机器学习：中英双语指南

Validation status / 复核结论： The original explanation was broadly correct and well structured. This edition corrects several boundary cases, sharpens algorithm-selection guidance, and distinguishes learning paradigms, task types, model families, and deployment requirements.
原说明整体正确、结构完整。本修订版进一步纠正边界问题，并明确区分学习范式、任务类型、模型家族及生产部署条件。

Machine learning should not be organized as one flat list. Four different layers are involved:

机器学习不能被理解成一个平铺的算法清单，而应分成四个层次：

1. Learning paradigm / 学习范式 — how the model receives feedback.
2. Task type / 任务类型 — what output the model must produce.
3. Algorithm or model family / 算法或模型家族 — how the relationship is learned.
4. Operational system / 生产系统 — how the model is evaluated, deployed, monitored, and used.

The basic pipeline is:

Business question → Learning paradigm → Task → Algorithm → Evaluation → Deployment → Monitoring
业务问题 → 学习范式 → 任务类型 → 算法选择 → 评估 → 部署 → 监控

Examples:

• Supervised learning / 监督学习 is a learning paradigm.
• Classification / 分类 is a task.
• Random forest / 随机森林 is an algorithm family.
• F1, calibration, latency, and drift monitoring / F1、概率校准、延迟和漂移监控 are evaluation and operational concerns.

One algorithm can solve several tasks. Random forests support classification and regression. Neural networks can support classification, regression, generation, representation learning, and reinforcement learning.

同一算法可以解决多种任务。例如随机森林可以分类，也可以回归；神经网络可以用于分类、回归、生成、表示学习及强化学习。

These categories can overlap. A system may use self-supervised pretraining, supervised fine-tuning, active labeling, and online monitoring in the same lifecycle.

这些范式并非互斥。一个系统可以先进行自监督预训练，再做监督微调，同时采用主动标注，并在部署后进行在线监控。

Supervised data contain input features (X) and known targets (y):

监督学习数据包含特征 (X) 和已知目标 (y)：

[ f(X) \rightarrow y ]

Classification predicts discrete categories or class probabilities.
分类用于预测离散类别或类别概率。

For a contract-win model, precision matters when bids are expensive and resources must be concentrated; recall matters when missing a viable opportunity is costly. If the model reports “70% chance of winning,” probability calibration must also be tested.

对于合同中标模型，如果投标成本高、资源需要集中，Precision更重要；如果漏掉可行机会的代价高，Recall更重要。如果模型输出“70%中标概率”，还必须检验概率校准。

Regression predicts a continuous number, such as cost, demand, duration, or contract value.
回归预测连续数值，例如成本、需求量、工期或合同金额。

Common metrics include MAE, MSE, RMSE, (R^2), MAPE, and pinball loss. MAPE becomes unstable when actual values are zero or close to zero. (R^2) alone does not prove accuracy, causality, or operational usefulness.

常见指标包括MAE、MSE、RMSE、(R^2)、MAPE及Pinball Loss。真实值为零或接近零时，MAPE会失真；单独的(R^2)不能证明模型准确、具有因果性或具备业务价值。

Ranking predicts relative order rather than a standalone category. It is useful for contract-opportunity prioritization, search, recommendations, candidate retrieval, and risk triage.

排序学习预测对象的相对顺序，而不是单独类别。适用于合同机会优先级、搜索、推荐、候选对象检索及风险排序。

Common models include RankNet, LambdaRank, LambdaMART, pairwise ranking, and boosted-tree ranking. Training requires query/user/task groups and relevance or preference labels. Evaluate with NDCG, MAP, or MRR. Historical click data require correction for exposure and position bias.

常用模型包括RankNet、LambdaRank、LambdaMART、成对排序及提升树排序。训练需要查询、用户或任务分组，以及相关性或偏好标签。通常使用NDCG、MAP或MRR评估；历史点击数据还要控制曝光偏差和位置偏差。

Time order must be preserved. Typical applications include budget execution, expenditure, inventory, demand, staffing, and server-load forecasts.

时间顺序必须保留。典型应用包括预算执行、支出、库存、需求、人员及服务器负载预测。

Use walk-forward or rolling-origin validation. A random row split usually leaks temporal information. Every feature must be available at the actual forecast time. More complex deep models do not automatically beat seasonal baselines or boosted trees.

