Thanks. Now I will talk about our model plan and the Decision Tree.
谢谢。接下来我讲一下 model plan 和决策树。
From the EDA, we made some improvements. First, we picked only the important features. We kept strong ones like work interference and family history, and dropped the weak ones.
Second, we changed the text answers into numbers so the models can use them. Third, we adjusted the settings for each model.
第二，我们把文字答案改成了数字，让模型能用。第三，我们对每个模型都调了一下设置。
All five models were tested in the same way, so the comparison is fair.
五个模型用同样的方法测试，所以比较是公平的。
🔁 过渡句 · Cindy 建议
讲完 slide 13 后说一句话，把话题交给介绍模型的组员
Next, my team members will each introduce one model.
[A] will go first with [模型 1],
then [B], then [C].
I will introduce the Decision Tree,
and Cindy will finish with the boosting model.
💡 黄色框 = 待填空。先和组员对一下：①谁是 A/B/C？②每人负责哪个模型？③你 DT 上场的顺序？
📝 如果顺序不一样，把句子里 "first / then / I will introduce" 的位置换一下就行。
② Model Fitting引出 DT、用 plain English 解释为什么选它
14
Models Used — Decision Tree 这一行
DT 的"招牌"：rule-based classification
≈ 20s
PPT 内容（只讲 DT 那一行）
Decision Tree — Rule-based classification
（其他四行不讲）
我要说的话
这一页对应 rubric 的 Model Fitting——"appropriate models clearly explained in simple terms"。
My model is the Decision Tree. We chose it because it gives us clear rules — you can read it from top to bottom like a flow chart, so users can see how a decision is made.
Interestingly, our tree's top splits are work interference and family history — the same two features the EDA said are the strongest. So the Decision Tree agrees with the EDA: work matters more than age or gender.
③ Model Improvement — DT 的调参把 Improvement 落到 DT 层面 · ⚠️ 强调 "impact shown"
15
Cross Validation and Tuning
10-fold CV + cp / maxdepth + 调参的 impact
≈ 15s
PPT 内容（只讲 DT 那一行）
Decision Tree: cp, maxdepth
（其他 3 行 + 左边 Validation Strategy 由别人讲）
我要说的话（精简版）
这一页对应 Model Improvement——只讲 DT 这一行，但要带上"调参 → 效果更好"，满足 rubric 的"impact shown"。
For the Decision Tree, we tuned cp and maxdepth using 10-fold cross-validationvalid（瓦利得）+ DAY（白天）+ shun → "valid-DAY-审", and picked the values with the best AUC.
对决策树，我们用 10 折交叉验证调了 cp 和 maxdepth，挑了 AUC 最好的组合。
After tuning, the tree works much better on new data than without.
The key point: our Decision Tree has the highest Precision out of all five models. When our tree says "this person needs treatmenttree（树）+ t-ment → "推特-门特"", it is usually right.
重点：我们的决策树在五个模型里 Precision 最高。当树说"这个人需要治疗"，它通常是对的。
17
ROC Curve Comparison — DT 那条线
用 ROC 收尾 + 顺带一句 Model Selection 的定位
≈ 12s
PPT 内容（只看 DT 那条曲线）
橘色那条 = Decision Tree
AUC = 0.859
简短指一下就行，别人讲他们的线
我要说的话（精简版）
这一页只指 DT 的线 + 收尾。其他 4 条曲线由别人讲。
On the ROC curve, the Decision Tree is the orange line. It goes high up, close to the top-left, with AUC = 0.859 — a bit lower than the other tree models, but still a good result.