Key Concepts in Machine Learning for IB & OCR Computer Science | IB与OCR计算机科学机器学习考点精讲

📚 Key Concepts in Machine Learning for IB & OCR Computer Science | IB与OCR计算机科学机器学习考点精讲

Machine learning is transforming the way we interact with technology. For IB and OCR Computer Science students, understanding the core principles of machine learning is essential. This article breaks down the key concepts, algorithms, and evaluation techniques you need to master for your exams.

机器学习正在改变我们与技术交互的方式。对于IB和OCR计算机科学课程的学生来说，掌握机器学习的核心原理至关重要。本文将拆解你需要掌握的考点，包括核心概念、算法和评估技术。

1. What is Machine Learning? | 什么是机器学习？

Machine learning is a subset of artificial intelligence that enables systems to learn patterns from data without being explicitly programmed. Instead of following static instructions, models improve their performance through experience.

机器学习是人工智能的一个子集，它使系统能够从数据中学习模式，而无需显式编程。与遵循静态指令不同，模型通过经验来提升性能。

A traditional program uses rules and data to produce answers. A machine learning system uses data and answers to produce rules. This shift allows computers to solve problems that are too complex for hand-coded logic, such as image recognition or language translation.

传统程序使用规则和数据来产生答案。机器学习系统则使用数据和答案来产生规则。这种转变使得计算机可以解决对于手工编码逻辑来说过于复杂的问题，如图像识别或语言翻译。

Key examples include spam filters, recommendation engines, and self-driving cars.

关键示例包括垃圾邮件过滤器、推荐引擎和自动驾驶汽车。
At its core, machine learning relies on data, a model, and a learning algorithm that adjusts the model’s parameters to reduce error.

核心在于，机器学习依赖于数据、模型以及一个可以调整模型参数以减少误差的学习算法。

In IB Computer Science, you are expected to explain how ML differs from traditional computing; OCR specifications require you to describe the fundamental process of training and inference.

在IB计算机科学中，你需要解释机器学习与传统计算的区别；OCR大纲要求你描述训练和推理的基本过程。

2. Types of Machine Learning | 机器学习的类型

Machine learning is broadly divided into three main paradigms: supervised learning, unsupervised learning, and reinforcement learning. Each is defined by the nature of the data and the feedback available during training.

机器学习大致分为三种主要范式：监督学习、无监督学习和强化学习。它们的区别在于训练数据的性质以及训练过程中可用的反馈。

Supervised learning uses labelled data. The algorithm learns a mapping from inputs to outputs based on example input-output pairs.

监督学习使用带标签的数据。算法根据示例输入-输出对，学习从输入到输出的映射。
Unsupervised learning works with unlabeled data. The algorithm identifies hidden patterns, groupings, or structures without explicit guidance.

无监督学习处理无标签数据。算法在没有明确指导的情况下识别隐藏的模式、分组或结构。
Reinforcement learning involves an agent learning to make decisions by interacting with an environment. The agent receives rewards or penalties and aims to maximize cumulative reward.

强化学习涉及一个智能体通过与环境的交互来学习决策。智能体根据行为获得奖励或惩罚，并旨在最大化累积奖励。

Understanding these categories helps you identify which approach suits a given problem. For your exam, be ready to classify real-world scenarios like predicting house prices (supervised regression) or grouping customers (unsupervised clustering).

理解这些类别有助于针对特定问题选择合适的方法。在考试中，要准备好对真实场景进行分类，例如预测房价（监督回归）或客户分组（无监督聚类）。

3. Supervised Learning in Detail | 监督学习详解

Supervised learning tasks are further split into regression and classification. The distinction is based on the type of target variable.

监督学习任务进一步分为回归和分类。区别在于目标变量的类型。

Regression predicts a continuous numeric value. For example, forecasting temperature or a student’s final score.

回归预测连续的数值。例如，预测温度或学生的最终成绩。
Classification predicts a discrete category. For example, whether an email is ‘spam’ or ‘not spam’, or recognising a handwritten digit.

分类预测离散的类别。例如，判断电子邮件是“垃圾邮件”还是“非垃圾邮件”，或者识别手写数字。

Common supervised algorithms include linear regression, k-nearest neighbours (k-NN), decision trees, and support vector machines. Each has its strengths and assumptions.

常见的监督学习算法包括线性回归、k-最近邻（k-NN）、决策树和支持向量机。每种算法都有其优势和假设。

In the training phase, the model sees many examples with known labels. A loss function measures the difference between predictions and true labels, and the learning algorithm adjusts parameters to minimise this loss.

