Machine Learning Core Concepts for Computer Science Interviews | 计算机面试考点：机器学习核心知识梳理

📚 Machine Learning Core Concepts for Computer Science Interviews | 计算机面试考点：机器学习核心知识梳理

Machine learning has become a cornerstone of modern computer science interviews, particularly for roles in data science, AI engineering, and software development. This article consolidates the essential machine learning concepts you need to master before walking into any technical interview, presented in a clear, question-oriented format.

机器学习已成为现代计算机科学面试的基石，尤其是数据科学、AI 工程和软件开发岗位。本文以清晰的、面向考题的形式，为您系统梳理进入任何技术面试前必须掌握的机器学习核心概念。

1. Supervised vs. Unsupervised Learning | 监督学习与无监督学习

Supervised learning trains models on labeled data, where each training example includes an input-output pair. The model learns a mapping from inputs to outputs and generalizes to unseen data. Common tasks include classification and regression.

监督学习使用带标签的数据训练模型，每个训练样本包含输入-输出对。模型学习从输入到输出的映射，并泛化到未见过的数据。常见任务包括分类和回归。

Unsupervised learning, by contrast, works with unlabeled data. The algorithm seeks hidden structures, patterns, or groupings within the data without any explicit output guidance. Typical tasks include clustering, dimensionality reduction, and anomaly detection.

相比之下，无监督学习处理的是无标签数据。算法在没有明确输出指导的情况下，寻找数据中隐藏的结构、模式或分组。典型任务包括聚类、降维和异常检测。

A simple comparison table:

一个简单的对比表格：

Aspect	Supervised	Unsupervised
Data labels	Required	Not required
Goal	Predict output	Discover structure
Examples	Linear regression, SVM	K-Means, PCA

2. The Bias-Variance Tradeoff | 偏差-方差权衡

Bias refers to the error introduced by approximating a complex real-world problem with a simplified model. High bias leads to underfitting, where the model fails to capture the underlying patterns in the training data.

偏差是指用简化模型逼近复杂现实问题所引入的误差。高偏差会导致欠拟合，即模型未能捕捉训练数据中的潜在规律。

Variance refers to the model’s sensitivity to fluctuations in the training data. High variance leads to overfitting, where the model learns noise and random details rather than the true signal, performing poorly on new data.

方差是指模型对训练数据波动的敏感程度。高方差会导致过拟合，即模型学习的是噪声和随机细节而非真实信号，在新数据上表现不佳。

Total error can be decomposed as:

总误差可以分解为：

Total Error = Bias² + Variance + Irreducible Error

In interviews, you may be asked how increasing model complexity affects bias and variance. As complexity increases, bias typically decreases while variance increases. The optimal model balances these two sources of error.

面试中可能会问增加模型复杂度如何影响偏差和方差。随着复杂度增加，偏差通常降低而方差升高。最优模型需要在两类误差之间取得平衡。

3. Overfitting and Underfitting | 过拟合与欠拟合

Overfitting occurs when a model fits the training data too closely, capturing noise and outliers as if they were true patterns. Symptoms include extremely high training accuracy but poor validation or test accuracy.

过拟合是指模型对训练数据拟合得过紧，将噪声和离群点当作真实模式捕捉。表现症状是训练精度极高，但验证集或测试集精度很差。

Underfitting occurs when a model is too simple to capture the underlying structure of the data. Both training and test performance are poor, indicating the model has not learned enough.

欠拟合是指模型过于简单，无法捕捉数据背后的结构。训练和测试表现都差，说明模型没有学到足够的信息。

Common strategies to mitigate overfitting include:

缓解过拟合的常见策略包括：

Collecting more training data | 收集更多训练数据
Applying regularization (L1, L2) | 应用正则化（L1、L2）
Using dropout (in neural networks) | 使用丢弃法（神经网络中）
Early stopping during training | 训练期间提前停止
Cross-validation to monitor generalization | 使用交叉验证监控泛化能力

4. Regularization: L1 vs. L2 | 正则化：L1 与 L2

Regularization adds a penalty term to the loss function to discourage overly complex models. L1 regularization (Lasso) adds the sum of the absolute values of the coefficients as a penalty. It encourages sparsity, driving some coefficients to exactly zero, which is useful for feature selection.

