所属模块：
M2 · 经典机器学习 (Classical Machine Learning)| 专题分类：SVM 与核方法 (Support Vector Machines & Kernels)| 难度等级：Hard
一、核心一句话结论 (One-Sentence Summary)
敏感：少数类支持向量少、噪声点被当作支持向量。缓解：类权重、软间隔、清理噪声。
SVM is highly sensitive to class imbalance (which shifts the hyperplane toward the minority class) and label noise near the margin; mitigated via class-weighted penalties ($C_+ ne C_-$) and slack variable bounds.
二、核心考点要义 (Key Insights)
- 📌 加权 C 使少数类更受重视
- 📌 SVM 不输出概率，需 Platt 校准
English Insights:
– Class Imbalance Sensitivity: The margin objective balances $|w|2^2$ against misclassification penalties; when majority class dominates, the hyperplane collapses toward the minority class, predicting all majority.
– Label Noise Sensitivity: Misclassified points near the boundary have slack $xi_i > 1$ and non-zero $alpha_i = C$, exerting maximum possible leverage on the hyperplane orientation.
– Class-Weighted SVM: Assigns asymmetric penalty weights $C+ = frac{N}{2 N_+} C$ and $C_- = frac{N}{2 N_-} C$ (class_weight='balanced').
三、核心数学原理与机理推导 (Mathematical Principles & Derivation)
$$text{class weight}: C_i=Ccdot w_{y_i}$$
敏感性的机制：① 类别不平衡——SVM 的目标是最小化 Σξᵢ（所有样本的违反之和），多数类样本数多故主导目标，模型会倾向’牺牲少数类’（把少数类样本判错以换取更多多数类样本正确），导致少数类召回极低。② 噪声标签——标签错误的样本若位于决策边界附近，会成为支持向量并直接扭曲边界（因为 SVM 只由支持向量决定）；尤其当噪声样本被强制正确分类时（硬间隔或大 C），边界会被严重拉扯。③ 类别重叠——当两类在特征空间高度重叠时，SVM 的线性边界（或核诱导的平滑边界）无法很好分离，性能受限。
📖 查看英文严格数学推导 (English Mathematical Derivation)
Class-Weighted SVM objective: $min_{w, b, xi} frac{1}{2}|w|_2^2 + C_+ sum_{i: y_i=+1} xi_i + C_- sum_{j: y_j=-1} xi_j$ subject to standard slack constraints. In the dual formulation: the box constraints on Lagrange multipliers become asymmetric: $0 le alpha_i le C_+$ for positive instances, and $0 le alpha_j le C_-$ for negative instances. Setting $C_+ = C cdot frac{N_-}{N_+}$ forces the optimizer to treat a minority false negative as far more costly than a majority false positive, pushing the decision boundary away from the minority cluster.
四、工业级落地权衡与工程考量 (Industrial Trade-offs)
缓解手段：① 类权重——把 C 改为 Cᵢ=C·w_{yᵢ}，少数类用更大权重（如 w∝1/类别频率），使目标平衡；sklearn 的 class_weight='balanced' 自动按频率倒数加权。② 软间隔与调小 C——允许更多违反，避免被噪声样本拉扯；C 越小对噪声越鲁棒（但可能欠拟合）。③ 核选择——RBF 核的局部性使其对远离边界的噪声较鲁棒，但对边界附近的噪声仍敏感。④ 数据层面——清理噪声标签（用交叉验证检测）、过采样少数类（SMOTE）、欠采样多数类（注意在 CV 折内做，防泄漏）。⑤ 概率输出与阈值调整——SVM 默认输出决策值而非概率，若需按业务调整阈值（如召回优先），需先用 Platt scaling 或 isotonic regression 在验证集上校准，再按代价矩阵选阈值。⑥ 替代方案——不平衡 + 噪声场景下，集成方法（RF/GBDT 配 class_weight 或 Focal loss）通常比 SVM 更易调优。
⚙️ 查看英文落地权衡分析 (English Systems & Trade-offs)
Remediation toolkit: (1) Class-weighted SVM: Built into scikit-learn via `class_weight=’balanced’`. (2) Threshold Moving: Predict continuous decision function scores $f(x) = w^T phi(x) + b$ and tune the classification threshold $tau$ using precision-recall curves rather than raw $f(x) ge 0$. (3) Under-sampling the majority class: Reduces training set size while balancing classes, providing $10times$ faster quadratic programming convergence.
五、常见面试避坑陷阱 (Common Pitfalls & Traps)
- ⚠️ 用硬间隔或不调 C 处理噪声标签
- ⚠️ 不做校准就把 SVM 决策值当概率使用
English Pitfalls:
– Using standard unweighted SVM on 99:1 imbalanced fraud data (model will achieve 99% accuracy by predicting zero fraud).
– Setting $C$ very large in the presence of label noise (forces SVM to twist the decision boundary to fit mislabeled points).
六、高频深度面试追问与预测 (Follow-Up Questions)
- SVM 如何输出概率？（Platt scaling）
- How does the $nu$-SVM formulation parameterize the fraction of margin errors using parameter $nu in (0, 1)$?
- 为什么 SVM 对重叠类效果一般？
- Why does Support Vector Data Description (SVDD / One-Class SVM) excel at extreme anomaly detection?
七、知识图谱对齐 (Knowledge Graph Anchor)
- 🔗 关联底层卡片：
支持向量机 SVM：几何间隔、软间隔、对偶性与 RBF 核技巧(Support Vector Machines (SVM), Dual Formulation & Kernels) - 🗺️ 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。