数据偏移与概念偏移
区分输入变化和目标变化
数据偏移与概念偏移 是 CoddyKit 上的免费 MLOps Academy 课时。 这是第 1 节课，共 4 节。 你可以在下方免费阅读本课时的完整内容 — 然后在浏览器中使用内置代码编辑器和全天候 AI 导师进行实践。 这是 MLOps Academy 学习路径的一部分，你的进度在网页和 CoddyKit 应用中同步。 MLOps Academy 课程共包含 4 节课。
本课时的部分内容尚未翻译，以英文显示。
Two Ways the World Shifts
Your model was trained on a snapshot of the past. When reality moves on, it can quietly go stale. There are two distinct ways this happens: data drift and concept drift. 🌍
What Data Drift Means
Data drift happens when the input data changes shape over time. The features your model sees now no longer look like the features it learned from.
A Data Drift Example
A loan model trained pre-pandemic suddenly sees very different incomes and balances. The inputs shifted, even though what they mean has not. That is pure data drift.
What Concept Drift Means
Concept drift is sneakier: the inputs may look the same, but the relationship between inputs and the target has changed under your feet.
A Concept Drift Example
A spam filter sees normal-looking emails, but spammers invent new tricks. The same words now mean something different, so the input-to-label mapping shifted. That is concept drift.
The Key Difference
Data drift changes P(X), the distribution of inputs. Concept drift changes P(y given X), the mapping from inputs to the label. Tell them apart by asking which part moved. 🔑
Why It Matters for You
Data drift can hurt accuracy, but sometimes the model still copes. Concept drift almost always hurts, because the truth itself has moved away from what you learned.
Detecting Each Type
You can spot data drift by watching input distributions, with no labels needed. Concept drift usually needs outcomes, since you must compare predictions against real results.
Naming Your Inputs
It helps to keep a clean reference of your training inputs versus current ones. A tiny config makes drift checks easy to wire up later.
reference = train_df[["income", "balance", "age"]]
current = live_df[["income", "balance", "age"]]
features = list(reference.columns)Label Drift Too
A close cousin is label drift: the mix of target classes changes, like fraud jumping from 1% to 5%. It can signal either data or concept drift behind the scenes.
Drift Is Normal, Not Rare
Drift is not a freak event; it is the default state of any live model. Your job is not to prevent it but to detect it early and respond on time. ✅
Quick Check
Let us make sure you can tell the two apart.
Recap
Two kinds of shift: data drift moves the inputs, concept drift moves the meaning. Ask which part changed, and remember drift is normal, so always be ready to detect it. 🎯
常见问题解答
「数据偏移与概念偏移」课时是免费的吗？
是的 — 「数据偏移与概念偏移」的完整文本可在网页上免费阅读。要进行交互式练习（内置代码编辑器和全天候 AI 导师）并解锁 MLOps Academy 课程的其余内容，请升级到 CoddyKit PRO。 MLOps Academy 课程共包含 4 节课。
「数据偏移与概念偏移」这节课中我会学到什么？
区分输入变化和目标变化 你通过在浏览器中直接运行的动手代码来练习 MLOps Academy，全天候 AI 导师会在你学习这节课的过程中回答你的问题。
学习 MLOps Academy 需要有经验吗？
无需任何先前经验。CoddyKit 上的 MLOps Academy 课程适合初学者到高级学习者，你可以从这里开始或从头开始，按照自己的节奏学习。 这是第 1 节课，共 4 节。
「数据偏移与概念偏移」课时需要多长时间？
大多数 CoddyKit 课程大约需要 5–10 分钟。每节课都很精短且互动，所以你能稳步进步，并在网页和应用中从离开的地方继续。
我能在这节 MLOps Academy 课中编写并运行代码吗？
能。每节 MLOps Academy 课都包含内置代码编辑器，你可以在浏览器中直接编写并运行真实代码，并获得即时 AI 反馈 — 无需本地设置。