What is path analysis in product analytics?产品分析中的路径分析是什么？
Path analysis is an exploratory method that orders pages, screens, or events by user and time, then summarizes the routes people take before or after a focal action. Unlike a funnel, it does not begin with one required sequence. It exposes common transitions, branches, loops, repeated actions, and exits so a team can form better hypotheses about navigation and friction.
路径分析是一种探索型方法：它按用户和时间对页面、屏幕或事件排序，再汇总人们在某个关键动作之前或之后走过的路线。它不像漏斗那样从唯一必经序列出发，而是呈现高频转移、分支、回环、重复动作和退出，帮助团队围绕导航与摩擦形成更可靠的假设。
This guide uses the product-analytics meaning of path analysis, also called user path analysis, path exploration, click path analysis, or navigation path analysis. Statistical path analysis is different: it estimates relationships among variables within a specified model and belongs to the structural-equation-modeling family. If your question concerns direct and indirect effects among variables rather than ordered user events, use statistical guidance instead.
本文讨论的是产品分析语境中的 path analysis，也常称用户路径分析、路径探索、点击路径分析或导航路径分析。统计学路径分析则完全不同：它在预先指定的模型中估计变量关系，属于结构方程模型方法。如果问题是变量间直接与间接效应，而不是按时间排序的用户事件，应采用统计学方法。
When user path analysis helps—and when it does not用户路径分析何时有用，何时不适用
Ask what users do after signup, before cancellation, around an error, or between two milestones when the sequence is not yet known.
当序列尚不明确时，观察用户注册后、取消前、错误前后或两个里程碑之间实际做了什么。
Identify repeated searches, back-and-forth navigation, dead ends, premature exits, and unexpected shortcuts that aggregate metrics hide.
识别重复搜索、来回跳转、死胡同、过早退出和汇总指标会掩盖的意外捷径。
A route associated with conversion may reflect stronger intent, eligibility, acquisition source, or tracking coverage rather than an effect of the route.
与转化相关的路线可能反映更强意图、资格条件、获客来源或追踪覆盖，而非路线本身造成结果。
If the required sequence is already defined and the task is to quantify completion and drop-off, a funnel is usually clearer and easier to audit.
如果必经序列已经定义，任务只是量化完成率与流失，漏斗通常更清晰，也更容易审计。
Path reports are most useful when tied to a decision: simplifying onboarding, investigating a support spike, locating where successful users discover a feature, or learning what happens after a failed payment. “Show every path” is not a decision. It usually produces a dense diagram with no agreed population, time boundary, or action threshold.
路径报告在绑定具体决策时最有价值，例如简化新手引导、调查支持请求上升、定位成功用户发现功能的位置，或理解支付失败后的行为。“展示所有路径”不是决策问题，通常只会产生没有统一人群、时间边界与行动阈值的密集图。
Prepare event data before mapping user paths绘制用户路径前先准备事件数据
A path chart is only as reliable as its ordering, identity, and event definitions. Keep raw records immutable, document transformations, and create a compact analytical event table. The minimum fields are a pseudonymous subject or session key, an event timestamp with timezone, and a stable event or page name. Useful optional fields include event version, session ID, platform, acquisition source, device, plan, account, experiment assignment, and an outcome flag.
路径图的可靠性取决于排序、身份与事件定义。应保留不可变原始记录、记录转换过程，并建立紧凑的分析事件表。最少字段包括假名化主体或会话键、带时区的事件时间戳，以及稳定的事件或页面名称。可选字段包括事件版本、会话 ID、平台、获客来源、设备、套餐、账户、实验分组与结果标记。
| Input输入 | Decision to document需要记录的决策 | Failure if ignored忽略后的问题 |
|---|---|---|
| Identity身份 | User, account, device, or session; merge and deletion rules用户、账户、设备或会话；合并与删除规则 | Different people merge or one person fragments不同人被合并，或同一人被拆分 |
| Event time事件时间 | Client vs server time, timezone, ties, late arrivals客户端或服务器时间、时区、同一时间与迟到事件 | The displayed order is not the experienced order展示顺序与真实体验顺序不一致 |
| Event taxonomy事件分类 | Stable names, versions, properties, grouped noise events稳定名称、版本、属性与噪声事件分组 | Renames look like behavior change; paths explode事件改名看似行为变化，路径数量爆炸 |
| Session rule会话规则 | Inactivity gap, hard boundary, cross-device policy不活跃间隔、硬边界与跨设备策略 | Separate visits join or one journey splits独立访问被拼接，或同一旅程被拆分 |
| Population and window人群与时间窗 | Eligibility, focal event, lookback/lookforward, exclusions资格、锚点事件、回看与前看窗口、排除条件 | Survivorship, maturity, and denominator bias幸存者、成熟度与分母偏差 |
Privacy is part of the design, not a cleanup step. Collect only fields needed for the question, avoid placing direct identifiers or sensitive free text in event properties, restrict access, respect consent and deletion requests, and define retention. A pseudonymous key reduces exposure but does not make behavioral data anonymous.
