7989f3a159
后端新增风险图谱算法模块、风险观察与反馈服务、规则 DSL 校验器和可解释性引擎，完善系统仪表盘和财务仪表盘统计， 优化 agent 运行和编排执行链路，清理旧开发文档，前端新增 系统趋势、负载热力图等多种仪表盘图表组件，完善操作反馈 对话框和工作台日期选择器，优化报销创建和审批详情交互， 补充单元测试覆盖。
1484 lines
48 KiB
HTML
1484 lines
48 KiB
HTML
<!doctype html>
|
||
<html lang="zh-CN">
|
||
<head>
|
||
<meta charset="utf-8">
|
||
<meta name="viewport" content="width=device-width, initial-scale=1">
|
||
<title>X-Financial 财务行为图谱风险引擎核心算法说明</title>
|
||
<style>
|
||
:root {
|
||
color-scheme: light;
|
||
--bg: #f5f7f8;
|
||
--surface: #ffffff;
|
||
--surface-soft: #f8fbfc;
|
||
--surface-strong: #edf3f5;
|
||
--ink: #172126;
|
||
--ink-soft: #44525a;
|
||
--muted: #6b7a83;
|
||
--line: #dce5ea;
|
||
--line-strong: #c2d0d8;
|
||
--blue: #235a8f;
|
||
--blue-deep: #153e66;
|
||
--blue-soft: #e8f1f8;
|
||
--teal: #16756f;
|
||
--teal-soft: #e7f4f2;
|
||
--green: #28784a;
|
||
--green-soft: #e9f5ed;
|
||
--amber: #9a6200;
|
||
--amber-soft: #fff4dd;
|
||
--red: #b43d32;
|
||
--red-soft: #fff0ee;
|
||
--violet: #5b4b9a;
|
||
--violet-soft: #f0eef9;
|
||
--shadow: 0 18px 44px rgba(24, 38, 48, 0.08);
|
||
--shadow-soft: 0 10px 26px rgba(24, 38, 48, 0.06);
|
||
--radius: 8px;
|
||
--font: "Microsoft YaHei UI", "Microsoft YaHei", "PingFang SC", "Segoe UI", sans-serif;
|
||
--mono: "Cascadia Code", "JetBrains Mono", Consolas, monospace;
|
||
}
|
||
|
||
* {
|
||
box-sizing: border-box;
|
||
}
|
||
|
||
html {
|
||
scroll-behavior: smooth;
|
||
}
|
||
|
||
body {
|
||
margin: 0;
|
||
background:
|
||
linear-gradient(180deg, rgba(232, 241, 248, 0.92), rgba(245, 247, 248, 0.98) 420px),
|
||
var(--bg);
|
||
color: var(--ink);
|
||
font-family: var(--font);
|
||
font-size: 14px;
|
||
line-height: 1.72;
|
||
letter-spacing: 0;
|
||
overflow-wrap: anywhere;
|
||
}
|
||
|
||
a {
|
||
color: inherit;
|
||
text-decoration: none;
|
||
}
|
||
|
||
code {
|
||
padding: 2px 6px;
|
||
border-radius: 6px;
|
||
background: rgba(35, 90, 143, 0.08);
|
||
color: var(--blue-deep);
|
||
font-family: var(--mono);
|
||
font-size: 0.92em;
|
||
}
|
||
|
||
pre {
|
||
margin: 0;
|
||
padding: 16px;
|
||
overflow: auto;
|
||
border: 1px solid #20313c;
|
||
border-radius: var(--radius);
|
||
background: #101920;
|
||
color: #eef7fb;
|
||
font-family: var(--mono);
|
||
font-size: 12.5px;
|
||
line-height: 1.72;
|
||
white-space: pre;
|
||
}
|
||
|
||
.page {
|
||
display: grid;
|
||
grid-template-columns: 278px minmax(0, 1fr) 260px;
|
||
min-height: 100dvh;
|
||
}
|
||
|
||
.sidebar,
|
||
.aside {
|
||
position: sticky;
|
||
top: 0;
|
||
height: 100dvh;
|
||
overflow: auto;
|
||
background: rgba(255, 255, 255, 0.92);
|
||
backdrop-filter: blur(16px);
|
||
}
|
||
|
||
.sidebar {
|
||
padding: 24px 18px;
|
||
border-right: 1px solid var(--line);
|
||
}
|
||
|
||
.aside {
|
||
padding: 24px 18px;
|
||
border-left: 1px solid var(--line);
|
||
}
|
||
|
||
main {
|
||
min-width: 0;
|
||
padding: 34px 42px 64px;
|
||
}
|
||
|
||
.brand {
|
||
display: grid;
|
||
gap: 12px;
|
||
padding-bottom: 20px;
|
||
border-bottom: 1px solid var(--line);
|
||
}
|
||
|
||
.brand-row {
|
||
display: flex;
|
||
align-items: center;
|
||
gap: 12px;
|
||
}
|
||
|
||
.brand-mark {
|
||
display: grid;
|
||
width: 44px;
|
||
height: 44px;
|
||
place-items: center;
|
||
border-radius: 8px;
|
||
background: #142c3f;
|
||
color: #fff;
|
||
font-size: 13px;
|
||
font-weight: 900;
|
||
box-shadow: 0 14px 28px rgba(20, 44, 63, 0.24);
|
||
}
|
||
|
||
.brand-title {
|
||
color: var(--ink);
|
||
font-size: 15px;
|
||
font-weight: 900;
|
||
line-height: 1.28;
|
||
}
|
||
|
||
.brand-subtitle {
|
||
color: var(--muted);
|
||
font-size: 12px;
|
||
line-height: 1.58;
|
||
}
|
||
|
||
.nav-label,
|
||
.aside-title {
|
||
margin: 22px 0 10px;
|
||
color: var(--muted);
|
||
font-size: 12px;
|
||
font-weight: 850;
|
||
}
|
||
|
||
.nav {
|
||
display: grid;
|
||
gap: 5px;
|
||
}
|
||
|
||
.nav a,
|
||
.mini-link {
|
||
display: flex;
|
||
align-items: center;
|
||
min-height: 34px;
|
||
padding: 7px 10px;
|
||
border-radius: 8px;
|
||
color: var(--ink-soft);
|
||
font-size: 13px;
|
||
font-weight: 760;
|
||
}
|
||
|
||
.nav a:hover,
|
||
.mini-link:hover {
|
||
background: var(--blue-soft);
|
||
color: var(--blue-deep);
|
||
}
|
||
|
||
.dot {
|
||
flex: 0 0 auto;
|
||
width: 7px;
|
||
height: 7px;
|
||
margin-right: 10px;
|
||
border-radius: 99px;
|
||
background: var(--line-strong);
|
||
}
|
||
|
||
.nav a:hover .dot,
|
||
.mini-link:hover .dot {
|
||
background: var(--blue);
|
||
}
|
||
|
||
.side-note,
|
||
.aside-card {
|
||
margin-top: 18px;
|
||
padding: 13px;
|
||
border: 1px solid var(--line);
|
||
border-radius: var(--radius);
|
||
background: var(--surface-soft);
|
||
color: var(--ink-soft);