应使用Walk-forward或Rolling-origin验证。随机拆分行通常造成时间泄漏。所有特征必须在真实预测时点可获得。复杂深度模型不会自动优于季节基线或提升树。

• Survival analysis / 生存分析: predicts time to an event while handling censored observations; examples include equipment failure, attrition, and project delay. Models include Cox regression and random survival forests.
• Structured prediction / 结构化预测: produces sequences or structured outputs, such as named entities, translations, speech transcripts, image segmentation, or document fields. Models include CRFs, transformers, and U-Net-style architectures.
• Count modeling / 计数建模: predicts event counts with Poisson, negative-binomial, or related models when ordinary regression assumptions are unsuitable.

Unsupervised learning discovers structure without human-provided targets. Results generally require stronger domain interpretation because there is no definitive label-based accuracy score.

无监督学习在没有人工目标标签的情况下发现结构。由于不存在天然的标签准确率，结果通常更依赖领域解释。

Silhouette, Davies–Bouldin, and Calinski–Harabasz scores assess geometric properties, not whether the clusters are real business segments. Validate stability, interpretability, reproducibility, and downstream usefulness.

Silhouette、Davies–Bouldin及Calinski–Harabasz衡量的是几何结构，不能证明群组是真实业务分群。还必须验证稳定性、可解释性、可重复性及下游价值。

Anomaly detection is a task, not exclusively an unsupervised paradigm. It may be:

异常检测是一种任务，并不专属于无监督学习。它可以是：

• supervised, when reliable anomaly labels exist;
• semi-supervised or one-class, when mostly normal examples are available;
• unsupervised, when labels do not exist.

An unusual record is not automatically fraud, error, or wrongdoing. The model flags deviation; investigation establishes meaning.

异常记录并不自动等于欺诈、错误或违规。模型只负责发现偏离，业务调查才能确定含义。

Association-rule methods such as Apriori and FP-Growth identify frequent co-occurrence. Key measures are support, confidence, and lift:

Apriori、FP-Growth等关联规则方法识别高频共现。关键指标包括支持度、置信度和提升度：

[ Support(A\rightarrow B)=P(A\cap B) ]

[ Confidence(A\rightarrow B)=P(B|A) ]

[ Lift(A\rightarrow B)=\frac{P(B|A)}{P(B)} ]

Association is not causation. High-confidence rules can simply reflect a very common item.

关联不等于因果。高置信度规则可能只是因为某项目本身非常常见。

Topic discovery methods include LDA, NMF, BERTopic, and embedding-plus-clustering pipelines. Good tokenization, sufficient documents, multilingual support, stable topics, and expert interpretation are required.

主题发现可采用LDA、NMF、BERTopic或“嵌入＋聚类”。需要良好分词、足够文档、多语言支持、稳定主题及专家解释。

Use it when labels are expensive but unlabeled data are abundant and come from a distribution similar to deployment data. Methods include pseudo-labeling, self-training, label propagation, consistency regularization, FixMatch, and MixMatch.

当标签昂贵、无标签数据丰富且与部署数据分布相近时，可使用半监督学习。常用方法包括伪标签、自训练、标签传播、一致性正则化、FixMatch及MixMatch。

Success requires:

1. A representative labeled seed set / 有代表性的初始标签集。
2. A fully human-labeled holdout test set / 完全由人工标注的独立测试集。
3. Confidence thresholds and class-balance checks / 置信度阈值与类别平衡检查。
4. Monitoring for confirmation bias / 监控确认偏差。
5. Similarity between unlabeled and deployment distributions / 无标签数据与部署数据分布相似。

Self-supervised learning creates targets from raw data: next-token prediction, masked-token reconstruction, contrastive learning, masked-image reconstruction, or speech-segment prediction. It powers much of modern LLM, vision, and speech pretraining.

自监督学习从原始数据自动构造训练目标，例如下一Token预测、遮挡Token重建、对比学习、遮挡图像重建或语音片段预测，是现代LLM、视觉及语音预训练的重要基础。

Effective pretraining requires broad, representative, cleaned data; a suitable objective; sufficient model capacity and compute; deduplication; and privacy, copyright, bias, and safety controls. The pretrained representation still needs task-specific evaluation and often prompting, retrieval, adapters, or fine-tuning.