隐私应在设计阶段处理，而不是最后补救。只收集问题所需字段，避免把直接身份信息或敏感自由文本写入事件属性，限制访问权限，遵守同意与删除请求，并明确保留期。假名化键能降低暴露风险，但不会让行为数据自动变成匿名数据。
Define focal event, direction, depth, and counting rules定义锚点事件、方向、深度与计数规则
Start with one focal event that matters to the decision: first value, upgrade, failed search, help request, cancellation, or purchase. A forward path asks what happens next; a reverse path asks what tends to happen before the focal event. An unrestricted start is useful for broad discovery but harder to interpret. Set the maximum steps and time window before viewing results so interesting-looking branches do not change the rules after the fact.
先选择一个与决策相关的锚点事件，例如首次获得价值、升级、搜索失败、请求帮助、取消或购买。前向路径观察接下来发生什么；反向路径观察锚点之前通常发生什么。无限制起点适合广泛探索，但解释更困难。应在查看结果前确定最大步骤数与时间窗，避免看到有趣分支后再改变规则。
- Unit: count unique users when reach matters, sessions when visit structure matters, or events when interaction volume matters. Label the denominator.单位：关注触达时统计去重用户，关注访问结构时统计会话，关注交互量时统计事件，并明确标注分母。
- Repeated events: decide whether consecutive repeats remain separate, collapse into one node, or become a repetition count. Each choice answers a different question.重复事件：决定连续重复是保留为独立节点、合并为一个节点，还是转换为重复次数。不同选择回答不同问题。
- Other and exit: retain an explicit Other bucket and define Exit. Hiding the long tail can make visible routes appear more dominant than they are.其他与退出：保留明确的“其他”桶并定义“退出”。隐藏长尾会让可见路线看起来比实际更占主导。
- Segmentation: compare only a few decision-relevant groups while preserving sample counts and missingness.分群：只比较少数与决策有关的群体，并保留样本量与缺失情况。
How to run path analysis step by step如何逐步执行路径分析
- Write the decision and observable question写下决策与可观察问题Replace “understand navigation” with a bounded question such as “What do first-week users do in the 20 minutes after their first failed search, and which routes end in a successful content view?” Name the owner and possible actions.把“理解导航”改成有边界的问题，例如：“首周用户首次搜索失败后 20 分钟内做了什么，哪些路线最终到达成功内容浏览？”同时明确负责人和可能行动。
- Freeze eligibility and time boundaries冻结资格与时间边界Define who can enter, the date range, maturity delay, focal event, lookback or lookforward window, session gap, and exclusions for employees, bots, test accounts, duplicates, or incomplete instrumentation.定义进入资格、日期范围、成熟期、锚点事件、前后观察窗、会话间隔，以及员工、机器人、测试账户、重复记录或追踪不完整数据的排除规则。
- Build and quality-check the event sequence构建并检查事件序列Normalize names and timestamps, resolve ties deterministically, apply identity and session rules, compare source totals with the analytical table, and manually trace representative records.规范名称与时间戳，用确定性规则解决同一时间事件，应用身份和会话规则，将来源总数与分析表核对，并人工追踪代表性记录。
- Choose useful nodes and limit complexity选择有用节点并限制复杂度Keep decision-relevant actions, outcomes, errors, and meaningful page groups. Group low-level telemetry, cap depth, collapse repeats when justified, keep Other visible, and record transformations.保留相关动作、结果、错误和页面组。合并低层遥测、限制深度、在合理时折叠重复、保留“其他”，并记录转换。
- Calculate transitions and coverage计算转移与覆盖率Report eligible counts, next-node counts, transition share, exit share, and cumulative coverage. When users can re-enter, state whether each occurrence or only one occurrence per user is counted.报告合格数量、下一节点数量、转移占比、退出占比和累计覆盖率。若用户可重复进入，应说明统计每次出现还是每位用户只统计一次。
- Compare segments and inspect the long tail比较分群并检查长尾Check whether a branch persists across dates, devices, sources, accounts, and releases. Inspect Other and missing events before declaring a path unusual.检查分支在日期、设备、来源、账户与版本间是否持续，并在判断异常前检查“其他”与缺失事件。