|
||
font-size: 12px;
|
||
line-height: 1.62;
|
||
}
|
||
|
||
.hero {
|
||
overflow: hidden;
|
||
border: 1px solid var(--line);
|
||
border-radius: var(--radius);
|
||
background:
|
||
linear-gradient(135deg, rgba(255,255,255,0.98), rgba(232,241,248,0.92)),
|
||
var(--surface);
|
||
box-shadow: var(--shadow);
|
||
}
|
||
|
||
.hero-inner {
|
||
padding: 34px;
|
||
}
|
||
|
||
.eyebrow {
|
||
display: inline-flex;
|
||
align-items: center;
|
||
gap: 8px;
|
||
margin-bottom: 14px;
|
||
padding: 5px 10px;
|
||
border: 1px solid rgba(35, 90, 143, 0.18);
|
||
border-radius: 999px;
|
||
background: var(--blue-soft);
|
||
color: var(--blue-deep);
|
||
font-size: 12px;
|
||
font-weight: 850;
|
||
}
|
||
|
||
h1 {
|
||
max-width: 900px;
|
||
margin: 0;
|
||
color: #101922;
|
||
font-size: clamp(30px, 4.5vw, 56px);
|
||
line-height: 1.08;
|
||
letter-spacing: 0;
|
||
}
|
||
|
||
.hero-lead {
|
||
max-width: 920px;
|
||
margin: 18px 0 0;
|
||
color: var(--ink-soft);
|
||
font-size: 16px;
|
||
line-height: 1.78;
|
||
}
|
||
|
||
.hero-metrics {
|
||
display: grid;
|
||
grid-template-columns: repeat(4, minmax(0, 1fr));
|
||
gap: 10px;
|
||
margin-top: 26px;
|
||
}
|
||
|
||
.metric {
|
||
min-height: 104px;
|
||
padding: 14px;
|
||
border: 1px solid var(--line);
|
||
border-radius: var(--radius);
|
||
background: rgba(255,255,255,0.88);
|
||
}
|
||
|
||
.metric-value {
|
||
color: var(--blue-deep);
|
||
font-size: 24px;
|
||
font-weight: 950;
|
||
line-height: 1.1;
|
||
}
|
||
|
||
.metric-label {
|
||
margin-top: 7px;
|
||
color: var(--ink);
|
||
font-size: 13px;
|
||
font-weight: 850;
|
||
}
|
||
|
||
.metric-desc {
|
||
margin-top: 5px;
|
||
color: var(--muted);
|
||
font-size: 12px;
|
||
line-height: 1.5;
|
||
}
|
||
|
||
section {
|
||
margin-top: 22px;
|
||
border: 1px solid var(--line);
|
||
border-radius: var(--radius);
|
||
background: var(--surface);
|
||
box-shadow: var(--shadow-soft);
|
||
}
|
||
|
||
.section-inner {
|
||
padding: 26px;
|
||
}
|
||
|
||
.section-head {
|
||
display: flex;
|
||
align-items: flex-start;
|
||
justify-content: space-between;
|
||
gap: 16px;
|
||
margin-bottom: 18px;
|
||
padding-bottom: 14px;
|
||
border-bottom: 1px solid var(--line);
|
||
}
|
||
|
||
h2 {
|
||
margin: 0;
|
||
color: #101922;
|
||
font-size: 24px;
|
||
line-height: 1.25;
|
||
letter-spacing: 0;
|
||
}
|
||
|
||
h3 {
|
||
margin: 0 0 10px;
|
||
color: #172126;
|
||
font-size: 17px;
|
||
line-height: 1.35;
|
||
}
|
||
|
||
h4 {
|
||
margin: 0 0 8px;
|
||
color: var(--ink);
|
||
font-size: 14px;
|
||
font-weight: 900;
|
||
}
|
||
|
||
.section-kicker {
|
||
margin-top: 8px;
|
||
color: var(--muted);
|
||
font-size: 13px;
|
||
line-height: 1.66;
|
||
}
|
||
|
||
.tag {
|
||
display: inline-flex;
|
||
align-items: center;
|
||
min-height: 26px;
|
||
padding: 0 9px;
|
||
border-radius: 999px;
|
||
background: var(--surface-strong);
|
||
color: var(--ink-soft);
|
||
font-size: 12px;
|
||
font-weight: 850;
|
||
white-space: nowrap;
|
||
}
|
||
|
||
.grid {
|
||
display: grid;
|
||
gap: 14px;
|
||
}
|
||
|
||
.grid.two {
|
||
grid-template-columns: repeat(2, minmax(0, 1fr));
|
||
}
|
||
|
||
.grid.three {
|
||
grid-template-columns: repeat(3, minmax(0, 1fr));
|
||
}
|
||
|
||
.grid.four {
|
||
grid-template-columns: repeat(4, minmax(0, 1fr));
|
||
}
|
||
|
||
.card {
|
||
padding: 16px;
|
||
border: 1px solid var(--line);
|
||
border-radius: var(--radius);
|
||
background: var(--surface-soft);
|
||
}
|
||
|
||
.card.blue { background: var(--blue-soft); border-color: #c8d9e8; }
|
||
.card.green { background: var(--green-soft); border-color: #c9dfd0; }
|
||
.card.amber { background: var(--amber-soft); border-color: #ead4a9; }
|
||
.card.red { background: var(--red-soft); border-color: #edc7c2; }
|
||
.card.violet { background: var(--violet-soft); border-color: #d7d1eb; }
|
||
.card.teal { background: var(--teal-soft); border-color: #c9dfdc; }
|
||
|
||
.card p,
|
||
.note p {
|
||
margin: 0;
|
||
color: var(--ink-soft);
|
||
}
|
||
|
||
.list {
|
||
display: grid;
|
||
gap: 8px;
|
||
margin: 0;
|
||
padding: 0;
|
||
list-style: none;
|
||
}
|
||
|
||
.list li {
|
||
position: relative;
|
||
padding-left: 18px;
|
||
color: var(--ink-soft);
|
||
}
|
||
|
||
.list li::before {
|
||
content: "";
|
||
position: absolute;
|
||
left: 0;
|
||
top: 12px;
|
||
width: 7px;
|
||
height: 7px;
|
||
border-radius: 99px;
|
||
background: var(--blue);
|
||
}
|
||
|
||
.flow {
|
||
display: grid;
|
||
gap: 10px;
|
||
}
|
||
|
||
.flow-row {
|
||
display: grid;
|
||
grid-template-columns: repeat(5, minmax(0, 1fr));
|
||
gap: 10px;
|
||
}
|
||
|
||
.flow-node {
|
||
min-height: 84px;
|
||
display: grid;
|
||