有效预训练需要广泛、代表性强且经过清理的数据，合适的训练目标，足够的模型容量与算力，数据去重，以及隐私、版权、偏见和安全控制。预训练后仍需进行具体任务评估，并可能配合提示、检索、适配器或微调。

Reinforcement learning learns a policy through interaction. It is most appropriate when actions affect future states and rewards—not merely when a labeled prediction is needed.

强化学习通过与环境交互学习策略。只有当行动会影响未来状态和奖励时才特别适合；如果只需要有标签预测，通常无需强化学习。

Core concepts are agent, environment, state, action, reward, policy, value, exploration, and exploitation. The objective is commonly a discounted return:

核心概念包括智能体、环境、状态、动作、奖励、策略、价值、探索和利用。目标通常是最大化折扣累计回报：

[ G_t=\sum_{k=0}^{\infty}\gamma^k r_{t+k+1} ]

Effective RL needs a well-defined state/action/reward system, adequate exploration or representative offline coverage, safe experimentation or a realistic simulator, reliable reward attribution, and guardrails against reward hacking. In high-stakes domains, off-policy evaluation, uncertainty analysis, and human oversight are essential.

有效强化学习需要清晰定义状态、动作和奖励，需要充分探索或有代表性的离线覆盖，需要安全试验环境或逼真模拟器，并需要可靠的奖励归因及防止奖励投机。在高风险领域，还必须进行离线策略评估、不确定性分析及人工监督。

• Feature extraction / 特征提取: freeze a pretrained encoder and train a smaller task model.
• Prompting / 提示: use existing model capability without changing weights.
• RAG / 检索增强生成: retrieve current or internal knowledge without encoding all facts into weights.
• LoRA/QLoRA: parameter-efficient adaptation.
• Full fine-tuning / 全量微调: update most or all weights when enough data, compute, and justification exist.

If the problem is missing or changing knowledge, RAG is often preferable. If the problem is persistent behavior, format, terminology, or domain adaptation, fine-tuning may help. These approaches can be combined.

如果问题是缺少知识或知识经常变化，通常优先考虑RAG；如果问题是稳定的行为、格式、术语或领域适配，微调可能更有效。两者可以组合使用。

The model selects samples whose labels would be most informative. Use uncertainty sampling, query-by-committee, diversity sampling, or core-set selection. It works only when an expert labeling loop exists. Uncertainty alone may repeatedly select noise, so representativeness and diversity are also needed.

模型选择最有信息价值的样本请求人工标注。可以采用不确定性采样、委员会查询、多样性采样或核心集选择。它要求存在专家标注闭环；仅选择最不确定样本可能不断抽到噪声，因此还需考虑代表性和多样性。

Online models update as data arrive. They require delayed-label tracking, drift detection, versioning, rollback, poisoning defenses, and clear retraining rules. Not every production model needs continuous weight updates; periodic batch retraining is often safer.

在线模型随数据到达持续更新，需要跟踪延迟标签、检测漂移、管理版本与回滚、防止数据投毒并制定重训规则。并非所有生产模型都需要实时更新权重；定期批量重训通常更安全。

Federated learning keeps raw data distributed while sharing model updates. It does not automatically guarantee privacy. Secure aggregation, differential privacy, identity/access control, participant trust, communication efficiency, and non-IID data handling may still be required.

联邦学习让原始数据保持分散，只共享模型更新，但它并不会自动保证隐私。仍可能需要安全聚合、差分隐私、身份与访问控制、参与方信任、通信效率及非独立同分布数据处理。

Graph methods are useful when relationships carry information: supplier–contract–agency networks, transaction graphs, supply chains, recommendations, knowledge graphs, and cyberattack paths.

当关系本身包含重要信息时，应考虑图机器学习，例如供应商—合同—机构网络、交易网络、供应链、推荐、知识图谱及网络攻击路径。

Tasks include node classification, link prediction, community detection, graph classification, and knowledge-graph completion. Models include PageRank, Node2Vec, GraphSAGE, GCN, GAT, and graph transformers. Accurate entity resolution and leakage-safe temporal or structural splits are crucial.