- Turn patterns into hypotheses and tests把模式转化为假设与验证Pair the path with support themes, usability research, or session evidence. Define an intervention, expected metric, guardrails, duration, and rollback. Use an experiment or credible causal design before claiming impact.把路径与支持主题、可用性研究或会话证据结合。定义干预、预期指标、护栏、持续时间与回滚，并在声称影响前采用实验或可信因果设计。
Path analysis vs funnel, journey, and statistical methods路径分析与漏斗、旅程及统计方法的区别
| Method方法 | Best question最适问题 | Structure结构 | Main caution主要注意事项 |
|---|---|---|---|
| User path analysis用户路径分析 | What sequences occur before or after an event?某事件前后出现哪些序列？ | Exploratory, many branches探索型，多分支 | Complexity and observational bias复杂度与观察偏差 |
| Funnel analysis漏斗分析 | Where does a defined sequence lose eligible users?既定序列在哪一步流失用户？ | Predefined ordered steps预定义有序步骤 | Optional behavior may look like failure可选行为可能被误判为失败 |
| Customer journey analytics客户旅程分析 | How do cross-channel interactions relate to outcomes?跨渠道互动如何关联结果？ | Longer, cross-source journey更长、跨来源旅程 | Identity stitching and unavailable touchpoints身份拼接与不可用触点 |
| Statistical path analysis统计学路径分析 | Are variable relationships consistent with a specified model?变量关系是否与指定模型一致？ | Regression-based path model基于回归的路径模型 | Assumptions and causal overclaiming假设与因果过度声称 |
A productive sequence is often path → funnel → cohort or experiment. Explore paths to find an unexpected branch; convert it into a precise funnel or metric; check whether the pattern is concentrated in a cohort or release; then test a change. Do not keep adding nodes when a narrower method answers the next question more clearly.
有效工作流常是“路径 → 漏斗 → cohort 或实验”。先用路径发现意外分支，再把它转换为精确漏斗或指标，检查模式是否集中在某个 cohort 或版本，最后测试改动。若更窄的方法能更清楚回答下一问题，就不要继续堆节点。
Example: investigating a failed-search user path示例：调查搜索失败后的用户路径
Hypothetical example: the events and numbers below illustrate reasoning only; they are not InfiniSynapse customer data or a benchmark.
假设示例：以下事件与数字只用于展示推理过程，不是 InfiniSynapse 客户数据，也不是行业基准。
A documentation product sees more support tickets mentioning search. The team defines eligible subjects as signed-in first-week users who trigger search_no_results. It examines the next five meaningful events within 20 minutes, counts one focal occurrence per user per day, collapses consecutive query edits, excludes internal accounts, and defines content_view_success as the outcome.
某文档产品发现提到搜索的支持请求增多。团队把合格主体定义为登录后的首周用户，且触发 search_no_results。分析观察随后 20 分钟内最多五个有意义事件，每位用户每天只统计一次锚点事件，折叠连续查询修改，排除内部账户，并把 content_view_success 定义为结果。
Suppose the illustrative summary shows 46% edit the query and view content; 24% open category navigation before viewing content; 18% repeat search, open help, and exit; 12% enter Other. These percentages do not show that navigation or help caused success or failure. They show where to inspect evidence. Session samples reveal a recent terminology change is absent from search synonyms, so the team proposes updated synonyms and clearer no-result suggestions.
假设汇总显示：46% 修改查询后成功浏览内容；24% 先打开分类导航再浏览；18% 重复搜索、打开帮助后退出；12% 进入“其他”。这些比例不能证明导航或帮助造成成功或失败，只能指明证据检查位置。会话抽样发现近期术语变更没有进入搜索同义词，于是团队提出更新同义词和改进无结果建议。
Before shipping, the team verifies the pattern on web and mobile, checks alternate 15- and 30-minute windows, reconciles source counts, traces raw sequences, and defines guardrails for relevance, latency, and support contacts. A controlled rollout tests whether the change improves content success without increasing irrelevant results.
