When I joined Square, I was embedded on a product that had been in-market for a year but didn’t have dedicated analytics support.

當我加入Square時，我被嵌入了已經上市一年但沒有專門的分析支持的產品。

As you might expect, the team had a large backlog of curiosities:

如您所料，該團隊積壓了很多好奇心：

• What metrics should we be tracking?

  我們應該跟蹤哪些指標？

• Why is this metric lower than I’d expect?

  為什么這個指標低于我的預期？

• Do we have data on XYZ?

  我們有關于XYZ的數據嗎？

• I have this idea for an A/B test! Can you help me implement it?

  我對A / B測試有這個想法！ 你能幫我實現嗎？

• What % of users use XYZ feature?

  百分之幾的用戶使用XYZ功能？

• Can we use machine learning? I heard clustering is a great way to understand our user base!

  我們可以使用機器學習嗎？ 我聽說集群是了解我們的用戶群的好方法！

Candidly, I was overwhelmed. Where should I start? What’s the most important thing to do right now? What should I be doing in three months, a year?

坦率地說，我不知所措。 我應該從哪里開始？ 現在最重要的事情是什么？ 一年三個月后我應該做什么？

Yearning for a framework to guide me, I came upon the AI hierarchy of needs, modeled after Maslow’s hierarchy of needs. While the specifics didn’t directly map to product analytics, I realized the overarching of a hierarchy did.

渴望有一個指導我的框架，我以馬斯洛的需求層次為模型，建立了AI需求 層次 。 盡管具體細節并沒有直接映射到產品分析，但我意識到層次結構的概述確實可以實現。

需求的分析層次結構 (The Analytics Hierarchy of Needs)

The general idea of the analytics hierarchy of needs is that you should not move up the hierarchy until you’ve done the basics in the prior step (i.e. no deep analysis before metrics are defined & tracked, no dashboards built before you’ve started collecting & cleaning your data, etc).

需求分析層次結構的總體思想是，在上一步完成基礎知識之前，您不應該上移層次結構(即，在定義和跟蹤指標之前不進行深入分析，在開始收集之前不構建儀表板)并清理您的數據等)。

1.收集 (1. Collect)

Wait, how are we not tracking usage of our flagship feature?!

等等，我們如何不跟蹤旗艦功能的使用？

This isn’t fun to hear. You can’t manufacture data that you wish existed in hindsight!

聽到這不好玩。 您無法制造出事后希望存在的數據！

Before doing anything else, an analyst’s first priority should be to ensure basic event logging & data modeling for key entities. This can be informed by your own product intuition and key questions stakeholders have of your data.

在執行其他任何操作之前，分析師的首要任務應該是確保關鍵實體的基本事件記錄和數據建模。 這可以通過您自己的產品直覺和利益相關者對您的數據的關鍵問題來了解。

Determining what to log and how to log it can be intimidating. Consider mapping out an Entity Relationship Diagram (ERD) for your key entities, and giving engineering the basic event logging guidance of LATAM (Log All Taps And [Relevant] Metadata).

確定要記錄什么以及如何記錄它可能會令人生畏。 考慮映射出一個實體關系圖(ERD)為您的實體按鍵，并給予工程LATAM的基本事件日志記錄的指導(L OG 者A 牛逼 APS 釹 [相關] 中號 etadata)。

2.清潔 (2. Clean)

Hey Ryan, can you quickly pull this for me?

嘿，瑞安，你能幫我拉一下嗎？

Sure, let me just SELECT * FROM some_pristine_table_that_does_not_exist;

當然，讓我選擇SELECT * FROM some_pristine_table_that_does_not_exist;

Your engineering team will likely emit data in disparate, narrow tables. Before dashboarding & diving into analysis, you should ETL your data into wider, standardized tables. This will make your analyses and dashboards far more scalable & easier to build.

您的工程團隊可能會在不同的狹窄表格中發出數據。 在進行儀表板分析和深入分析之前，您應該將數據ETL到更大的標準化表中。 這將使您的分析和儀表板更具可擴展性，并且更易于構建。

Consider a per-user-per-day summary table as a place to start. Your stakeholders can inform the columns: ask them what common questions they wish they knew but are unable to answer.

考慮將“每用戶每天”摘要表作為開始的地方。 您的利益相關者可以告知各列：詢問他們希望他們知道但無法回答的常見問題。

3.定義和跟蹤 (3. Define & Track)

Why is revenue lower than I expected this month?

