Suzy (Ziwei) Su苏紫薇 Suzy Su

My work我的工作

JAFFLE_SHOP / FROM SOURCE TO REPORTJAFFLE_SHOP / 从数据源到报表

I turned the three marts from the dbt project into a dashboard: a cohort-retention heatmap, monthly revenue and customer segments. Published on Tableau Public, and rebuilt in Power BI Desktop with a relationship between orders and customers, three DAX measures (revenue, orders, repeat rate) and a heatmap made from a matrix with conditional formatting.

我把 dbt 项目的三张 mart 表做成看板：cohort 留存热力图、月度收入、客户分层。发布在 Tableau Public 上；又在 Power BI Desktop 里重做一遍，建了订单表和客户表的关联、三个 DAX 度量值（收入、订单数、复购率），热力图用矩阵加条件格式做出。

Public training data · 99 orders · Power BI Desktop, not a production BI deployment.公开练习数据 · 99 笔订单 · Power BI Desktop，非生产 BI 部署。

My work我的工作

I modelled public jaffle_shop sample data in three layers so each question (revenue by order, customer value, monthly retention) is answered by one tested mart. Order sequence and cohort month come from window functions. 56 data tests pass, including three reconciliation checks: revenue in the marts equals raw payments, every order is counted once, and month-zero retention is exactly 100%. Generated docs and a CI build that fails on any test failure.

我用 dbt 官方公开样例数据 jaffle_shop 分三层建模，让「每单收入、客户价值、月度留存」三个问题各由一张有测试的 mart 表回答。订单序号和 cohort 月份用窗口函数算出。56 个数据测试全部通过，其中三条是对账测试：mart 里的总收入等于原始支付总额、每单只计一次、首月留存恰好是 100%。附自动生成的文档，CI 里任何测试失败都会让构建失败。

My work我的工作

I compared validation strategies for rock-type classification from well logs on a public dataset (3,164 rows, 7 wells, 9 classes). Scored twice: shuffled K-fold, and holding out whole wells, which is how the model would really be used. Random K-fold made LightGBM look like a 0.71 macro-F1 model, clearly ahead of logistic regression; held out by well it scores 0.38 and ties it. The leak inflated the score and would have picked the wrong model. 26 tests.

我用公开测井数据（3,164 行、7 口井、9 类岩相）做岩相分类，评估两次：随机打乱的 K 折，以及整口井留出（这才是模型真实的用法）。随机 K 折下 LightGBM 的 macro-F1 是 0.71，明显好过逻辑回归；按井留出后只有 0.38，和逻辑回归打平。数据泄漏抬高了分数，还会让人选错模型。26 个测试。

My work我的工作

I compared three pity rules with 20,000 seeded trials each, measuring acquisition cost and total issuance separately, and stated what the experiment cannot tell you.

我用三种保底规则各跑 2 万次固定种子模拟，把获取成本和总投放分开测量，并写明这个实验回答不了什么。

My work我的工作

I analysed audience attention versus shot content in films, grounding the analysis in real ClickHouse queries. The obvious story did not survive: attention rises toward the end of almost every film regardless of content, so raw correlations are confounded. Across three films and 18 tests against Bonferroni-corrected thresholds, one result cleared the floor by a hair and did not replicate. The write-up also publishes my own errors: end credits had been scored as film, which produced the original 4.40× headline (1.28× on the film itself).

我分析电影中观众注意力和镜头内容的关系，通过真实的 ClickHouse 查询为回答提供依据。最直观的结论没有站住：几乎每部片子后半段的注意力都会上升，和内容无关，所以直接做相关会被这个趋势干扰。三部片、18 项检验按 Bonferroni 校正后的阈值判断，只有一项勉强过线，而且没有复现。文档也公开了我自己的错误：片尾字幕曾被当作正片打分，最初“后半段是前半段 4.40 倍”的结论由此而来（正片本身是 1.28 倍）。