📚 Introduction to Machine Learning for IB/AQA Computer Science | 机器学习入门：IB/AQA 计算机考点精讲
Machine learning (ML) is transforming how computer systems acquire intelligence without being explicitly programmed. For IB and AQA Computer Science students, mastering the fundamental concepts—from supervised learning and neural networks to overfitting and ethical implications—is essential. This revision guide breaks down each key topic with clear explanations, real-world analogies, and assessment-relevant detail, helping you build a solid foundation for exam success.
机器学习正在改变计算机系统无需显式编程即可获得智能的方式。对于 IB 和 AQA 计算机科学学生而言，掌握从监督学习和神经网络到过拟合与伦理影响的基本概念至关重要。本考点精讲通过清晰的解释、贴近生活的类比以及紧扣评估的细节，逐一剖析关键主题，助你打下坚实基础，从容应对考试。
1. What is Machine Learning? | 什么是机器学习？
Machine learning is a branch of artificial intelligence where systems learn patterns from data, improving their performance on a task over time. Rather than following hard-coded rules, an ML model identifies statistical relationships in training examples and uses them to make predictions or decisions on new, unseen data. This data-driven approach underpins everything from spam filters to voice assistants.
机器学习是人工智能的一个分支，系统从数据中学习模式，并随时间改进在任务上的表现。机器学习模型并非遵循固定编码的规则，而是在训练样本中识别统计关系，并用其对新数据进行预测或决策。这种数据驱动的方法支撑着从垃圾邮件过滤器到语音助手等各种应用。
A classic definition by Tom Mitchell states: “A computer program is said to learn from experience E with respect to some class of tasks T and performance measure P, if its performance at tasks in T, as measured by P, improves with experience E.” This framework highlights the core components: task, experience, and performance metric.
Tom Mitchell 的经典定义指出：“对于某类任务 T 和性能度量 P，如果一个计算机程序在任务 T 上的性能（用 P 衡量）随着经验 E 而提高，则称该程序从经验 E 学习。”该框架突显了核心要素：任务、经验和性能度量。
2. Types of Machine Learning | 机器学习的类型
ML algorithms are typically categorised into supervised, unsupervised, and reinforcement learning. In supervised learning, the training data includes input–output pairs (labelled data), and the goal is to learn a mapping from inputs to outputs. Unsupervised learning deals with unlabelled data, discovering hidden structure such as clusters or associations. Reinforcement learning involves an agent that learns to act in an environment by receiving rewards or penalties.
机器学习算法通常分为监督学习、无监督学习和强化学习。在监督学习中，训练数据包含输入–输出对（带标签数据），目标是学习从输入到输出的映射。无监督学习处理无标签数据，发现隐藏的结构（如聚类或关联）。强化学习则涉及一个智能体，通过接收奖励或惩罚来学习如何在环境中行动。
For IB and AQA exams, you may also encounter semi-supervised learning, which uses a small amount of labelled data alongside a large amount of unlabelled data, and self-supervised learning, where a model generates its own supervisory signal from the input data.
在 IB 和 AQA 考试中，你或许还会遇到半监督学习（使用少量带标签数据和大量无标签数据）和自监督学习（模型从输入数据中自行生成监督信号）。
3. Supervised Learning: Key Concepts | 监督学习的关键概念
Supervised learning tasks are divided into regression (predicting continuous values) and classification (predicting discrete categories). A dataset is composed of features (independent variables) and labels (the target output). For instance, predicting house prices uses features like area and number of bedrooms to output a continuous price—a regression task. Classifying emails as ‘spam’ or ‘not spam’ uses features like word frequencies to output a binary label—a classification task.