align-content: start;
|
||
gap: 6px;
|
||
padding: 13px;
|
||
border: 1px solid var(--line);
|
||
border-radius: var(--radius);
|
||
background: #fff;
|
||
}
|
||
|
||
.flow-node strong {
|
||
color: var(--ink);
|
||
font-size: 13px;
|
||
}
|
||
|
||
.flow-node span {
|
||
color: var(--muted);
|
||
font-size: 12px;
|
||
line-height: 1.45;
|
||
}
|
||
|
||
.arrow {
|
||
display: grid;
|
||
place-items: center;
|
||
min-height: 28px;
|
||
color: var(--muted);
|
||
font-size: 12px;
|
||
font-weight: 850;
|
||
}
|
||
|
||
.pipeline {
|
||
display: grid;
|
||
gap: 12px;
|
||
}
|
||
|
||
.pipeline-stage {
|
||
display: grid;
|
||
grid-template-columns: 168px minmax(0, 1fr);
|
||
gap: 12px;
|
||
align-items: stretch;
|
||
}
|
||
|
||
.stage-label {
|
||
display: grid;
|
||
place-items: center;
|
||
min-height: 92px;
|
||
padding: 14px;
|
||
border-radius: var(--radius);
|
||
background: #132c3e;
|
||
color: #fff;
|
||
text-align: center;
|
||
font-weight: 900;
|
||
}
|
||
|
||
.stage-content {
|
||
display: grid;
|
||
grid-template-columns: repeat(3, minmax(0, 1fr));
|
||
gap: 10px;
|
||
}
|
||
|
||
.formula {
|
||
padding: 18px;
|
||
border: 1px solid #bfd2df;
|
||
border-radius: var(--radius);
|
||
background: #f5f9fc;
|
||
}
|
||
|
||
.formula-main {
|
||
padding: 16px;
|
||
overflow-x: auto;
|
||
border-radius: var(--radius);
|
||
background: #102033;
|
||
color: #eff8ff;
|
||
font-family: var(--mono);
|
||
font-size: 14px;
|
||
line-height: 1.75;
|
||
white-space: nowrap;
|
||
}
|
||
|
||
.formula-note {
|
||
margin-top: 10px;
|
||
color: var(--ink-soft);
|
||
font-size: 13px;
|
||
}
|
||
|
||
.matrix {
|
||
overflow: auto;
|
||
border: 1px solid var(--line);
|
||
border-radius: var(--radius);
|
||
}
|
||
|
||
table {
|
||
width: 100%;
|
||
min-width: 760px;
|
||
border-collapse: collapse;
|
||
background: #fff;
|
||
}
|
||
|
||
th,
|
||
td {
|
||
padding: 12px 13px;
|
||
border-bottom: 1px solid var(--line);
|
||
color: var(--ink-soft);
|
||
text-align: left;
|
||
vertical-align: top;
|
||
}
|
||
|
||
th {
|
||
background: #f3f7f9;
|
||
color: var(--ink);
|
||
font-size: 12px;
|
||
font-weight: 900;
|
||
white-space: nowrap;
|
||
}
|
||
|
||
tr:last-child td {
|
||
border-bottom: 0;
|
||
}
|
||
|
||
.asset-map {
|
||
display: grid;
|
||
grid-template-columns: repeat(5, minmax(0, 1fr));
|
||
gap: 10px;
|
||
}
|
||
|
||
.asset {
|
||
min-height: 138px;
|
||
padding: 14px;
|
||
border: 1px solid var(--line);
|
||
border-radius: var(--radius);
|
||
background: #fff;
|
||
}
|
||
|
||
.asset-num {
|
||
display: inline-grid;
|
||
width: 28px;
|
||
height: 28px;
|
||
place-items: center;
|
||
margin-bottom: 10px;
|
||
border-radius: 7px;
|
||
background: var(--blue);
|
||
color: #fff;
|
||
font-weight: 900;
|
||
}
|
||
|
||
.evidence-path {
|
||
display: grid;
|
||
grid-template-columns: repeat(7, minmax(0, 1fr));
|
||
gap: 8px;
|
||
align-items: stretch;
|
||
}
|
||
|
||
.evidence-item {
|
||
min-height: 86px;
|
||
padding: 12px;
|
||
border: 1px solid var(--line);
|
||
border-radius: var(--radius);
|
||
background: #fff;
|
||
}
|
||
|
||
.evidence-item strong {
|
||
display: block;
|
||
color: var(--ink);
|
||
font-size: 12px;
|
||
}
|
||
|
||
.evidence-item span {
|
||
display: block;
|
||
margin-top: 5px;
|
||
color: var(--muted);
|
||
font-size: 11px;
|
||
line-height: 1.45;
|
||
}
|
||
|
||
.split {
|
||
display: grid;
|
||
grid-template-columns: minmax(0, 1.05fr) minmax(0, 0.95fr);
|
||
gap: 16px;
|
||
align-items: start;
|
||
}
|
||
|
||
.callout {
|
||
padding: 15px 16px;
|
||
border-left: 4px solid var(--blue);
|
||
border-radius: var(--radius);
|
||
background: #f4f9fc;
|
||
color: var(--ink-soft);
|
||
}
|
||
|
||
.callout strong {
|
||
color: var(--ink);
|
||
}
|
||
|
||
.step-list {
|
||
display: grid;
|
||
gap: 12px;
|
||
counter-reset: steps;
|
||
}
|
||
|
||
.step {
|
||
display: grid;
|
||
grid-template-columns: 42px minmax(0, 1fr);
|
||
gap: 12px;
|
||
padding: 14px;
|
||
border: 1px solid var(--line);
|
||
border-radius: var(--radius);
|
||
background: var(--surface-soft);
|
||
}
|
||
|
||
.step::before {
|
||
counter-increment: steps;
|
||
content: counter(steps);
|
||
display: grid;
|
||
width: 34px;
|
||
height: 34px;
|
||
place-items: center;
|
||
border-radius: 8px;
|
||
background: #132c3e;
|
||
color: #fff;
|
||
font-weight: 900;
|
||
}
|
||
|
||
.step strong {
|
||
color: var(--ink);
|
||
}
|
||
|
||
.step p {
|
||
margin: 4px 0 0;
|
||
color: var(--ink-soft);
|
||
}
|
||
|
||
.sources {
|
||
display: grid;
|
||
gap: 8px;
|
||
margin: 0;
|
||
padding: 0;
|
||
list-style: none;
|
||
}
|
||
|
||
.sources a {
|
||
display: block;
|
||
padding: 10px 12px;
|
||
