"text":2836},{"id":2845,"depth":3104,"text":2846},{"id":2863,"depth":3104,"text":2864},"md",true,{"uuid":3143,"slots":3144},"f3561dd0-e86a-11f0-8fd1-914b1152d8ed",{},35,{"title":1085,"description":1093},"posts\u002F2026\u002F2026-01-03-论文实训草稿",[230,231,232,233],"OJYroFHRncV3kO6Mb3QALZrS6vq9VAN0taKxW4xSe0w",1790443286648]