监督学习任务分为回归（预测连续值）和分类（预测离散类别）。数据集由特征（自变量）和标签（目标输出）组成。例如，预测房价用面积和卧室数等特征输出一个连续价格——这是回归任务。将电子邮件分为“垃圾邮件”或“非垃圾邮件”则用词频等特征输出二元标签——这是分类任务。
Models learn a function f such that y ≈ f(x), where x is the feature vector and y is the label. The learning algorithm adjusts parameters to minimise a loss function, such as mean squared error for regression or cross-entropy for classification.
模型学习一个函数 f，使得 y ≈ f(x)，其中 x 为特征向量，y 为标签。学习算法调整参数以最小化损失函数，如回归的均方误差或分类的交叉熵。
Mean Squared Error (MSE) = (1/n) Σ (yᵢ − ŷᵢ)²
4. Common Supervised Algorithms | 常见监督学习算法
k-Nearest Neighbours (k-NN) classifies a data point based on the majority class among its k closest neighbours in feature space. It is a lazy learner that makes no assumptions about data distribution but can be computationally heavy during prediction.
k-最近邻 (k-NN) 根据特征空间中 k 个最近邻的多数类别对数据点进行分类。它是一种惰性学习器，对数据分布不做假设，但预测时计算量大。
Decision trees split data recursively using feature thresholds to maximise information gain or minimise Gini impurity. They are interpretable but prone to overfitting. Ensemble methods like random forests combine many trees to improve robustness.
决策树 通过特征阈值递归分割数据，以最大化信息增益或最小化基尼不纯度。它们可解释性强但容易过拟合。像随机森林这样的集成方法组合多棵树以提高鲁棒性。
Linear regression fits a line (or hyperplane) by minimising the sum of squared residuals. Logistic regression, despite its name, is used for binary classification by modelling the probability of class membership with a sigmoid function.
线性回归 通过最小化残差平方和来拟合一条直线（或超平面）。逻辑回归 虽然名为回归，但通过使用 Sigmoid 函数对类别归属概率建模，用于二元分类。
5. Unsupervised Learning: Clustering & Association | 无监督学习：聚类与关联
Unsupervised learning discovers patterns without labelled responses. Clustering groups similar data points, while association rule learning uncovers relationships between variables in large datasets.
无监督学习在没有标签响应的情况下发现模式。聚类将相似的数据点分组，而关联规则学习揭示大数据集中变量之间的关系。
K-means clustering partitions data into k clusters by iteratively minimising the distance between points and their assigned cluster centroids. It requires specifying k in advance and is sensitive to initial centroid positions. The algorithm alternates between assigning points to the nearest centroid and updating centroids as the mean of assigned points.
K-均值聚类 通过迭代最小化数据点与其分配簇质心之间的距离，将数据划分为 k 个簇。需预先指定 k，并对初始质心位置敏感。算法在将点分配给最近质心和基于所分配点的均值更新质心之间交替进行。
Apriori algorithm for association mining finds frequent itemsets and generates rules like “if {milk, bread} then {butter}”. Confidence and support metrics determine rule strength.
用于关联挖掘的Apriori 算法 发现频繁项集并生成规则，如“如果{牛奶, 面包}则{黄油}”。置信度和支持度指标决定规则强度。
6. Reinforcement Learning Basics | 强化学习基础
In reinforcement learning, an agent interacts with an environment by performing actions and receiving rewards. It learns a policy —a mapping from states to actions—that maximises cumulative reward over time. No pre-labelled data is provided; the agent explores and exploits to learn successfully.
在强化学习中，智能体通过执行动作并与环境交互来接收奖励。它学习一个策略——从状态到动作的映射——以最大化随时间累积的奖励。没有提供预标记的数据；智能体通过探索和利用进行学习。
The process is modelled as a Markov decision process (MDP) with states, actions, transition probabilities, and reward functions. Temporal-difference methods such as Q-learning enable agents to update value estimates based on current experiences. Exploration vs exploitation is a fundamental trade-off: should the agent try new actions or stick to known rewarding ones?