border: 1px solid var(--line);
|
||
border-radius: var(--radius);
|
||
background: var(--surface-soft);
|
||
color: var(--blue-deep);
|
||
font-size: 12px;
|
||
font-weight: 760;
|
||
}
|
||
|
||
.sources a:hover {
|
||
background: var(--blue-soft);
|
||
}
|
||
|
||
.print-note {
|
||
margin-top: 18px;
|
||
color: var(--muted);
|
||
font-size: 12px;
|
||
}
|
||
|
||
@media (max-width: 1320px) {
|
||
.page {
|
||
grid-template-columns: 250px minmax(0, 1fr);
|
||
}
|
||
.aside {
|
||
display: none;
|
||
}
|
||
.hero-metrics,
|
||
.grid.four,
|
||
.asset-map {
|
||
grid-template-columns: repeat(2, minmax(0, 1fr));
|
||
}
|
||
.evidence-path {
|
||
grid-template-columns: repeat(4, minmax(0, 1fr));
|
||
}
|
||
}
|
||
|
||
@media (max-width: 980px) {
|
||
.page {
|
||
display: block;
|
||
}
|
||
.sidebar {
|
||
position: static;
|
||
height: auto;
|
||
border-right: 0;
|
||
border-bottom: 1px solid var(--line);
|
||
}
|
||
main {
|
||
padding: 24px 18px 44px;
|
||
}
|
||
.hero-inner,
|
||
.section-inner {
|
||
padding: 20px;
|
||
}
|
||
.grid.two,
|
||
.grid.three,
|
||
.grid.four,
|
||
.hero-metrics,
|
||
.flow-row,
|
||
.stage-content,
|
||
.asset-map,
|
||
.split {
|
||
grid-template-columns: 1fr;
|
||
}
|
||
.pipeline-stage {
|
||
grid-template-columns: 1fr;
|
||
}
|
||
.evidence-path {
|
||
grid-template-columns: 1fr;
|
||
}
|
||
}
|
||
|
||
@media print {
|
||
body {
|
||
background: #fff;
|
||
}
|
||
.page {
|
||
display: block;
|
||
}
|
||
.sidebar,
|
||
.aside {
|
||
display: none;
|
||
}
|
||
main {
|
||
padding: 0;
|
||
}
|
||
section,
|
||
.hero {
|
||
box-shadow: none;
|
||
break-inside: avoid;
|
||
}
|
||
}
|
||
</style>
|
||
</head>
|
||
<body>
|
||
<div class="page">
|
||
<aside class="sidebar">
|
||
<div class="brand">
|
||
<div class="brand-row">
|
||
<div class="brand-mark">RGE</div>
|
||
<div>
|
||
<div class="brand-title">财务行为图谱风险引擎</div>
|
||
<div class="brand-subtitle">Risk Graph Engine</div>
|
||
</div>
|
||
</div>
|
||
<div class="brand-subtitle">
|
||
X-Financial 核心算法说明文档。版本：2026-05-29。
|
||
</div>
|
||
</div>
|
||
|
||
<div class="nav-label">目录</div>
|
||
<nav class="nav" aria-label="文档目录">
|
||
<a href="#summary"><span class="dot"></span>核心定义</a>
|
||
<a href="#barrier"><span class="dot"></span>不可复制壁垒</a>
|
||
<a href="#architecture"><span class="dot"></span>总体架构</a>
|
||
<a href="#ontology"><span class="dot"></span>本体桥接</a>
|
||
<a href="#observation"><span class="dot"></span>风险观察</a>
|
||
<a href="#scoring"><span class="dot"></span>评分公式</a>
|
||
<a href="#evidence"><span class="dot"></span>证据链</a>
|
||
<a href="#deep"><span class="dot"></span>深度算法层</a>
|
||
<a href="#gate"><span class="dot"></span>人机共审门控</a>
|
||
<a href="#dashboard"><span class="dot"></span>产品体现</a>
|
||
<a href="#roadmap"><span class="dot"></span>实施路线</a>
|
||
<a href="#sources"><span class="dot"></span>参考资料</a>
|
||
</nav>
|
||
|
||
<div class="side-note">
|
||
本文档是给业务、算法、后端、前端和管理层共同阅读的核心算法介绍。它不替代 `CONCEPT.md` 与 `TODO.md`，而是把核心算法讲成完整方案。
|
||
</div>
|
||
</aside>
|
||
|
||
<main>
|
||
<article class="hero" id="summary">
|
||
<div class="hero-inner">
|
||
<div class="eyebrow">核心中的核心算法</div>
|
||
<h1>财务行为图谱风险引擎</h1>
|
||
<p class="hero-lead">
|
||
这套算法不是单一模型，而是把专有本体、对象中心事件日志、财务行为图谱、画像基线、规则命中、人工反馈和可回放评测集组合成一套持续进化的风险控制系统。它的目标不是“给一个风险分”，而是解释清楚谁异常、哪里异常、依据是什么、历史如何处理、现在应该怎么做。
|
||
</p>
|
||
<div class="hero-metrics">
|
||
<div class="metric">
|
||
<div class="metric-value">5</div>
|
||
<div class="metric-label">不可复制资产</div>
|
||
<div class="metric-desc">专有本体、事件日志、反馈池、人机共审、回放评测。</div>
|
||
</div>
|
||
<div class="metric">
|
||
<div class="metric-value">8</div>
|
||
<div class="metric-label">第一版风险信号</div>
|
||
<div class="metric-desc">重复票据、超标准、预算超额、附件缺失等。</div>
|
||
</div>
|
||
<div class="metric">
|
||
<div class="metric-value">10+</div>
|
||
<div class="metric-label">算法模块</div>
|
||
<div class="metric-desc">本体、实体解析、过程挖掘、图特征、异常检测。</div>
|
||
</div>
|
||
<div class="metric">
|
||
<div class="metric-value">1</div>
|
||
<div class="metric-label">统一产物</div>
|
||
<div class="metric-desc">所有风险最终沉淀为 `RiskObservation`。</div>
|
||
</div>
|
||
</div>
|
||
</div>
|
||
</article>
|
||
|
||
<section id="barrier">
|
||
<div class="section-inner">
|
||
<div class="section-head">
|
||
<div>
|
||
<h2>不可复制壁垒</h2>
|
||
<div class="section-kicker">
|
||
复杂算法本身可以被模仿，真正难复制的是长期运行中积累的语义、过程、反馈和评测资产。
|
||
</div>
|
||
</div>
|
||
<span class="tag">Core Barrier</span>
|
||
</div>
|
||
|
||
<div class="asset-map">
|
||
<div class="asset">
|
||
<span class="asset-num">1</span>
|
||
<h3>专有财务语义本体</h3>
|
||
<p>把制度条款、费用类型、风险信号、审批场景、预算科目和票据要求归一为公司自己的语义协议。</p>
|
||
</div>
|
||
<div class="asset">
|
||
<span class="asset-num">2</span>
|
||