在训练阶段，模型会接触大量带有已知标签的示例。损失函数衡量预测值与真实标签之间的差异，学习算法则调整参数以最小化这种损失。

4. Unsupervised Learning and Clustering | 无监督学习与聚类

Unsupervised learning finds patterns in data without predefined labels. The most common technique is clustering, where the algorithm groups similar data points together.

无监督学习在没有预定义标签的数据中寻找模式。最常见的技术是聚类，即算法将相似的数据点归为一组。

K-means clustering is a classic example. It partitions data into k distinct clusters by minimising the distance between points and cluster centres. The algorithm iterates: assign points to the nearest centroid, then recompute centroids.

K-means 聚类是一个经典示例。它通过最小化数据点与聚类中心之间的距离，将数据划分为 k 个不同的簇。该算法迭代进行：将点分配到最近的质心，然后重新计算质心。

Another method is dimensionality reduction, such as principal component analysis (PCA), which simplifies data while retaining important variance. This helps visualise high-dimensional data and speeds up training.

另一种方法是降维，如主成分分析（PCA），它在简化数据的同时保留重要的方差。这有助于高维数据的可视化，并能加速训练。

Applications include market segmentation, anomaly detection, and recommendation systems (often combining unsupervised and supervised elements).

应用场景包括市场细分、异常检测以及推荐系统（常常结合了无监督和监督的元素）。

5. Reinforcement Learning | 强化学习

Reinforcement learning (RL) trains an agent to take actions in an environment to maximise a cumulative reward. Unlike supervised learning, there is no correct answer provided at each step; the agent learns through trial and error.

强化学习（RL）训练一个智能体在环境中采取行动，以最大化累积奖励。与监督学习不同，每一步没有提供正确答案；智能体通过试错来学习。

The key components of an RL problem are the agent, environment, state, action, reward, and policy. The policy maps states to actions and is continuously improved.

强化学习问题的关键组成部分包括智能体、环境、状态、动作、奖励和策略。策略将状态映射到动作，并不断改进。

Q-learning is a model-free algorithm that learns the value of taking an action in a given state. Deep reinforcement learning uses neural networks to approximate these value functions, enabling solutions to complex tasks like playing Go or controlling robots.

Q-learning 是一种免模型算法，用于学习在给定状态下采取某个动作的价值。深度强化学习使用神经网络来逼近这些价值函数，从而能够解决复杂任务，如下围棋或控制机器人。

IB and OCR courses may introduce RL as an application of AI. You should be able to explain the exploration vs exploitation trade-off: should the agent try new actions to discover potentially better rewards, or stick with actions known to yield high rewards?

IB和OCR课程可能将强化学习作为人工智能的应用进行介绍。你应当能够解释探索与利用的权衡：智能体应该尝试新动作以发现潜在更好的奖励，还是坚持已知能带来高回报的动作？

6. Training, Validation and Test Data | 训练、验证与测试数据

A crucial concept in machine learning is splitting data into three independent sets: training, validation, and test. This ensures the model generalises well to unseen data.

机器学习中的一个关键概念是将数据划分成三个独立的集合：训练集、验证集和测试集。这能确保模型对未见过的数据具有良好的泛化能力。

Training set: used to fit the model parameters.

训练集：用于拟合模型参数。
Validation set: used to tune hyperparameters and evaluate during training, helping to detect overfitting.

验证集：用于调整超参数并在训练过程中进行评估，有助于检测过拟合。
Test set: used only once at the very end to report the final performance. Never used in training or hyperparameter tuning.

测试集：仅在最终阶段使用一次，报告最终性能。绝不用于训练或超参数调优。

A typical split ratio is 70% training, 15% validation, 15% test. Cross-validation, such as k-fold cross-validation, is a technique where the training data is partitioned into k subsets and the model is trained and validated multiple times, each time using a different subset for validation.

典型的划分比例是 70% 训练、15% 验证、15% 测试。交叉验证，例如 k 折交叉验证，是一种将训练数据分成 k 个子集的技术，模型多次训练和验证，每次使用不同的子集作为验证。

For exams, explain why we need separate sets: to avoid optimistic bias and ensure the model works in the real world.

考试中，需解释为何需要独立的数据集：避免乐观偏差，确保模型在真实世界中能够正常工作。

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

Overfitting occurs when a model learns the training data too well, including its noise and outliers. It performs excellently on training data but poorly on new data because it fails to generalise.

过拟合发生于模型对训练数据学习过度，包括其中的噪声和异常值。它在训练数据上表现优秀，但在新数据上表现糟糕，因为无法泛化。

Underfitting happens when a model is too simple to capture the underlying structure of the data. It performs badly on both training and test data.

欠拟合发生在模型过于简单，无法捕捉数据的底层结构时。它在训练数据和测试数据上都表现很差。

We aim for a good fit where the model captures the real patterns without memorising noise. Strategies to combat overfitting include: reducing model complexity, adding more training data, using regularisation, and early stopping during training.