任务包括节点分类、链接预测、社区发现、整图分类及知识图谱补全。模型包括PageRank、Node2Vec、GraphSAGE、GCN、GAT及图Transformer。准确的实体解析以及避免泄漏的时间或结构拆分非常关键。

Generative models learn aspects of a data distribution and generate text, code, images, audio, video, or synthetic records.

生成模型学习数据分布的某些特征，并生成文本、代码、图像、音频、视频或合成记录。

Evaluation must be task-specific. Fluency or visual quality alone does not establish factuality, usefulness, safety, privacy, or absence of bias. For deployed generative systems, test task success, groundedness, hallucination rate, safety, latency, cost, and human acceptance.

评估必须针对具体任务。语言流畅或视觉质量高并不能证明事实准确、有用、安全、保护隐私或不存在偏见。生产生成系统还应测试任务成功率、依据性、幻觉率、安全性、延迟、成本及用户接受度。

Clarify the prediction target, prediction time, user, downstream action, and costs of false positives and false negatives. A technically accurate model can still be useless if nobody can act on it.

必须明确预测目标、预测时间、使用者、后续行动以及误报和漏报成本。即使技术准确，如果无人能够据此行动，模型仍然没有价值。

Check label definitions, annotator agreement, missing labels, delayed outcomes, historical policy effects, and whether the label actually represents the business objective. Label noise often limits performance more than algorithm choice.

检查标签定义、标注者一致性、缺失标签、延迟结果、历史政策影响，以及标签是否真正代表业务目标。标签噪声对性能的限制往往大于算法选择。

Training data should cover the future population, periods, agencies, regions, classes, edge cases, and policy or market regimes. More data do not fix systematic sampling bias.

训练数据应覆盖未来人群、时间段、机构、地区、类别、边缘案例以及政策或市场环境。增加数据量无法修复系统性采样偏差。

There is no universal sample-size rule. Requirements depend on dimensionality, noise, model capacity, effect size, class rarity, desired confidence, and use of pretrained representations. Learning curves and uncertainty intervals are more defensible than a fixed “samples per feature” rule.

不存在通用样本量规则。需求取决于维度、噪声、模型容量、效应大小、少数类稀有程度、置信要求及是否采用预训练表示。学习曲线和不确定性区间比固定的“每特征多少样本”规则更可靠。

Every feature must exist at prediction time. Leakage occurs when future outcomes, post-decision fields, duplicates, group identity, or preprocessing information crosses the train/test boundary.

所有特征必须在真实预测时点可获得。如果未来结果、决策后字段、重复记录、群组身份或预处理信息跨越训练/测试边界，就会发生数据泄漏。

Fit scalers, imputers, feature selection, resampling, and target encoding inside each training fold—not on the complete dataset.

缩放、缺失填补、特征选择、重采样和目标编码必须在每个训练折内部拟合，不能先在完整数据集上完成。

Use class weights, appropriate sampling, focal loss, anomaly methods, or threshold adjustment where justified. SMOTE is not automatically beneficial and must be applied only inside training folds. Preserve realistic prevalence in validation/test data when estimating operational performance.

可以根据需要使用类别权重、适当采样、Focal Loss、异常检测或阈值调整。SMOTE并非自动有效，而且只能在训练折内部使用。验证集和测试集应尽量保持真实发生率，以估计实际表现。

• Scaling is important for SVM, KNN, PCA, regularized linear models, and neural networks.

• Tree models usually do not require scaling.

• Missingness may itself carry information, but handling must reflect deployment.

• Categorical encoding must avoid target leakage.

• Text, image, and audio augmentation must preserve the label.

• SVM、KNN、PCA、正则化线性模型及神经网络通常需要缩放。

• 树模型通常不需要缩放。

• 缺失本身可能包含信息，但处理方式必须与部署一致。

• 类别编码必须防止目标泄漏。

• 文本、图像及音频增强不能改变真实标签。

Underfitting means poor training and validation performance. Overfitting means training performance is much better than performance on unseen data. Control complexity through regularization, pruning, early stopping, data augmentation, feature reduction, more representative data, or a simpler model.