<h3>对象中心事件日志</h3>
|
||
<p>把申请、预算、票据、审批、付款、归档、复盘统一沉淀为可回放事件。</p>
|
||
</div>
|
||
<div class="asset">
|
||
<span class="asset-num">3</span>
|
||
<h3>风险观察反馈池</h3>
|
||
<p>每条风险都有证据路径、规则命中、画像偏离、图谱异常、人工结果和算法版本。</p>
|
||
</div>
|
||
<div class="asset">
|
||
<span class="asset-num">4</span>
|
||
<h3>人机共审行为数据</h3>
|
||
<p>记录采纳、驳回、补件、退回、升级审批、标记误报和候选规则生成。</p>
|
||
</div>
|
||
<div class="asset">
|
||
<span class="asset-num">5</span>
|
||
<h3>可回放评测资产</h3>
|
||
<p>用同一批历史单据、本体版本、规则版本、算法版本和反馈标签评估新算法。</p>
|
||
</div>
|
||
</div>
|
||
|
||
<div class="callout" style="margin-top: 16px;">
|
||
<strong>结论：</strong>别人可以复制页面、名词和公开算法，但复制不了真实审批过程、公司制度语义、历史处理结果和长期回放集。算法壁垒必须从第一天开始沉淀这些资产。
|
||
</div>
|
||
</div>
|
||
</section>
|
||
|
||
<section id="architecture">
|
||
<div class="section-inner">
|
||
<div class="section-head">
|
||
<div>
|
||
<h2>总体架构</h2>
|
||
<div class="section-kicker">
|
||
整体架构分为语义、事件、图谱、画像、推理、观察、反馈七层。每层只解决一个问题，避免规则中心、图谱引擎和数字员工职责混乱。
|
||
</div>
|
||
</div>
|
||
<span class="tag">Seven Layers</span>
|
||
</div>
|
||
|
||
<div class="pipeline">
|
||
<div class="pipeline-stage">
|
||
<div class="stage-label">语义层</div>
|
||
<div class="stage-content">
|
||
<div class="card blue"><h4>本体解析</h4><p>识别场景、意图、费用类型、风险信号、约束条件。</p></div>
|
||
<div class="card blue"><h4>制度条款</h4><p>把制度转成可引用、可版本化、可匹配的条款。</p></div>
|
||
<div class="card blue"><h4>标准词典</h4><p>统一供应商、费用类型、票据类型和风险信号。</p></div>
|
||
</div>
|
||
</div>
|
||
<div class="pipeline-stage">
|
||
<div class="stage-label">事件层</div>
|
||
<div class="stage-content">
|
||
<div class="card teal"><h4>业务事件</h4><p>申请、提交、上传票据、审批、退回、付款、归档。</p></div>
|
||
<div class="card teal"><h4>多对象日志</h4><p>一个事件同时绑定单据、票据、员工、供应商和预算。</p></div>
|
||
<div class="card teal"><h4>运行追踪</h4><p>关联 AgentRun、数字员工任务和工具调用结果。</p></div>
|
||
</div>
|
||
</div>
|
||
<div class="pipeline-stage">
|
||
<div class="stage-label">图谱层</div>
|
||
<div class="stage-content">
|
||
<div class="card green"><h4>节点</h4><p>员工、部门、供应商、票据、单据、制度条款、规则。</p></div>
|
||
<div class="card green"><h4>边</h4><p>提交、包含、使用票据、关联供应商、命中规则。</p></div>
|
||
<div class="card green"><h4>证据路径</h4><p>形成可展示、可审计、可回放的风险解释链。</p></div>
|
||
</div>
|
||
</div>
|
||
<div class="pipeline-stage">
|
||
<div class="stage-label">推理层</div>
|
||
<div class="stage-content">
|
||
<div class="card amber"><h4>规则命中</h4><p>确定性规则提供强证据，仍由规则中心治理。</p></div>
|
||
<div class="card amber"><h4>画像偏离</h4><p>员工、部门、供应商和费用类型的同类基线。</p></div>
|
||
<div class="card amber"><h4>图谱异常</h4><p>供应商集中、重复票据、审批链异常、时空冲突。</p></div>
|
||
</div>
|
||
</div>
|
||
<div class="pipeline-stage">
|
||
<div class="stage-label">反馈层</div>
|
||
<div class="stage-content">
|
||
<div class="card violet"><h4>人工确认</h4><p>采纳、误报、忽略、退回、补件、升级审批。</p></div>
|
||
<div class="card violet"><h4>离线回放</h4><p>用历史样本评估规则和算法版本。</p></div>
|
||
<div class="card violet"><h4>候选优化</h4><p>生成规则候选、制度更新候选和抽审策略调整。</p></div>
|
||
</div>
|
||
</div>
|
||
</div>
|
||
</div>
|
||
</section>
|
||
|
||
<section id="ontology">
|
||
<div class="section-inner">
|
||
<div class="section-head">
|
||
<div>
|
||
<h2>本体与风险图谱桥接</h2>
|
||
<div class="section-kicker">
|
||
本体不是一个辅助字段，而是风险图谱的语义骨架。它决定节点类型、边类型、规则适用范围、风险信号口径和看板聚合维度。
|
||
</div>
|
||
</div>
|
||
<span class="tag">Ontology First</span>
|
||
</div>
|
||
|
||
<div class="split">
|
||
<div class="card">
|
||
<h3>本体输出最小协议</h3>
|
||
<pre>ontology_parse_id
|
||
ontology_version
|
||
domain
|
||
scenario
|
||
intent
|
||
entities
|
||
constraints
|
||
risk_signals
|
||
confidence</pre>
|
||
</div>
|
||
<div class="card">
|
||
<h3>图谱节点标准字段</h3>
|
||
<pre>node_type
|
||
ontology_type
|
||
canonical_key
|
||
canonical_id
|
||
source_object_type
|
||
source_object_id
|
||
ontology_parse_id
|
||
ontology_version</pre>
|
||
</div>
|
||
</div>
|
||
|
||
<div class="matrix" style="margin-top: 16px;">
|
||
<table>
|
||
<thead>
|
||
<tr>
|
||
<th>本体实体</th>
|
||
<th>图谱节点</th>
|
||
<th>作用</th>
|
||
<th>示例</th>
|
||
</tr>
|
||
</thead>
|
||
<tbody>
|
||
<tr>
|
||
<td><code>expense_type</code></td>
|
||
<td>费用类型节点</td>
|
||
<td>统一规则范围、画像基线和风险看板聚合。</td>
|
||
<td><code>hotel</code>, <code>travel</code>, <code>entertainment</code></td>
|
||
</tr>
|
||
<tr>
|
||
<td><code>risk_signal</code></td>
|
||
<td>风险信号节点 / 风险观察</td>
|
||
<td>把“住宿超标”“酒店超标”等近义说法合并。</td>
|
||
<td><code>accommodation_standard_deviation</code></td>
|
||
</tr>
|
||
<tr>
|
||
<td><code>document_type</code></td>
|
||
<td>单据、票据、合同、流水</td>
|
||
<td>支撑多凭证一致性校验。</td>
|
||
<td><code>invoice</code>, <code>payment_record</code></td>
|
||
</tr>
|
||
<tr>
|
||
<td><code>policy_clause</code></td>
|
||
<td>制度条款节点</td>
|
||
<td>让风险结论能引用制度依据。</td>
|
||
<td>差旅住宿标准第 3 条</td>
|
||
</tr>
|
||
<tr>
|
||
<td><code>vendor</code></td>
|
||