正则化在损失函数中加入惩罚项，以抑制过于复杂的模型。L1 正则化（Lasso）以系数绝对值之和作为惩罚项，鼓励稀疏性，使部分系数精确变为零，可用于特征选择。

L2 regularization (Ridge) adds the sum of the squared coefficients as a penalty. It shrinks coefficients toward zero without forcing them to exactly zero, which helps distribute weight more evenly across features and reduces sensitivity to individual features.

L2 正则化（Ridge）以系数平方和作为惩罚项，将系数向零压缩但不强制为零，有助于在特征间更均匀地分配权重，降低对单个特征的敏感度。

Mathematically, for a linear model with loss L:

从数学角度看，对于损失 L 的线性模型：

L1 Loss: L + λ ∑|wᵢ|

L2 Loss: L + λ ∑wᵢ²

Here, λ is the regularization strength. A larger λ imposes a stronger penalty, increasing bias but reducing variance.

其中 λ 为正则化强度。λ 越大，惩罚越强，偏差增加但方差减小。

5. Gradient Descent and Its Variants | 梯度下降及其变体

Gradient descent is the most widely used optimization algorithm in machine learning. It iteratively updates model parameters in the direction opposite to the gradient of the loss function, thereby minimizing the loss.

梯度下降是机器学习中使用最广泛的优化算法。它迭代地沿损失函数梯度的反方向更新模型参数，从而最小化损失。

The parameter update rule is:

参数更新规则为：

θ = θ − η · ∇J(θ)

where η is the learning rate and ∇J(θ) is the gradient of the loss function with respect to the parameters.

其中 η 是学习率，∇J(θ) 是损失函数关于参数的梯度。

Three main variants are commonly discussed in interviews:

面试中常讨论的三种主要变体：

Batch Gradient Descent — computes the gradient using the entire dataset. Accurate but slow for large datasets. | 批量梯度下降 — 使用整个数据集计算梯度，准确但大数据集上速度慢。
Stochastic Gradient Descent (SGD) — uses one random sample per update. Fast but noisy and less stable. | 随机梯度下降（SGD） — 每次更新使用一个随机样本，速度快但噪声大、稳定性较差。
Mini-batch Gradient Descent — uses a small batch of samples per update, balancing speed and stability. | 小批量梯度下降 — 每次更新使用一小批样本，兼顾速度与稳定性。

Learning rate selection is critical. Too high a rate may cause divergence; too low a rate results in extremely slow convergence.

学习率的选择至关重要。学习率过大会导致发散，过小则收敛极慢。

6. Model Evaluation Metrics | 模型评估指标

Different tasks require different evaluation metrics. For classification, the most fundamental metrics include accuracy, precision, recall, and F1-score.

不同任务需要不同的评估指标。对于分类任务，最基础的指标包括准确率、精确率、召回率和 F1 分数。

A confusion matrix is a valuable tool for understanding classifier performance:

混淆矩阵是理解分类器表现的重要工具：

	Predicted Positive	Predicted Negative
Actual Positive	True Positive (TP)	False Negative (FN)
Actual Negative	False Positive (FP)	True Negative (TN)

Key formulas:

关键公式：

Accuracy = (TP + TN) / (TP + TN + FP + FN)

Precision = TP / (TP + FP)

Recall = TP / (TP + FN)

F1 Score = 2 × (Precision × Recall) / (Precision + Recall)

Precision measures how many of the predicted positives are actually positive; recall measures how many actual positives were correctly identified. The F1 score is the harmonic mean of precision and recall, useful when class distribution is imbalanced.