发布前，团队验证该模式在网页和移动端都存在，检查 15 与 30 分钟窗口，核对来源计数、追踪原始序列，并为相关性、延迟与支持联系定义护栏。最后通过受控发布测试改动能否改善内容成功且不增加无关结果。
Common path analysis mistakes and limits路径分析的常见错误与局限
Many branches without cumulative coverage hide grouped or excluded behavior. Report displayed coverage and Other.
大量分支却不报告累计覆盖率，会掩盖被合并或排除的行为。应报告展示覆盖率和“其他”。
Users, sessions, accounts, and events produce different shares. Choose one primary grain and label secondary metrics.
用户、会话、账户和事件会产生不同占比。选择一个主要粒度并标注次要指标。
Client clocks, batch ingestion, retries, and ties can reorder actions. Preserve source time, receipt time, and tie rules.
客户端时钟、批量摄取、重试和同序会改变动作顺序。应保留来源时间、接收时间与同序规则。
Starting only from converters removes failed paths. Define eligibility before the outcome when comparing success.
只从转化者开始会排除失败路径。比较成功时，应在结果发生前定义资格。
Renamed events or newly instrumented screens can look like behavior change. Version schemas and annotate releases.
事件改名或新追踪屏幕可能看似行为变化。应版本化 schema 并标注发布。
A prominent route is not necessarily desirable. Link interpretation to outcomes, alternatives, and tests.
高频路线不一定理想。应把解释与结果、替代解释及测试相连。
Cross-device journeys may fragment without a lawful shared identity. Offline or blocked events may be absent. Sampling can suppress rare but important routes. Long windows mix separate goals; short windows cut valid journeys. Visualization thresholds can change apparent dominance. Document these boundaries beside the result.
没有合法共享身份时，跨设备旅程会被拆分；线下或被拦截事件可能缺失；抽样会压制罕见但重要的路线；窗口过长会混合不同目标，过短会切断有效旅程；阈值也会改变表面主导路线。应把这些边界写在结果旁。
Validate a path report before making a product decision做产品决策前验证路径报告
- Reconcile counts: compare source events, cleaned events, eligible users, focal occurrences, and displayed steps for the same window.核对计数：在同一窗口比较来源事件、清洗事件、合格用户、锚点次数与展示步骤。
- Trace records: manually reconstruct common, looped, exited, and Other sequences from raw data.追踪记录：从原始数据人工重建高频、回环、退出与“其他”序列。
- Run sensitivity checks: vary session gaps, windows, repeat collapsing, grouping, and identity rules. Separate stable findings from definition-dependent ones.执行敏感性检查：改变会话间隔、窗口、重复折叠、分组与身份规则，区分稳定发现与依赖定义的发现。
- Compare populations: review device, version, market, source, tenure, and account segments with sample sizes and missingness.比较人群：按设备、版本、市场、来源、使用时长与账户检查，并报告样本量与缺失情况。
- Triangulate: pair aggregate routes with support evidence, usability research, or privacy-respecting session evidence.三角验证：把汇总路线与支持证据、可用性研究或尊重隐私的会话证据结合。
- Test action: predefine the intervention, target metric, guardrails, decision threshold, duration, and rollback.测试行动：预先定义干预、目标指标、护栏、决策阈值、持续时间与回滚。
A trustworthy result states who was eligible, how events were ordered and grouped, what share is visible, what is missing, whether the pattern survives reasonable definitions, and which evidence would change the conclusion. Export both the visualization and a tabular transition summary so another analyst can reproduce the counts.
可信结果应说明谁符合资格、事件如何排序与分组、可见占比、缺失内容、模式是否经得起合理定义变化，以及什么证据会改变结论。应同时导出可视化和表格化转移汇总，让其他分析师能够复现计数。
Use InfiniSynapse for an evidence-first path analysis workflow使用 InfiniSynapse 构建证据优先的路径分析流程
Before opening the tool, prepare a database table or file containing the pseudonymous subject key, timestamp, event name, optional session key, relevant segment fields, and a written definition of the focal event and window. Remove unnecessary personal data and keep a source-total query for reconciliation.
打开工具前，准备包含假名化主体键、时间戳、事件名称、可选会话键与相关分群字段的数据库表或文件，并写明锚点事件和窗口定义。删除不必要的个人数据，同时保留用于总数核对的来源查询。
Use InfiniSynapse Online to connect or provide supported data sources, ask bounded analytical questions in natural language, inspect generated evidence, and iterate on definitions. Verify results against source totals; InfiniSynapse is not described here as an event collector, session-replay recorder, dedicated path-chart builder, or experimentation platform.
