所属模块：
M7 · 检索、排序与推荐系统 (Retrieval, Ranking & RecSys)| 专题分类：多目标与约束 (Multi-Objective Ranking & Optimization)| 难度等级：Easy
一、核心一句话结论 (One-Sentence Summary)
加权和（简单）、乘法/排序公式（’都必须满足’）、约束优化（主目标 + 约束）、多任务模型 + 融合层。
Multi-objective systems simultaneously optimize competing metrics (CTR, CVR, dwell time, long-term retention) through weighted linear scalarization, multiplicative ranking formulas, constrained optimization, and multi-task neural architectures (MMoE/PLE).
二、核心考点要义 (Key Insights)
- 📌 加权和：简单但权重难调、目标可互补
- 📌 乘法/排序公式：表示’都必须满足’（如 CTR×CVR）
- 📌 约束优化：主目标 + 其他目标不低于阈值
- 📌 多任务模型（MMoE）+ 融合层：共享底层、各自预测、再组合
English Insights:
– The multi-objective challenge: Direct business objectives (e.g., clicks vs. revenue vs. user satisfaction) frequently exhibit negative correlation and conflicting gradients.
– Methodology spectrum: Spans linear weighted sums, multiplicative value formulas (eCPM), constrained optimization (Lagrangian relaxation), and Pareto ranking.
– Architectural foundation: Deploys multi-task networks (MMoE, PLE) to share representations while decoupling task-specific output predictions.
三、核心数学原理与机理推导 (Mathematical Principles & Derivation)
$$text{weighted}: sum w_ihat y_i;qquad text{product}: prod hat y_i;qquad text{constrained}: max y_1 text{s.t.} y_2ge c$$
数学机理：四类做法——(1) 加权和（weighted sum）——score=Σ w_i·ŷ_i；语义——各目标可互补（一个高可以弥补另一个低）；优点——简单、可导、易优化；缺点——(a) 权重难调（不同目标量纲不同，需归一化）；(b) 无法表达’必须满足’（某目标为 0 时仍可能被其他目标’救回’）；(c) 帕累托前沿的’凹部’无法达到（加权和只能达到帕累托前沿的凸部分）。(2) 乘法/排序公式——score=Π ŷ_i 或 ŷ_1·ŷ_2·…；语义——‘都必须满足’（任一为 0 则总分为 0）；例——电商的 ‘pCTR × pCVR × price’（必须点击且转化）；优点——表达’必要条件’；缺点——(a) 对’小概率’敏感（一个很小的概率会主导）；(b) 需各目标校准（因为乘积依赖概率的绝对值）。(3) 约束优化（constrained optimization）——max 主目标 s.t. 其他目标 ≥ 阈值；实现——(a) 拉格朗日松弛（把约束加入目标：max y₁ − λ·max(0, c−y₂)）；(b) 分阶段（先满足约束、再优化主目标）；(c) 投影（每步把解投影到约束集）；优点——语义明确（’在主目标最优的前提下满足约束’）；缺点——实现复杂、λ 需调。(4) 多任务模型 + 融合层——(a) MMoE/PLE 输出多个目标的预测；(b) 再用’融合公式’（加权/乘法/规则）组合；优点——共享底层（互相促进）、融合灵活；缺点——融合仍需设计。其他做法——(a) 帕累托多目标优化（直接搜索帕累托前沿）；(b) 强化学习（把长期目标作为奖励）；(c) 规则层（在重排阶段强制约束）。选择依据——(a) 目标可互补 → 加权和；(b) 目标必须同时满足（如’点击且转化’） → 乘法；(c) 有硬性下限（如’多样性 ≥ X’） → 约束优化或规则；(d) 多目标 + 复杂交互 → 多任务模型 + 融合。权重/参数的调优——(a) 在线 A/B（离线不可靠）；(b) 帕累托前沿可视化（帮助决策）；(c) 按场景/用户分组（不同群体不同权重）；(d) 动态调整（如新品期提升探索权重）。实践建议——(a) 明确每个目标的语义（互补还是必需）；(b) 乘法用于’必要条件’、加权用于’可互补’；(c) 硬约束用规则层（重排阶段）；(d) 权重用在线实验调；(e) 监控各目标（防此消彼长）。度量——(a) 各目标的指标；(b) 端到端业务指标；(c) 帕累托前沿；(d) 约束的满足率。
📖 查看英文严格数学推导 (English Mathematical Derivation)
Mathematical & Optimization Formulation: Four Multi-Objective Paradigms.
(1) Weighted Linear Sum (Scalarization):
$$text{Score}(u, i) = sum_{k=1}^K w_k cdot hat{y}_k(u, i), quad w_k > 0, , sum w_k = 1$$
– Semantics: Additive compensation (a low score on Task 1 can be compensated by an exceptionally high score on Task 2).
– Limitation: Cannot find solutions on non-convex regions of the Pareto frontier, and hyperparameter tuning of $w_k$ is brittle across shifting traffic.
(2) Multiplicative / Value Ranking Formulas (e.g., Ad Auctions):
$$text{Score}(u, i) = ptext{CTR}(u, i)^alpha times ptext{CVR}(u, i)^beta times text{Bid}_i + gamma cdot text{Quality}(i)$$
– Semantics: Conjunctive requirement (both click and conversion are strictly necessary; if either $ptext{CTR} = 0$ or $ptext{CVR} = 0$, the commercial valuation drops to zero).