為什么收入低于我本月的預期？

Now that your data is organized, it’s time to break down your funnel.

現在，您的數據已經組織好了，是時候分解渠道了。

Usually people care most about active users and $$. In order to understand what’s driving these, you need to map your product funnel to a flow chart. Consider this basic SaaS product as an example:

通常，人們最關心活躍用戶和$$。 為了了解驅動這些因素的原因，您需要將產品渠道映射到流程圖。 以這個基本的SaaS產品為例：

> Revenue = (active users) * (average subscription cost per user)

>收入=(活躍用戶)*(每位用戶的平均訂閱費用)

>> Active users = (new conversions) + (retained users)

>>有效用戶=(新轉化)+(保留用戶)

>>> New conversions = (free trials) * (average conversion %)

>>>新轉化=(免費試用)*(平均轉化百分比)

And so on. Once you’ve mapped out this flow chart and built dashboards tracking each metric, you’re ready to analyze your funnel.

等等。 繪制完該流程圖并構建了跟蹤每個指標的儀表板之后，就可以分析渠道了。

4.分析 (4. Analyze)

Interesting churn analysis, Ryan, but our retention is best in class. Conversion is top of mind for me right now.

有趣的客戶流失分析，Ryan，但我們的保留率是同類中最好的。 現在，轉換是我的首要任務。

Once my team had each of our funnel metrics defined and tracked, we assigned a Red/Amber/Green status to each. We did this by comparing each metric to internal benchmarks, external benchmarks, and our gut; this helped us determine our biggest areas of opportunity.

一旦我的團隊定義并跟蹤了每個渠道指標，我們便為每個渠道分配了紅色/琥珀色/綠色狀態 。 為此，我們將每個指標與內部基準，外部基準和我們的直覺進行了比較； 這有助于我們確定最大的機會領域。

For example: why do a deep dive into retention if it’s beating benchmarks?

例如：如果要超越基準率，為什么還要深入研究保留率？

Once you have your top area(s) of opportunity, you can do some basic heuristic analysis on each. There’s often a noteworthy trend when visualizing a metric by cohort, acquisition channel, and basic user demographics.

一旦有了最重要的機會領域，就可以對每個機會進行一些基本的啟發式分析。 通過同類群組，獲取渠道和基本用戶人口統計數據可視化指標時，通常會有一個值得注意的趨勢。

5.優化和預測 (5. Optimize & Predict)

So, you’re finally going to do that clustering analysis I heard about?

那么，您最終將要進行我聽說過的聚類分析？

You’re collecting data and you’ve cleaned it. You’ve defined your metrics and you’re tracking them. You’ve analyzed your metrics and outlined key areas for improvement.

您正在收集數據，并且已清理數據。 您已經定義了指標并正在跟蹤它們。 您已經分析了指標并概述了需要改進的關鍵領域。

Congratulations: you can now do the ‘sexy’ stuff. 😎

恭喜：您現在可以做“性感”的事情。 😎

A/B testing can give you confidence that one user experience is more optimal than another. If you’ve been following the analytics hierarchy of needs, this is well-timed: as the more users you have, the higher statistical confidence your experiments will yield.

A / B測試可以使您確信一種用戶體驗比另一種用戶體驗更為理想。 如果您一直遵循需求的分析層次結構，那么時機就很好：隨著用戶的增多，實驗將產生更高的統計信心。

Supervised machine learning can help you understand what is predictive of user behavior — for example, what actions or demographics lead to conversion or sustained usage.

有監督的機器學習可以幫助您了解什么可以預測用戶的行為，例如，哪些操作或人口統計信息會導致轉化或持續使用。

Unsupervised machine learning can help you understand patterns among your users you weren’t aware of (hello, clustering!).

無監督機器學習可以幫助您了解您不了解的用戶中的模式( 您好，集群！ )。

However, a word of caution: just because you’ve reached the top of the hierarchy does not mean that machine learning is the best solution to all your questions. Simple heuristics can often yield a better outcome, as they’re quicker to implement and usually more digestible.

但是，請注意：僅僅因為您已達到層次結構的最高層， 并不意味著機器學習是所有問題的最佳解決方案。 簡單的啟發式方法通常可以產生更好的結果，因為它們實施起來更快，并且通常更易于消化。

綜上所述 (In summary)

Don’t be the analyst explaining the SHAP values of your XGBoost model while your new flagship feature isn’t being tracked.

當您的新旗艦功能沒有被追蹤時，不要當分析師來解釋您的XGBoost模型的SHAP值。