<td>供应商 / 商户节点</td>
|
||
<td>支撑供应商集中、别名归一、关联交易识别。</td>
|
||
<td>上海某酒店管理有限公司</td>
|
||
</tr>
|
||
</tbody>
|
||
</table>
|
||
</div>
|
||
|
||
<div class="callout" style="margin-top: 16px;">
|
||
<strong>关键约束：</strong>图谱边必须来自白名单，例如 <code>submitted</code>、<code>uses_invoice</code>、<code>paid_to_vendor</code>、<code>matches_policy_clause</code>、<code>triggered_rule</code>。数字员工不能自由创造运行时边类型。
|
||
</div>
|
||
</div>
|
||
</section>
|
||
|
||
<section id="observation">
|
||
<div class="section-inner">
|
||
<div class="section-head">
|
||
<div>
|
||
<h2>RiskObservation 是唯一风险产物</h2>
|
||
<div class="section-kicker">
|
||
规则命中、画像偏离、图谱异常、制度条款和人工反馈最终都必须合成风险观察。它是单据详情、工作记录、风险看板和算法回放的统一事实表。
|
||
</div>
|
||
</div>
|
||
<span class="tag">Single Output</span>
|
||
</div>
|
||
|
||
<div class="grid three">
|
||
<div class="card blue">
|
||
<h3>输入</h3>
|
||
<ul class="list">
|
||
<li>本体解析结果</li>
|
||
<li>规则命中结果</li>
|
||
<li>画像基线偏离</li>
|
||
<li>图谱关系异常</li>
|
||
<li>历史反馈和相似案例</li>
|
||
</ul>
|
||
</div>
|
||
<div class="card green">
|
||
<h3>输出</h3>
|
||
<ul class="list">
|
||
<li>风险分和风险等级</li>
|
||
<li>证据链和制度依据</li>
|
||
<li>建议动作和自动化模式</li>
|
||
<li>算法版本和置信度</li>
|
||
<li>人工反馈状态</li>
|
||
</ul>
|
||
</div>
|
||
<div class="card amber">
|
||
<h3>用途</h3>
|
||
<ul class="list">
|
||
<li>单据详情解释风险</li>
|
||
<li>工作记录展示任务产出</li>
|
||
<li>风险看板聚合态势</li>
|
||
<li>规则优化和误报分析</li>
|
||
<li>离线回放和算法评测</li>
|
||
</ul>
|
||
</div>
|
||
</div>
|
||
|
||
<div class="card" style="margin-top: 16px;">
|
||
<h3>核心结构</h3>
|
||
<pre>RiskObservation
|
||
subject_type / subject_id
|
||
risk_type / risk_signals_json
|
||
severity / score / confidence_score
|
||
ontology_parse_id / ontology_version
|
||
domain / scenario / intent
|
||
evidence_items_json / evidence_path_json
|
||
related_policy_clauses_json
|
||
comparable_baseline_json
|
||
suggested_actions_json
|
||
source_type / source_id
|
||
automation_mode / control_stage / control_mode
|
||
algorithm_version / feedback_status</pre>
|
||
</div>
|
||
</div>
|
||
</section>
|
||
|
||
<section id="scoring">
|
||
<div class="section-inner">
|
||
<div class="section-head">
|
||
<div>
|
||
<h2>风险评分模型</h2>
|
||
<div class="section-kicker">
|
||
第一版采用可解释加权模型，不用黑盒模型直接给最终结论。复杂模型只提供候选证据，强结论必须能解释。
|
||
</div>
|
||
</div>
|
||
<span class="tag">Explainable Score</span>
|
||
</div>
|
||
|
||
<div class="formula">
|
||
<div class="formula-main">
|
||
risk_score = clip(
|
||
0.35 * S_rule
|
||
+ 0.25 * S_anomaly
|
||
+ 0.20 * S_graph
|
||
+ 0.15 * S_policy
|
||
+ 0.05 * S_history,
|
||
0, 100
|
||
)
|
||
</div>
|
||
<div class="formula-note">
|
||
<strong>S_rule</strong> 来自规则中心命中，<strong>S_anomaly</strong> 来自画像基线偏离，<strong>S_graph</strong> 来自图谱关系异常，<strong>S_policy</strong> 来自制度语义，<strong>S_history</strong> 来自历史反馈。
|
||
</div>
|
||
</div>
|
||
|
||
<div class="grid two" style="margin-top: 16px;">
|
||
<div class="formula">
|
||
<h3>同类基线偏离</h3>
|
||
<div class="formula-main">deviation = (x - median(peer)) / max(IQR(peer), epsilon)
|
||
S_anomaly = 100 * sigmoid(k * (deviation - tau))</div>
|
||
<div class="formula-note">用于金额、频次、天数、退回率等指标。样本不足时必须记录降级口径。</div>
|
||
</div>
|
||
<div class="formula">
|
||
<h3>图谱异常分</h3>
|
||
<div class="formula-main">S_graph = min(100, sum(w_i * g_i))</div>
|
||
<div class="formula-note">候选信号包括重复票据、供应商集中、审批链异常、时空冲突和相似单据异常。</div>
|
||
</div>
|
||
</div>
|
||
|
||
<div class="matrix" style="margin-top: 16px;">
|
||
<table>
|
||
<thead>
|
||
<tr>
|
||
<th>分数项</th>
|
||
<th>证据来源</th>
|
||
<th>可解释依据</th>
|
||
<th>是否可单独强拦截</th>
|
||
</tr>
|
||
</thead>
|
||
<tbody>
|
||
<tr>
|
||
<td><code>S_rule</code></td>
|
||
<td>规则中心 / 决策表</td>
|
||
<td>规则版本、命中条件、输入快照。</td>
|
||
<td>强制度规则可拦截。</td>
|
||
</tr>
|
||
<tr>
|
||
<td><code>S_anomaly</code></td>
|
||
<td>画像基线</td>
|
||
<td>同部门、同职级、同费用类型分位偏离。</td>
|
||
<td>不可单独拦截。</td>
|
||
</tr>
|
||
<tr>
|
||
<td><code>S_graph</code></td>
|
||
<td>图谱路径</td>
|
||
<td>供应商集中、重复票据、异常审批路径。</td>
|
||
<td>高证据覆盖时可复核拦截。</td>
|
||
</tr>
|
||
<tr>
|
||
<td><code>S_policy</code></td>
|
||
<td>制度条款</td>
|
||
<td>费用说明、票据、制度条款是否一致。</td>
|
||
<td>需结合规则和证据。</td>
|
||
</tr>
|
||
<tr>
|
||
<td><code>S_history</code></td>
|
||
<td>人工反馈</td>
|
||
<td>相似单据退回率、确认率、误报率。</td>
|
||
<td>不单独拦截，影响权重和门控。</td>
|
||
</tr>
|
||
</tbody>
|
||
</table>
|
||
</div>
|
||
</div>
|
||
</section>
|
||
|
||
<section id="evidence">
|
||
<div class="section-inner">
|
||
<div class="section-head">
|
||
<div>
|
||
<h2>证据链如何展示</h2>
|
||
<div class="section-kicker">
|
||
图谱不要默认展示成全量关系网。产品上要展示小范围、场景化、可解释的证据链。
|
||
</div>
|
||
</div>
|
||
<span class="tag">Evidence Path</span>
|
||
</div>
|
||
|
||
<div class="evidence-path">
|
||
<div class="evidence-item"><strong>员工</strong><span>张三，P8，研发中心</span></div>