该过程被建模为马尔可夫决策过程 (MDP)，包含状态、动作、转移概率和奖励函数。诸如 Q-learning 的时间差分方法使智能体能基于当前经验更新值估计。探索与利用是一个根本性的权衡：智能体应当尝试新动作还是坚持已知的高回报动作？
7. Training, Validation and Test Data | 训练、验证与测试数据
A critical ML practice is splitting data into three non-overlapping sets: training set (to fit the model), validation set (to tune hyperparameters and prevent overfitting), and test set (to evaluate final performance). The test set must only be used once to avoid information leakage.
一个关键的机器学习实践是将数据划分为三个互不重叠的集合：训练集（用于拟合模型）、验证集（用于调整超参数并防止过拟合）和测试集（用于评估最终性能）。测试集只能使用一次，以免信息泄露。
Cross-validation, especially k-fold cross-validation, provides a more reliable estimate by repeatedly partitioning the data into folds, training on k−1 folds and validating on the remaining one, then averaging the results. This reduces variance due to a single random split.
交叉验证，尤其是 k 折交叉验证，通过反复将数据划分为若干折，在 k−1 折上训练并在剩余一折上验证，然后对结果取平均，从而提供更可靠的估计。这降低了由单次随机划分引起的方差。
8. Overfitting and Underfitting | 过拟合与欠拟合
An overfitted model captures noise in the training data rather than the true underlying pattern, resulting in excellent training accuracy but poor generalisation to new data. It often has high variance. Underfitting occurs when the model is too simple to capture the pattern, yielding poor performance on both training and test data (high bias).
过拟合 的模型捕获了训练数据中的噪声而非真实底层模式，导致训练精度很高但泛化到新数据时表现差。它通常具有高方差。欠拟合 发生在模型过于简单无法捕捉模式时，导致在训练和测试数据上都表现不佳（高偏差）。
Techniques to combat overfitting include: using more training data, reducing model complexity, applying regularisation (L1/L2 penalties on weights), and early stopping (halting training when validation error starts to rise). In decision trees, pruning removes branches that provide little predictive power.
防止过拟合的技术包括：使用更多训练数据、降低模型复杂度、应用正则化（对权重施加 L1/L2 惩罚）以及早停（当验证误差开始上升时停止训练）。在决策树中，剪枝 移除提供很少预测能力的分支。
9. Performance Metrics | 性能评估指标
Selecting appropriate evaluation metrics is essential and depends on the problem type. For classification, accuracy is common but can be misleading for imbalanced datasets. A confusion matrix lays out true positives, true negatives, false positives, and false negatives.
选择合适的评估指标至关重要，并取决于问题类型。对于分类，准确率很常用，但在不平衡的数据集中可能具有误导性。混淆矩阵 展示了真正例、真负例、假正例和假负例。
From the confusion matrix we derive: Precision = TP/(TP+FP) (exactness), Recall = TP/(TP+FN) (completeness), and F1-score = 2 × (Precision × Recall)/(Precision + Recall) (harmonic mean). For regression, Root Mean Squared Error (RMSE) and R² are commonly used.
由混淆矩阵可推导出：精确率 = TP/(TP+FP)（精确性），召回率 = TP/(TP+FN)（完备性），以及 F1 得分 = 2 × (精确率 × 召回率)/(精确率 + 召回率)（调和平均值）。在回归中，常用均方根误差 (RMSE) 和 R²。
IB/AQA questions often ask you to interpret precision-recall trade-offs or suggest a metric given a scenario (e.g., high recall is vital in disease screening).
IB/AQA 题目常要求你解释精确率–召回率的权衡，或根据情景建议指标（例如，在疾病筛查中高召回率至关重要）。
10. Data Preprocessing | 数据预处理
Real-world data is messy. Preprocessing steps include handling missing values (deletion or imputation using mean/median/mode), normalising or standardising features to bring them onto similar scales, and encoding categorical variables (one-hot encoding or label encoding).
