所属模块：
M2 · 经典机器学习 (Classical Machine Learning)| 专题分类：逻辑回归与 GLM (Logistic Regression & GLM)| 难度等级：Medium
一、核心一句话结论 (One-Sentence Summary)
GLM = 随机成分（指数族分布）+ 系统成分（线性预测子）+ 链接函数。
A Generalized Linear Model (GLM) extends linear regression to non-Gaussian distributions within the exponential family; composed of (1) Random Component (exponential family distribution), (2) Systematic Component (linear predictor $eta = Xbeta$), and (3) Link Function ($g(mu) = eta$).
二、核心考点要义 (Key Insights)
- 📌 线性回归：Gaussian + identity
- 📌 逻辑回归：Bernoulli + logit
- 📌 Poisson 回归：Poisson + log，用于计数率
English Insights:
– 1. Random Component: The target variable $y$ follows an exponential family distribution (Gaussian, Bernoulli, Poisson, Gamma) with mean $mu = E[ymid x]$.
– 2. Systematic Component: Linear predictor $eta = w^T x = sum_{j=1}^d w_j x_j$.
– 3. Link Function: A monotonic, differentiable function $g(cdot)$ that connects the conditional mean to the linear predictor: $eta = g(mu) iff mu = g^{-1}(eta)$.
三、核心数学原理与机理推导 (Mathematical Principles & Derivation)
$$mathbb E[Y]=mu=g^{-1}(Xw)$$
GLM 的三个组件：① 随机成分——指定 Y 的分布属于指数族（Gaussian、Bernoulli、Poisson、Gamma、Inverse Gaussian 等），指数族的统一形式为 f(y;θ)=exp[(yθ−b(θ))/a(φ)+c(y,φ)]，其中 b'(θ)=E[Y]、b”(θ)=Var[Y]·a(φ)，这使方差与均值通过方差函数 V(μ) 关联；② 系统成分——线性预测子 η=Xw；③ 链接函数——单调可微的 g 使 g(μ)=η，即 μ=g⁻¹(Xw)。三个经典组合：线性回归（Gaussian + identity）、逻辑回归（Bernoulli + logit）、Poisson 回归（Poisson + log）。
📖 查看英文严格数学推导 (English Mathematical Derivation)
The canonical link function equates the linear predictor $eta$ directly to the natural parameter $theta$ of the exponential family $p(ymid theta) = h(y)exp(theta y – A(theta))$. Since $mu = E[y] = A'(theta)$, the canonical link is $g(mu) = (A’)^{-1}(mu)$. Derivations of classical models: (1) Linear Regression: Normal distribution with $A(theta) = frac{theta^2}{2} implies A'(theta) = theta = mu$. The canonical link is identity: $g(mu) = mu = w^T x$. (2) Logistic Regression: Bernoulli distribution with $A(theta) = log(1 + e^theta) implies A'(theta) = frac{e^theta}{1+e^theta} = mu$. The canonical link is logit: $g(mu) = logleft(frac{mu}{1-mu}right) = w^T x$. (3) Poisson Regression: Poisson distribution with $A(theta) = e^theta implies A'(theta) = e^theta = mu$. The canonical link is log: $g(mu) = log(mu) = w^T x implies mu = e^{w^T x}$.
四、工业级落地权衡与工程考量 (Industrial Trade-offs)
GLM 的统一性带来三个实践优势：① 损失函数自动确定——由分布假设推出负对数似然，无需人工选择（如 Poisson 回归的损失是 Σ(μ−y log μ)，而非 MSE）；② 方差随均值变化——Poisson 的方差函数 V(μ)=μ（均值越大方差越大），这自动处理了计数数据的异方差，无需加权；③ 系数解释通过链接函数——log 链接下 e^{βⱼ} 是’率比’（rate ratio）。为什么 CTR 建模有时用 Poisson/Gamma：CTR 是’曝光中点击的比例’，若把曝光视为’暴露量’，点击数服从 Poisson（λ=曝光×CTR），用 Poisson 回归配合 offset 项 log(曝光) 可以正确处理不同曝光量下的方差差异——这比直接用二项逻辑回归在曝光量差异大时更稳。此外，过度离散的计数数据可用负二项 GLM（引入额外离散参数）或 Quasi-Poisson（只放宽方差为 φμ）。
⚙️ 查看英文落地权衡分析 (English Systems & Trade-offs)
GLMs eliminate ad-hoc non-linear transformations: (1) If modeling positive count data (e.g. Taxi trips or customer purchases), fitting OLS to $log(y)$ fails when $y=0$. Poisson or Negative Binomial regression directly models counts with exact discrete likelihoods. (2) Canonical links guarantee that the negative log-likelihood is strictly convex and that the Hessian equals the Fisher Information Matrix, ensuring fast, stable IRLS convergence.
五、常见面试避坑陷阱 (Common Pitfalls & Traps)
- ⚠️ 把 GLM 当作’任何非线性模型’（它要求指数族 + 链接函数）
- ⚠️ 对过度离散的计数数据直接用 Poisson GLM
English Pitfalls:
– Using Poisson regression when the variance vastly exceeds the mean (overdispersion; must use Negative Binomial or Quasi-Poisson GLM).
– Confusing the link function $g(mu)$ with transforming the raw response variable $g(y)$ (GLM models $g(E[y])$, not $E[g(y)]$; by Jensen’s inequality, $g(E[y]) ne E[g(y)]$).
六、高频深度面试追问与预测 (Follow-Up Questions)
- 为什么 CTR 建模常用 Poisson/Gamma？
- Why does Jensen’s inequality explain why $log(E[y])$ in Poisson regression is fundamentally different from fitting $E[log(y)]$ in OLS?
- log 链接如何保证预测非负？
- What is Tweedie regression and why is it the gold standard in insurance claims and ad-spend LTV modeling with mass at zero?
七、知识图谱对齐 (Knowledge Graph Anchor)
- 🔗 关联底层卡片：
逻辑回归 Sigmoid、Log-Odds 对数几率与广义线性模型(Logistic Regression, Log-Odds & Generalized Linear Models) - 🗺️ 知识图谱模块：
经典机器学习思维导图
🔬 算法科学家与机器学习深度考察全量题库 (Science Depth)
本题收录于 TalentMe 算法科学家深度考察真题库 (Science Depth)。全库共 856 道硬核考点，深度覆盖数学统计、经典ML、深度学习、Transformer、大语言模型、多模态、推荐系统与 MLOps。支持 Jev 面经智能匹配、一键离线单文件 HTML 手册导出并直连 Obsidian 本地记忆。