我们追求的是良好的拟合，即模型捕捉真实模式而不记忆噪声。对抗过拟合的策略包括：降低模型复杂度、增加更多训练数据、使用正则化以及在训练过程中提前停止。

Concept	Training Error	Test Error
Underfitting	High	High
Overfitting	Low	High
Good Fit	Low	Low

Both IB and OCR exams expect you to define these terms and suggest remedies. A common question is sketching learning curves showing training and validation error over epochs to diagnose overfitting.

IB和OCR考试都要求你定义这些术语并提出补救措施。常见题型是根据训练和验证误差随周期变化的学习曲线来诊断过拟合。

8. Evaluation Metrics | 评估指标

Evaluating a model’s performance goes beyond simple accuracy, especially for imbalanced datasets. You must understand precision, recall, F1-score, and confusion matrices.

评估模型性能不能仅看简单的准确率，尤其对于不平衡数据集。你必须理解精确率、召回率、F1 分数和混淆矩阵。

The confusion matrix is a table showing true positives (TP), true negatives (TN), false positives (FP), and false negatives (FN).

混淆矩阵是一个显示真正例（TP）、真负例（TN）、假正例（FP）和假负例（FN）的表格。

Accuracy = (TP + TN) / (total samples). It can be misleading if classes are skewed.

准确率 = (TP + TN) / (总样本数)。在类别分布不均衡时可能产生误导。
Precision = TP / (TP + FP). It answers: of all predicted positives, how many were truly positive?

精确率 = TP / (TP + FP)。它回答：在所有被预测为正例的样本中，有多少是真正的正例？
Recall = TP / (TP + FN). It answers: of all actual positives, how many did we capture?

召回率 = TP / (TP + FN)。它回答：在所有实际正例中，我们捕捉到了多少？
F1-score = 2 × (Precision × Recall) / (Precision + Recall). It is the harmonic mean, balancing precision and recall.

F1 分数 = 2 × (精确率 × 召回率) / (精确率 + 召回率)。它是调和平均值，平衡了精确率和召回率。

For regression tasks, common metrics are mean squared error (MSE) and root mean squared error (RMSE).

对于回归任务，常用指标为均方误差（MSE）和均方根误差（RMSE）。

OCR and IB exam questions may provide a confusion matrix and ask you to compute these values. Know the trade-off between precision and recall.

OCR和IB考试题目可能会提供一个混淆矩阵，要求你计算这些值。需了解精确率与召回率之间的权衡。

9. Feature Engineering | 特征工程

Feature engineering is the process of selecting, transforming, and creating variables (features) that help a machine learning model learn more effectively. The quality of features often matters more than the choice of algorithm.

特征工程是选择、转换并创建变量（特征）的过程，以帮助机器学习模型更有效地学习。特征的质量往往比算法的选择更为重要。

Steps include handling missing values, encoding categorical variables (e.g., one-hot encoding), scaling numerical features to a similar range, and creating interaction features. Normalisation (scaling to 0-1 range) and standardisation (scaling to zero mean, unit variance) are typical techniques.

步骤包括处理缺失值、对分类变量进行编码（例如独热编码）、将数值特征缩放到相似范围，以及创建交互特征。归一化（缩放到 0-1 范围）和标准化（缩放到零均值、单位方差）是典型的技术。

Dimensionality reduction techniques like PCA also form part of feature extraction. Feature selection removes irrelevant or redundant features to improve model performance and reduce overfitting.

像 PCA 这样的降维技术也属于特征提取的一部分。特征选择移除不相关或冗余的特征，以提升模型性能、减少过拟合。

Good feature engineering can make a simple linear model outperform a complex neural network on tabular data. Your syllabus likely expects you to discuss why data preprocessing is critical.

优秀的特征工程可以让简单的线性模型在表格数据上超越复杂的神经网络。大纲可能要求你讨论为什么数据预处理至关重要。

10. Common Algorithms: k-NN and Decision Trees | 常见算法：k-最近邻与决策树

The k-nearest neighbours (k-NN) algorithm is a simple, instance-based method used for classification and regression. It stores all training examples and predicts based on the majority vote (classification) or average (regression) of its k nearest neighbours.

k-最近邻（k-NN）算法是一种简单的、基于实例的方法，用于分类和回归。它储存所有训练样本，并根据其 k 个最近邻的多数投票（分类）或平均值（回归）进行预测。

Choosing k is critical. Small k (e.g., 1) leads to overfitting because the model is too sensitive to noise. Large k smooths the decision boundary but may underfit. Euclidean distance is commonly used to measure closeness.