|
||
<div class="evidence-item"><strong>单据</strong><span>差旅报销 4,860 元</span></div>
|
||
<div class="evidence-item"><strong>票据</strong><span>酒店发票，1,260 元/晚</span></div>
|
||
<div class="evidence-item"><strong>供应商</strong><span>上海某酒店</span></div>
|
||
<div class="evidence-item"><strong>基线</strong><span>同级别 P90 为 760 元/晚</span></div>
|
||
<div class="evidence-item"><strong>制度</strong><span>住宿标准第 3 条</span></div>
|
||
<div class="evidence-item"><strong>历史</strong><span>相似 12 笔，8 笔退回</span></div>
|
||
</div>
|
||
|
||
<div class="grid three" style="margin-top: 16px;">
|
||
<div class="card blue">
|
||
<h3>单据详情</h3>
|
||
<p>面向审批人，解释当前单据为什么异常、应该怎么处理。</p>
|
||
</div>
|
||
<div class="card green">
|
||
<h3>工作记录详情</h3>
|
||
<p>面向管理员，解释数字员工本次任务处理了哪些实体、发现了哪些风险。</p>
|
||
</div>
|
||
<div class="card amber">
|
||
<h3>风险看板</h3>
|
||
<p>面向财务和管理层，展示整体风险态势、算法效果和抽审结果。</p>
|
||
</div>
|
||
</div>
|
||
</div>
|
||
</section>
|
||
|
||
<section id="deep">
|
||
<div class="section-inner">
|
||
<div class="section-head">
|
||
<div>
|
||
<h2>深度算法层</h2>
|
||
<div class="section-kicker">
|
||
这些模块构成算法复杂度，但必须全部回到可解释证据和 `RiskObservation`，不能变成不可审计黑盒。
|
||
</div>
|
||
</div>
|
||
<span class="tag">Algorithm Stack</span>
|
||
</div>
|
||
|
||
<div class="grid two">
|
||
<div class="card">
|
||
<h3>对象中心过程挖掘</h3>
|
||
<p>把申请、票据、预算、审批、付款、供应商放入多对象事件日志，识别跳步审批、流程绕行、返工循环和付款前异常。</p>
|
||
</div>
|
||
<div class="card">
|
||
<h3>实体解析与主数据归一</h3>
|
||
<p>归一供应商、商户、酒店、银行户名、员工姓名，避免同一主体被拆成多个节点。</p>
|
||
</div>
|
||
<div class="card">
|
||
<h3>异构图与时序图特征</h3>
|
||
<p>计算员工、部门、供应商、票据、规则、制度之间的元路径、中心性、邻域风险密度和关系突增。</p>
|
||
</div>
|
||
<div class="card">
|
||
<h3>多模型异常检测组合</h3>
|
||
<p>组合稳健统计、孤立森林、局部离群、时间突变和周期偏离，但只作为候选风险证据。</p>
|
||
</div>
|
||
<div class="card">
|
||
<h3>决策建模与策略即代码</h3>
|
||
<p>把规则中心的确定性规则拆成决策表、策略运行和决策追踪，保证版本化和可审计。</p>
|
||
</div>
|
||
<div class="card">
|
||
<h3>反事实风险建议</h3>
|
||
<p>给出降低风险分的可执行补救动作，例如补充酒店水单、选择协议酒店或增加预算负责人复核。</p>
|
||
</div>
|
||
<div class="card">
|
||
<h3>数据血缘与质量门禁</h3>
|
||
<p>记录风险观察使用了哪些表、文档、OCR、AgentRun、规则版本和本体版本；低质量数据不触发强风控。</p>
|
||
</div>
|
||
<div class="card">
|
||
<h3>可解释与不确定性控制</h3>
|
||
<p>保存特征贡献、置信度、低置信度原因和人工复核要求，让每个结论都能被复查。</p>
|
||
</div>
|
||
</div>
|
||
</div>
|
||
</section>
|
||
|
||
<section id="gate">
|
||
<div class="section-inner">
|
||
<div class="section-head">
|
||
<div>
|
||
<h2>人机共审门控</h2>
|
||
<div class="section-kicker">
|
||
系统不应让 AI 直接通过或驳回单据。自动化必须由置信度、风险等级、证据覆盖、历史误报率和金额影响共同决定。
|
||
</div>
|
||
</div>
|
||
<span class="tag">Human in the Loop</span>
|
||
</div>
|
||
|
||
<div class="formula">
|
||
<div class="formula-main">
|
||
if confidence >= theta_auto
|
||
and severity <= medium
|
||
and evidence_coverage >= 2
|
||
and false_positive_rate <= alpha:
|
||
automation_mode = "auto"
|
||
elif confidence >= theta_semi
|
||
and evidence_coverage >= 1:
|
||
automation_mode = "semi_auto"
|
||
else:
|
||
automation_mode = "assist"
|
||
</div>
|
||
<div class="formula-note">
|
||
`assist` 只给建议，`semi_auto` 可半自动但保留抽审，`auto` 只允许低风险、高置信、证据充分、历史误报低的场景。
|
||
</div>
|
||
</div>
|
||
|
||
<div class="grid four" style="margin-top: 16px;">
|
||
<div class="card red"><h3>高风险</h3><p>全量复核，必须展示完整证据链和制度依据。</p></div>
|
||
<div class="card amber"><h3>中风险</h3><p>按风险类型、金额和历史误报率进入抽审。</p></div>
|
||
<div class="card green"><h3>低风险</h3><p>可进入小比例随机抽审，用于监控漏报。</p></div>
|
||
<div class="card blue"><h3>低置信</h3><p>只生成候选观察，不触发强拦截。</p></div>
|
||
</div>
|
||
</div>
|
||
</section>
|
||
|
||
<section id="dashboard">
|
||
<div class="section-inner">
|
||
<div class="section-head">
|
||
<div>
|
||
<h2>产品体现</h2>
|
||
<div class="section-kicker">
|
||
核心算法不能只存在后端。它必须在用户能理解的地方体现：详情解释单点风险，看板解释整体态势，工作记录解释数字员工产出。
|
||
</div>
|
||
</div>
|
||
<span class="tag">Product Surface</span>
|
||
</div>
|
||
|
||
<div class="flow">
|
||
<div class="flow-row">
|
||
<div class="flow-node"><strong>单据详情</strong><span>风险证据链、基线对比、制度条款、建议动作。</span></div>
|
||
<div class="flow-node"><strong>工作记录详情</strong><span>任务范围、处理数量、风险观察、异常关系。</span></div>
|
||
<div class="flow-node"><strong>风险看板</strong><span>总览、分布、趋势、排行、算法效果。</span></div>
|
||
<div class="flow-node"><strong>规则中心</strong><span>规则定义、版本、测试、发布、候选规则审核。</span></div>
|
||
<div class="flow-node"><strong>画像详情</strong><span>员工、部门、供应商和费用类型长期基线。</span></div>
|
||
</div>
|
||
</div>
|
||
|
||
<div class="matrix" style="margin-top: 16px;">
|
||
<table>
|
||
<thead>
|
||
<tr>
|
||
<th>页面</th>
|
||
<th>用户</th>
|
||
<th>回答的问题</th>
|
||
<th>数据来源</th>
|
||
</tr>
|
||
</thead>
|
||
<tbody>
|
||
<tr>
|
||
<td>单据详情</td>
|
||
<td>审批人</td>
|
||
<td>这张单为什么异常，应该怎么处理。</td>
|
||
<td>RiskObservation + evidence_path</td>
|
||
</tr>
|
||
<tr>
|
||
<td>数字员工工作记录</td>
|
||
<td>管理员</td>
|
||