精确率衡量预测为正类的样本中有多少确实是正类；召回率衡量实际正类中有多少被正确识别。F1 分数是精确率和召回率的调和平均数，适合类别分布不平衡的场景。

7. Cross-Validation | 交叉验证

Cross-validation assesses how well a model generalizes to an independent dataset. The most common method is k-fold cross-validation: the data is split into k equal folds; the model is trained on k−1 folds and validated on the remaining fold. This process repeats k times with each fold serving as the validation set exactly once.

交叉验证用于评估模型对独立数据集的泛化能力。最常用的方法是 k 折交叉验证：将数据分成 k 个相等的子集；用 k−1 个子集训练模型，在剩余的子集上验证。该过程重复 k 次，每个子集恰好作为一次验证集。

The final performance metric is the average across all k iterations. This approach is especially valuable when the dataset is small, as it maximizes both training and validation data usage.

最终性能指标是 k 次迭代的平均值。当数据集较小时，该方法尤其有价值，因为它最大化了训练和验证数据的使用效率。

Leave-one-out cross-validation (LOOCV) is an extreme case where k equals the number of samples. It is computationally expensive but provides a nearly unbiased estimate of model performance.

留一交叉验证（LOOCV）是 k 等于样本数的极端情况。它的计算成本高，但能提供近乎无偏的模型性能估计。

8. Core Algorithms: Linear and Logistic Regression | 核心算法：线性回归与逻辑回归

Linear regression models the relationship between input features and a continuous target variable using a linear function. Its objective is to minimize the mean squared error (MSE) between predictions and actual values.

线性回归使用线性函数对输入特征与连续目标变量之间的关系进行建模。其目标是最小化预测值与真实值之间的均方误差（MSE）。

J(w) = (1/2m) ∑(h(x⁽ⁱ⁾) − y⁽ⁱ⁾)²

Logistic regression, despite its name, is used for classification tasks. It applies a sigmoid function to the linear combination of inputs, producing a probability score between 0 and 1.

逻辑回归虽然名字中带”回归”，实际用于分类任务。它对输入的线性组合施加 sigmoid 函数，输出一个介于 0 和 1 之间的概率分数。

h(x) = 1 / (1 + e⁻ᶻ) , where z = wᵀx + b

The decision boundary is typically set at 0.5: if h(x) ≥ 0.5, predict class 1; otherwise, predict class 0. Logistic regression uses the log-loss (cross-entropy) as its objective function rather than MSE.

决策边界通常设为 0.5：如果 h(x) ≥ 0.5，预测为类别 1；否则预测为类别 0。逻辑回归使用对数损失（交叉熵）作为目标函数，而非均方误差。

9. Core Algorithms: Decision Trees and Random Forests | 核心算法：决策树与随机森林

Decision trees split the feature space recursively into regions, selecting the best feature and split point at each node based on criteria such as Gini impurity or information gain. They are intuitive and require little data preprocessing, but are prone to overfitting.

决策树递归地将特征空间划分为若干区域，在每个节点基于基尼不纯度或信息增益等准则选择最佳特征和分割点。它们直观且几乎不需要数据预处理，但容易过拟合。

Information gain for splitting on a feature is defined as the reduction in entropy:

按某个特征分裂的信息增益定义为熵的减少量：

Information Gain = Entropy(parent) − ∑ (nₖ/n) · Entropy(childₖ)

Random forests address the overfitting problem by constructing many decision trees on bootstrapped samples of the data and using random subsets of features at each split. The final prediction is the majority vote (classification) or average (regression) of all trees.

随机森林通过在数据的自助抽样样本上构建多棵决策树，并在每次分裂时使用随机特征子集，从而解决过拟合问题。最终预测是所有树的多数投票（分类）或平均值（回归）。

Random forests improve robustness and generalization by reducing variance while maintaining low bias, making them a strong baseline model in many competitions.

随机森林通过降低方差同时保持低偏差来提高鲁棒性和泛化能力，使其成为许多竞赛中强劲的基线模型。

10. Support Vector Machines and K-Nearest Neighbors | 支持向量机与 K 近邻

Support Vector Machines (SVM) aim to find the hyperplane that maximizes the margin between classes. The support vectors are the data points closest to the hyperplane, which alone determine the decision boundary. SVM can handle non-linear boundaries via the kernel trick, which implicitly maps data into a higher-dimensional space without explicitly computing the transformation.