欠拟合表现为训练和验证性能都差；过拟合表现为训练性能远好于未见数据。可以通过正则化、剪枝、早停、数据增强、减少特征、增加代表性数据或使用更简单模型来控制复杂度。

Probability calibration asks whether events predicted at 70% occur about 70% of the time. Use reliability diagrams, expected calibration error, Brier score, Platt scaling, isotonic regression, or temperature scaling as appropriate. Also quantify uncertainty from limited data, model instability, and distribution shift.

概率校准检查“预测70%的事件是否大约70%会发生”。可以使用可靠性图、期望校准误差、Brier分数、Platt缩放、等距回归或温度缩放。同时还要量化小样本、模型不稳定及分布变化带来的不确定性。

Feature importance, SHAP, LIME, and partial-dependence plots explain model behavior, not causal effects. If the question is “What action will change the outcome?”, causal inference or experimentation may be required.

特征重要性、SHAP、LIME及部分依赖图解释的是模型行为，而不是因果效应。如果问题是“采取什么行动会改变结果”，可能需要因果推断或实验。

Fairness evaluation should compare relevant error rates, calibration, access, and downstream impact across legally and operationally meaningful groups. Fairness definitions can conflict, so the correct criterion depends on context and policy.

公平性评估应比较具有法律及业务意义群体之间的错误率、校准、机会获得及下游影响。不同公平标准可能冲突，因此必须结合具体环境和政策选择。

Monitor input quality, missingness, prediction distributions, latency, cost, real outcomes, subgroup performance, data drift, concept drift, and feedback effects. Define alert thresholds, ownership, retraining triggers, rollback plans, and model retirement criteria before deployment.

应监控输入质量、缺失情况、预测分布、延迟、成本、真实结果、群体表现、数据漂移、概念漂移及反馈效应。在部署前就应明确告警阈值、责任人、重训触发条件、回滚方案及模型退役标准。

1. Define the decision and error costs. / 定义决策及错误成本。
2. Audit labels, sampling, timing, and leakage. / 审查标签、抽样、时间及泄漏。
3. Build a naive baseline. / 建立朴素基线。
4. Train a simple interpretable model. / 训练简单且可解释的模型。
5. Add a strong task-appropriate model. / 增加适合任务的强模型。
6. Use cross-validation or backtesting that matches deployment. / 使用与部署环境一致的交叉验证或回测。
7. Compare accuracy, calibration, stability, fairness, latency, and cost. / 比较准确性、校准、稳定性、公平性、延迟及成本。
8. Pilot inside the real workflow. / 在真实工作流程中试点。
9. Monitor outcomes and feedback loops. / 监控真实结果及反馈闭环。
10. Retrain, revise, or retire under predefined rules. / 按预定规则重训、修订或退役。

For most contract, financial, operational, and government tabular problems, begin with logistic/linear regression and boosted trees. Deep learning becomes compelling when raw text, image, audio, graph, massive-scale, or representation-learning requirements justify the added complexity.

对于大多数合同、财务、运营及政府表格数据问题，应优先从逻辑/线性回归及提升树开始。当任务涉及原始文本、图像、音频、图结构、超大规模数据或表示学习时，深度学习的额外复杂性才更有价值。

The original framework was fundamentally sound. The crucial refinements are:

原框架基本正确，关键修订如下：

1. Learning paradigms, tasks, and algorithms are different layers. / 学习范式、任务和算法属于不同层次。
2. Paradigms can overlap within one system lifecycle. / 一个系统生命周期可以同时采用多种范式。
3. Anomaly detection is not exclusively unsupervised. / 异常检测并不专属于无监督学习。
4. Time-series validation must preserve time. / 时间序列验证必须保持时间顺序。
5. Clustering scores do not prove business truth. / 聚类指标不能证明业务分群真实。
6. No universal sample-size rule exists. / 不存在通用样本量规则。
7. Explainability does not establish causality. / 可解释性不等于因果性。
8. Deployment quality depends as much on data, validation, workflow, and monitoring as on the algorithm. / 生产质量不仅取决于算法，也同样取决于数据、验证、工作流程及监控。

The practical formula is:

实用公式是：

[ Useful\ ML = Good\ Problem + Representative\ Data + Reliable\ Labels + Proper\ Validation + Operational\ Integration ]

[ 有效机器学习 = 正确问题 + 代表性数据 + 可靠标签 + 正确验证 + 业务流程集成 ]

Source: Machine_Learning_Bilingual_Validated_Guide.md