真实世界的数据杂乱无章。预处理步骤包括处理缺失值（删除或用均值/中位数/众数填补）、对特征进行归一化或标准化以使其具有相似的尺度，以及对分类变量进行编码（独热编码或标签编码）。
Feature scaling is critical for distance-based algorithms like k-NN and gradient descent in neural networks. Standardisation (z-score normalisation) subtracts the mean and divides by the standard deviation, while min-max scaling squashes values into a [0,1] range. Outlier detection and removal can also improve model robustness.
特征缩放 对于 k-NN 等基于距离的算法和神经网络中的梯度下降至关重要。标准化（z 分数归一化）减去均值并除以标准差，而最小-最大缩放将值压缩至 [0,1] 范围。异常值检测与处理也能提升模型的稳健性。
11. Neural Networks and Deep Learning | 神经网络与深度学习简介
Artificial neural networks are inspired by the biological brain. They consist of layers of interconnected nodes (neurons) where each connection has a weight. A neuron computes a weighted sum of its inputs, adds a bias, and passes the result through a non-linear activation function such as ReLU or sigmoid.
人工神经网络受到生物大脑的启发，由相互连接的节点（神经元）层组成，每条连接具有一个权重。神经元计算其输入的加权和，加上偏置，并将结果通过非线性激活函数（如 ReLU 或 Sigmoid）传递。
Training a neural network uses backpropagation to calculate gradients of the loss with respect to each weight, combined with an optimiser like stochastic gradient descent (SGD) to update weights. Deep learning refers to networks with many hidden layers, capable of learning hierarchical features. Convolutional neural networks (CNNs) excel at image data, while recurrent neural networks (RNNs) suit sequential data.
训练神经网络使用反向传播 计算损失对每个权重的梯度，并结合随机梯度下降 (SGD) 等优化器来更新权重。深度学习指具有多个隐藏层的网络，能学习层次化特征。卷积神经网络 (CNN) 擅长图像数据，而循环神经网络 (RNN) 适合序列数据。
12. Ethical Considerations | 伦理考量
Machine learning systems can inadvertently perpetuate societal biases present in training data, leading to unfair outcomes in hiring, lending, and criminal justice. Algorithmic fairness and transparency are pressing concerns, and model interpretability (e.g., using SHAP or LIME) helps stakeholders understand decision rationales.
机器学习系统可能无意中延续训练数据中存在的社会偏见，导致在招聘、贷款和刑事司法等领域产生不公平的结果。算法公平 和透明度 是紧迫的问题，模型可解释性（例如使用 SHAP 或 LIME）有助于利益相关者理解决策依据。
Privacy is another critical dimension; models trained on personal data must adhere to regulations such as GDPR. Techniques like differential privacy and federated learning aim to protect individual data. Moreover, the environmental footprint of training large models and the accountability for autonomous decisions are topics frequently discussed in IB/AQA exams.
隐私是另一个关键维度；基于个人数据训练的模型必须遵守 GDPR 等法规。差分隐私 和联邦学习 等技术旨在保护个体数据。此外，训练大模型的环境足迹以及自主决策的责任归属，是 IB/AQA 考试中经常讨论的话题。
Students should be prepared to discuss how biases emerge (e.g., skewed sampling, proxy variables) and propose mitigation strategies such as diverse training sets, bias audits, and human-in-the-loop oversight.
学生应准备好讨论偏见如何产生（如抽样偏差、代理变量），并提出缓解策略，如使用多样化的训练集、进行偏见审计和保持人机协同监督。
Published by TutorHao | Computer Science Revision Series | aleveler.com
更多咨询请联系16621398022（同微信）
屏轩国际教育cambridge primary/secondary checkpoint, cat4, ukiset,ukcat,igcse,alevel,PAT,STEP,MAT, ibdp,ap,ssat,sat,sat2课程辅导，国外大学本科硕士研究生博士课程论文辅导