支持向量机（SVM）旨在找到最大化类别间隔的超平面。支持向量是距离超平面最近的数据点，仅由它们决定决策边界。SVM 可以通过核技巧处理非线性边界，该技巧隐式地将数据映射到更高维空间，而无需显式计算变换。

Common kernels include linear, polynomial, and Radial Basis Function (RBF). The choice of kernel and its hyperparameters (e.g., C, γ) significantly affects performance.

常用核函数包括线性核、多项式核和径向基函数（RBF）核。核函数及其超参数（如 C、γ）的选择显著影响性能。

K-Nearest Neighbors (KNN) is a lazy learning algorithm that classifies a new sample by the majority label among its k nearest neighbors in feature space. The distance metric (e.g., Euclidean distance) and k value are critical choices. KNN requires no training phase but is computationally expensive at inference for large datasets.

K 近邻（KNN）是一种惰性学习算法，通过特征空间中 k 个最近邻的多数标签对新样本进行分类。距离度量（如欧氏距离）和 k 值是关键选择。KNN 无需训练阶段，但在大数据集上推理时计算开销大。

11. Clustering-Based Learning | 基于聚类的学习

K-Means is the most iconic clustering algorithm. It partitions data into k clusters by iteratively assigning each point to the nearest centroid and recomputing centroids as the mean of assigned points. The objective is to minimize the within-cluster sum of squares (inertia).

K-Means 是最具代表性的聚类算法。它通过迭代地将每个点分配给最近的质心并重新计算质心为分配点的均值，将数据划分为 k 个簇。其目标是最小化簇内平方和（惯性）。

The algorithm terminates when centroids stop changing significantly. Key limitations include sensitivity to initial centroid selection and the requirement to specify k in advance. The elbow method is commonly used to select k by plotting inertia against k and finding the “elbow” point.

算法在质心不再显著变化时终止。主要局限包括对初始质心选择敏感，以及需要预先指定 k 值。常用肘部法选择 k：绘制惯性随 k 变化的曲线，找到”肘部”点。

Interview questions often explore: how to initialize centroids (K-Means++), how to choose k, and how K-Means differs from hierarchical clustering or DBSCAN.

面试题常考察：如何初始化质心（K-Means++）、如何选择 k，以及 K-Means 与层次聚类或 DBSCAN 的区别。

12. Interview Tips and Common Pitfalls | 面试技巧与常见误区

Beyond knowing the algorithms, interviewers evaluate your ability to reason about model selection, diagnose errors, and communicate tradeoffs clearly. Always clarify whether the problem is classification or regression, and ask about data size, feature types, and evaluation criteria before proposing a solution.

除了理解算法本身，面试官还会评估你推理模型选择、诊断错误以及清晰沟通权衡的能力。在提出方案前，务必确认问题是分类还是回归，并询问数据规模、特征类型和评估标准。

Common pitfalls in interviews include confusing precision with recall, omitting the learning rate from gradient descent discussions, failing to mention regularization when discussing overfitting, and neglecting to consider computational cost when comparing algorithms.

面试中的常见误区包括：混淆精确率和召回率、讨论梯度下降时遗漏学习率、讨论过拟合时未提及正则化，以及比较算法时忽略计算成本。

Finally, always ground your answers in practical reasoning. For example, prefer logistic regression for a small, interpretable model with clean features, but consider random forests or gradient boosting for complex, mixed-type data with moderate sample sizes.

最后，务必以实际推理为基础来作答。例如，当模型需要小而可解释且特征干净时，优先选择逻辑回归；当数据复杂、类型混合且样本量中等时，则可考虑随机森林或梯度提升。

Published by TutorHao | Computer Science Revision Series | aleveler.com

更多咨询请联系16621398022（同微信）

Machine Learning Core Concepts for Computer Science Interviews | 计算机面试考点：机器学习核心知识梳理