选择 k 值至关重要。k 值小（如 1）会导致过拟合，因为模型对噪声过于敏感。k 值大会使决策边界平滑，但可能欠拟合。欧氏距离通常用于测量邻近程度。

Decision trees split data based on feature values to create a tree-like model of decisions. Each internal node tests an attribute, each branch represents an outcome, and leaf nodes hold class labels or continuous values.

决策树基于特征值对数据进行划分，创建树状的决策模型。每个内部节点测试一个属性，每个分支代表一个结果，而叶节点包含类别标签或连续值。

The splitting criterion aims to maximise information gain or minimise impurity (Gini impurity, entropy). Decision trees are interpretable but prone to overfitting; pruning and setting a maximum depth help control complexity.

划分标准旨在最大化信息增益或最小化不纯度（基尼不纯度、熵）。决策树可解释性强，但容易过拟合；剪枝和设置最大深度有助于控制复杂度。

Both algorithms appear in IB and OCR specifications. You should compare their strengths: k-NN is lazy (no training phase) and scales poorly with large datasets; trees are fast to train but can be unstable.

这两种算法都出现在IB和OCR大纲中。你需要比较它们的优势：k-NN 是惰性学习（没有训练阶段），在大数据集上扩展性差；树训练速度快，但可能不稳定。

11. Neural Networks and Deep Learning | 神经网络与深度学习

Neural networks are computing systems inspired by biological neural networks. An artificial neuron receives inputs, applies weights and a bias, passes the sum through an activation function, and produces an output.

神经网络是受生物神经网络启发的计算系统。人工神经元接收输入，应用权重和偏置，将总和通过激活函数传递，然后产生输出。

The learning process uses backpropagation to compute gradients of the loss with respect to weights, and an optimiser (e.g., gradient descent) updates the weights to minimise error.

学习过程利用反向传播计算损失相对于权重的梯度，并使用优化器（例如梯度下降）更新权重以最小化误差。

Common activation functions include sigmoid, tanh, and ReLU (Rectified Linear Unit). ReLU, defined as f(x) = max(0, x), helps alleviate the vanishing gradient problem in deep networks.

常见的激活函数包括 sigmoid、tanh 和 ReLU（修正线性单元）。ReLU 定义为 f(x) = max(0, x)，有助于缓解深层网络中的梯度消失问题。

Deep learning refers to neural networks with many hidden layers. These models excel at tasks involving unstructured data like images (convolutional neural networks, CNNs) and sequences (recurrent neural networks, RNNs and Transformers).

深度学习指具有许多隐藏层的神经网络。这些模型在处理非结构化数据方面表现出色，如图像（卷积神经网络，CNNs）和序列（循环神经网络，RNNs 和 Transformers）。

For IB and OCR, focus on the basic structure, the role of weights, activation functions, and the general idea of training via error minimisation. You are not expected to code a full neural network but can explain its components.

针对IB和OCR，重点在于基本结构、权重的作用、激活函数，以及通过最小化误差进行训练的整体思想。不要求编写完整的神经网络代码，但需要能解释其组成部分。

12. Bias-Variance Trade-off | 偏差-方差权衡

The bias-variance trade-off is a fundamental concept explaining the error of a learning algorithm. Total prediction error can be decomposed into bias, variance, and irreducible error.

偏差-方差权衡是解释学习算法误差的一个基本概念。总预测误差可以分解为偏差、方差和不可约误差。

Bias is the error due to overly simplistic assumptions. High bias can cause underfitting.

偏差是由于模型假设过于简单而产生的误差。高偏差可能导致欠拟合。
Variance is the error due to sensitivity to small fluctuations in the training set. High variance can cause overfitting.

方差是因为模型对训练集中微小波动的敏感性而产生的误差。高方差可能导致过拟合。

Simplifying a model reduces variance but increases bias. Complexifying a model reduces bias but increases variance. The goal is to find a sweet spot that minimises total error.

简化模型会降低方差，但会增加偏差。使模型更复杂会降低偏差，但会增加方差。目标是找到最小化总误差的最佳平衡点。

Models like decision trees with unlimited depth have low bias and high variance. Linear regression has high bias and low variance. Ensemble methods like random forests and gradient boosting combine multiple weak learners to balance bias and variance.

像深度不受限制的决策树，偏差低、方差高。线性回归偏差高、方差低。集成方法如随机森林和梯度提升，组合了多个弱学习器以平衡偏差和方差。

Your exam may ask you to relate this trade-off to underfitting/overfitting and to justify algorithm choices.

考试可能要求你将这种权衡与欠拟合/过拟合联系起来，并证明算法选择的合理性。

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

更多咨询请联系16621398022（同微信）

Key Concepts in Machine Learning for IB & OCR Computer Science | IB与OCR计算机科学机器学习考点精讲