<td>这次任务处理了什么，产出了什么风险。</td>
|
||
<td>AgentRun + RiskObservation</td>
|
||
</tr>
|
||
<tr>
|
||
<td>风险看板</td>
|
||
<td>财务 / 管理层</td>
|
||
<td>整体风险态势、确认率、误报率、抽审效率。</td>
|
||
<td>RiskObservation 聚合</td>
|
||
</tr>
|
||
<tr>
|
||
<td>规则中心</td>
|
||
<td>管理员</td>
|
||
<td>规则如何定义、测试、发布和回滚。</td>
|
||
<td>risk_rules + rule_runs</td>
|
||
</tr>
|
||
</tbody>
|
||
</table>
|
||
</div>
|
||
</div>
|
||
</section>
|
||
|
||
<section id="roadmap">
|
||
<div class="section-inner">
|
||
<div class="section-head">
|
||
<div>
|
||
<h2>实施路线</h2>
|
||
<div class="section-kicker">
|
||
第一版不要一次引入全部复杂算法。先打通最小闭环，然后逐步增强图谱、过程挖掘和评测体系。
|
||
</div>
|
||
</div>
|
||
<span class="tag">Execution Plan</span>
|
||
</div>
|
||
|
||
<div class="step-list">
|
||
<div class="step">
|
||
<div>
|
||
<strong>建立本体桥接和 RiskObservation</strong>
|
||
<p>把本体解析、规则命中、画像偏离、图谱路径和人工反馈统一进风险观察模型。</p>
|
||
</div>
|
||
</div>
|
||
<div class="step">
|
||
<div>
|
||
<strong>实现第一批 8 个高价值风险信号</strong>
|
||
<p>重复票据、金额超标准、预算超额、附件缺失、报销发票金额不一致、申请报销不一致、时间地点不一致、供应商集中异常。</p>
|
||
</div>
|
||
</div>
|
||
<div class="step">
|
||
<div>
|
||
<strong>落地三个产品入口</strong>
|
||
<p>单据详情风险证据链、数字员工工作记录详情、分析看板风险看板。</p>
|
||
</div>
|
||
</div>
|
||
<div class="step">
|
||
<div>
|
||
<strong>沉淀反馈与回放集</strong>
|
||
<p>记录人工采纳、误报、退回、补件和升级审批；建立历史样本回放能力。</p>
|
||
</div>
|
||
</div>
|
||
<div class="step">
|
||
<div>
|
||
<strong>引入深度算法模块</strong>
|
||
<p>逐步加入实体解析、对象中心过程挖掘、异构图特征、异常检测组合和反事实建议。</p>
|
||
</div>
|
||
</div>
|
||
</div>
|
||
</div>
|
||
</section>
|
||
|
||
<section id="sources">
|
||
<div class="section-inner">
|
||
<div class="section-head">
|
||
<div>
|
||
<h2>参考资料</h2>
|
||
<div class="section-kicker">
|
||
以下资料用于公开方法论和可借鉴技术来源。内部实现仍以 X-Financial 自有数据、本体、规则和反馈闭环为准。
|
||
</div>
|
||
</div>
|
||
<span class="tag">References</span>
|
||
</div>
|
||
|
||
<ul class="sources">
|
||
<li><a href="https://mks.yybip.com/group1/M00/07/EB/CgoRC2JVTMGAPdWmAEdtt5GGOf0756.pdf">用友 YonBIP 财务云智能费控服务白皮书</a></li>
|
||
<li><a href="https://mks.yybip.com/group1/M00/0A/29/CgoRC2XvFQuAKvNtACX8GJS9Zgo009.pdf">用友数智化财务资料：商旅费控、事项法会计与 AI 能力</a></li>
|
||
<li><a href="https://www.ekuaibao.com/hoseai.html">合思 AI 官网：Agent 与财务审核能力</a></li>
|
||
<li><a href="https://www.ekuaibao.com/solutionsr/control.html">合思企业内控解决方案</a></li>
|
||
<li><a href="https://www.ocel-standard.org/">OCEL 2.0 Object-Centric Event Log</a></li>
|
||
<li><a href="https://www.tf-pm.org/resources/manifesto">IEEE Task Force Process Mining Manifesto</a></li>
|
||
<li><a href="https://arxiv.org/abs/2006.10637">Temporal Graph Networks</a></li>
|
||
<li><a href="https://papers.neurips.cc/paper/6703-inductive-representation-learning-on-large-graphs">GraphSAGE</a></li>
|
||
<li><a href="https://www.kdd.org/kdd2017/papers/view/metapath2vec-scalable-representation-learning-for-heterogeneous-networks">metapath2vec</a></li>
|
||
<li><a href="https://arxiv.org/abs/1903.07293">Heterogeneous Graph Attention Network</a></li>
|
||
<li><a href="https://cs.nju.edu.cn/zhouzh/zhouzh.files/publication/icdm08b.pdf">Isolation Forest</a></li>
|
||
<li><a href="https://arxiv.org/abs/1705.07874">SHAP: A Unified Approach to Interpreting Model Predictions</a></li>
|
||
<li><a href="https://www.omg.org/dmn/">OMG Decision Model and Notation</a></li>
|
||
<li><a href="https://www.openpolicyagent.org/docs/latest">Open Policy Agent Documentation</a></li>
|
||
<li><a href="https://openlineage.io/">OpenLineage</a></li>
|
||
<li><a href="https://docs.greatexpectations.io/docs/guides/validation/validate_data_overview/">Great Expectations Data Validation</a></li>
|
||
</ul>
|
||
|
||
<div class="print-note">
|
||
本页为静态 HTML 文档，可直接浏览或打印为 PDF。
|
||
</div>
|
||
</div>
|
||
</section>
|
||
</main>
|
||
|
||
<aside class="aside">
|
||
<div class="aside-title">关键算法模块</div>
|
||
<a class="mini-link" href="#ontology"><span class="dot"></span>本体标准化</a>
|
||
<a class="mini-link" href="#observation"><span class="dot"></span>风险观察池</a>
|
||
<a class="mini-link" href="#scoring"><span class="dot"></span>风险评分</a>
|
||
<a class="mini-link" href="#deep"><span class="dot"></span>图谱特征</a>
|
||
<a class="mini-link" href="#gate"><span class="dot"></span>人机门控</a>
|
||
|
||
<div class="aside-card">
|
||
<strong>核心公式</strong>
|
||
<p style="margin-top: 8px;">风险分由规则、画像、图谱、制度、历史反馈五部分组成，所有分量必须有证据来源。</p>
|
||
</div>
|
||
|
||
<div class="aside-card">
|
||
<strong>边界原则</strong>
|
||
<p style="margin-top: 8px;">规则中心管规则，图谱引擎管风险观察，数字员工管持续运行，风险看板管整体态势。</p>
|
||
</div>
|
||
|
||
<div class="aside-card">
|
||
<strong>第一版优先级</strong>
|
||
<p style="margin-top: 8px;">先做本体桥接、RiskObservation、8 个风险信号、详情证据链、工作记录详情、风险看板和反馈回放。</p>
|
||
</div>
|
||
</aside>
|
||
</div>
|
||
</body>
|
||
</html>
|