为何SVM被视为“黑箱”机器学习算法？（基于吴恩达Coursera课程疑问）

Great question! Let's unpack why SVMs earn the "black box" label, especially when compared to more interpretable models like logistic regression.

Core Reason: Decision Process Is Less Intuitive & Hard to Translate to Plain Language #

The term "black box" here doesn't mean SVMs are completely ununderstandable—it refers to how their internal decision logic is abstract, making it tough to explain why a prediction was made using simple, business-friendly terms. Let's break down the key differences from logistic regression:

• Non-linear mapping via kernel functions (the biggest culprit)
  Logistic regression is a linear model: its decision boundary is a straight line (or hyperplane in higher dimensions), and each coefficient directly corresponds to a feature's impact on the prediction. For example, you can say, "Every 1-unit increase in feature A raises the probability of a positive class by 15%"—that's straightforward.
  SVMs, though, often use kernel functions (like RBF or polynomial kernels) to implicitly map raw features into extremely high-dimensional spaces to find an optimal separating hyperplane. This mapping is invisible to you—you can't visualize or quantify how individual features change in that high-dimensional space. When a kernelized SVM makes a prediction, you can't trace it back to a simple feature-weight relationship; you can only say it's "close to the support vectors of the positive class."

• Dependency on a small set of "support vectors"
  Logistic regression uses every training sample to tune its coefficients, so every data point contributes to the final model. SVMs, by contrast, only care about support vectors—the small subset of samples that lie closest to the decision boundary. This means the model's decisions are driven by a tiny fraction of your data, but it's rarely obvious why those specific samples are the critical ones, especially when working with high-dimensional feature sets. You can't easily link these support vectors to meaningful patterns in your data.

• Lack of native probability interpretation
  Logistic regression outputs a direct probability (e.g., "This sample has a 78% chance of being positive"), which is inherently easy to explain to stakeholders. Standard SVMs output a "distance from the decision boundary" (a confidence score), not a probability. While you can add extra steps like Platt scaling to convert this to probabilities, the result doesn't have the same clear statistical interpretation as logistic regression's outputs.

Is SVM Less Intuitive Than Logistic Regression? #

Absolutely. Logistic regression's linear structure and probabilistic outputs make it naturally interpretable—you can tie every prediction back to specific feature impacts. Kernelized SVMs, however, rely on abstract mathematical transformations that don't translate to simple, actionable explanations. Even linear SVMs (which don't use kernels) are slightly less intuitive than logistic regression, since their coefficients represent margin maximization rather than direct probability shifts.

That said, it's worth noting: linear SVMs can be interpreted similarly to logistic regression—their weights indicate feature importance for the decision boundary. The "black box" reputation mostly applies to kernelized SVMs, which are the most commonly used in real-world scenarios (and likely what Andrew Ng was referencing).

内容的提问来源于stack exchange，提问作者Stephen