通过 InfiniSynapse Online 连接或提供受支持的数据源，以自然语言提出有边界的分析问题、检查生成证据并迭代定义。所有结果都应与来源总数核对；本文不会把 InfiniSynapse 描述成事件采集器、会话回放录制器、专用路径图构建器或实验平台。
Open InfiniSynapse Online打开 InfiniSynapse OnlineFor product routes and installable command tools, review the InfiniSynapse tool directory. For broader product context, visit the InfiniSynapse feature overview. This article is a local delivery artifact; its canonical and bilingual routes require deployment before they become live pages.
如需查看产品入口和可安装命令工具，请浏览 InfiniSynapse 工具目录；如需了解更广泛的产品背景，请访问 InfiniSynapse 功能概览。本文目前是本地交付物，其 canonical 与双语地址需要部署后才会成为在线页面。
Path analysis FAQ路径分析常见问题
What is path analysis in product analytics?
产品分析中的路径分析是什么？
Path analysis orders page, screen, or event records by user and time, then summarizes the routes people take before or after a focal event. It reveals frequent sequences, branches, loops, and exits without requiring one predefined funnel.
路径分析按用户和时间排列页面、屏幕或事件记录，再汇总人们在锚点事件前后走过的路线。它无需预设唯一漏斗，就能呈现高频序列、分支、回环与退出。
What data do I need for user path analysis?
用户路径分析需要哪些数据？
At minimum, use a pseudonymous user or session key, an event timestamp, and a stable event or page name. Add event properties, user segments, session rules, and outcome flags only when they support the decision.
至少需要假名化用户或会话键、事件时间戳，以及稳定的事件或页面名称。仅在支持决策时添加事件属性、用户分群、会话规则与结果标记。
How is path analysis different from funnel analysis?
路径分析与漏斗分析有什么区别？
Path analysis discovers many observed sequences around a focal point. Funnel analysis measures progress through a defined sequence and conversion window. Use paths to discover behavior and funnels to quantify a specific hypothesis.
路径分析发现锚点周围的多种真实序列；漏斗分析度量用户通过既定序列与转化窗口的进度。用路径发现行为，用漏斗量化具体假设。
Can a common user path prove what caused conversion?
高频用户路径能证明转化原因吗？
No. A frequent path is an observational association shaped by intent, eligibility, tracking, and product design. Use qualitative evidence and a controlled experiment or credible causal design before claiming an event caused an outcome.
不能。高频路径是受意图、资格、追踪和产品设计影响的观察关联。在声称某事件造成结果前，应结合定性证据，并采用受控实验或可信因果设计。
Why does a path diagram become unreadable?
为什么路径图会变得难以阅读？
Too many low-level events, unlimited depth, repeated events, mixed populations, and inconsistent sessions create a spaghetti diagram. Start from a decision, group equivalent events, cap depth, expose an Other bucket, and segment deliberately.
低层事件过多、深度无限、重复事件、混合人群和不一致会话会形成“意大利面图”。应从决策出发，合并等价事件、限制深度、显示“其他”桶，并谨慎分群。
Official sources and next steps官方来源与下一步
For a first-party implementation reference, read the AWS clickstream path-analysis guidance, which documents metrics, session definitions, nodes, filters, repeated-event merging, and an example. The Google Analytics Path exploration documentation explains forward and reverse exploration, nodes, starting points, ending points, and product limits. These sources describe their own products; they do not prove that one configuration fits every organization.
如需第一方实现参考，可阅读 AWS 点击流路径分析指南，其中记录了指标、会话定义、节点、筛选、重复事件合并与示例；Google Analytics 路径探索文档说明前向与反向探索、节点、起点、终点及产品限制。这些来源描述各自产品，并不能证明某一种配置适合所有组织。
The next useful step is not a larger diagram. It is a smaller, auditable question: select one focal event, document one population and window, reconcile one event table, inspect the dominant routes and long tail, and write one hypothesis with a validation plan.
下一步最有价值的工作不是绘制更大的图，而是提出更小、可审计的问题：选择一个锚点事件，记录一个人群与窗口，核对一张事件表，检查主导路线与长尾，并写下一个带验证计划的假设。