(3) Constrained Optimization via Lagrangian Relaxation:
Maximize primary business revenue while enforcing hard constraints on secondary ecosystem guardrails:
$$max_theta mathbb{E}[text{Revenue}(theta)] quad text{s.t.} quad mathbb{E}[text{DwellTime}(theta)] ge T_0, quad mathbb{E}[text{Unsubscribe}(theta)] le epsilon_0$$
Formulate Lagrangian with dual multipliers $lambda_1, lambda_2 ge 0$:
$$mathcal{L}(theta, lambda) = mathbb{E}[text{Revenue}(theta)] + lambda_1 big( mathbb{E}[text{DwellTime}(theta)] – T_0 big) – lambda_2 big( mathbb{E}[text{Unsubscribe}(theta)] – epsilon_0 big)$$
Dual variables $lambda$ are updated dynamically online using subgradient descent to satisfy constraints automatically.
(4) Multi-Task Network Integration (MMoE / PLE):
Shared expert representations route via task-specific gates to output calibrated predictions $(hat{y}_{text{CTR}}, hat{y}_{text{CVR}}, hat{y}_{text{dwell}})$, feeding directly into the downstream value fusion layer.
四、工业级落地权衡与工程考量 (Industrial Trade-offs)
深度剖析与工程权衡：① ‘加权和 vs 乘法的语义差异’是关键——加权表示’可互补’、乘法表示’都必须满足’；面试中能指出这一点是深度理解的标志。② ‘加权和无法达到帕累托前沿的凹部’——这是多目标优化的理论限制（凸包限制）；故有帕累托方法。③ ‘乘法需校准’——因为乘积依赖概率绝对值；故需各目标校准良好。④ ‘硬约束放规则层’——如’多样性/合规’不适合放进排序模型（会扭曲 CTR 目标）；放重排阶段更清晰。⑤ ‘权重用在线实验调’——离线调权重不可靠（与在线行为有差距）。⑥ 面试要点——被问’多目标怎么优化’，应给出’四类做法（加权/乘法/约束/多任务 + 融合）+ 语义差异 + 权重在线调 + 硬约束放规则层‘；能指出’加权和无法达到帕累托凹部’是深度理解的标志。
⚙️ 查看英文落地权衡分析 (English Systems & Trade-offs)
In-Depth Analysis & Engineering Trade-offs: ① The linear sum calibration failure—linear weighting ($w_1 hat{y}_1 + w_2 hat{y}_2$) breaks down if scores occupy different numerical distributions (e.g., CTR $in [0.01, 0.15]$ vs. dwell time $in [0, 600text{ s}]$); raw outputs must undergo quantile transformation, z-score standardization, or probability calibration before linear combination. ② Lagrangian relaxation vs. manual weight tuning—manual grid-searching of weights takes weeks of A/B testing and breaks whenever seasonal traffic shifts; Lagrangian relaxation dynamically raises $lambda_1$ if dwell time drops below SLA, automatically adjusting weights in real time. ③ Pareto dominance vs. Scalarization—in offline evaluation, models that achieve Pareto improvements (improving at least one metric without degrading any other) are prioritized; if metrics trade off, executive business strategy dictates the acceptable substitution rate. ④ Multi-task loss balancing during training—multi-task networks fail if one task’s loss gradient dominates the backward pass; deploying GradNorm or dynamic weight averaging ensures equal learning velocity across all prediction heads. ⑤ Latency overhead of multi-objective fusion—evaluating 5 task heads simultaneously on a GPU tensor adds $< 1text{ ms}$; the primary compute cost lies in the shared trunk, making multi-task architectures highly latency-efficient. ⑥ Interview takeaway—structure multi-objective approaches into linear sum, multiplicative formulas, constrained Lagrangian optimization, and multi-task neural modeling, explain why probability calibration is mandatory, and detail automated Lagrangian multiplier updates.
五、常见面试避坑陷阱 (Common Pitfalls & Traps)
- ⚠️ 用加权和表达’必须同时满足’（应为乘法）
- ⚠️ 把硬约束塞进排序模型（扭曲主目标）
English Pitfalls:
– Directly summing unnormalized predictions with disparate physical units (e.g., probability of click + video watch seconds), allowing the larger-magnitude variable to dominate.
– Using linear scalarization when all objectives must be satisfied simultaneously, failing to utilize multiplicative value formulations.
– Hardcoding static linear weights w_k across disparate seasons and promotions without automated Lagrangian constraint adjustments.
六、高频深度面试追问与预测 (Follow-Up Questions)
- 加权和 vs 乘法的语义差异？
- How does online subgradient descent update Lagrangian dual multipliers to satisfy guardrail constraints in real-time ranking?
- 约束优化怎么实现？
- Why is multiplicative scoring (pCTR * pCVR) strictly required for e-commerce ad auctions rather than linear score addition?
七、知识图谱对齐 (Knowledge Graph Anchor)
- 🔗 关联底层卡片：
多任务多目标学习：Shared-Bottom、MMoE 软门控专家网络与 PLE 渐进分流(Multi-Task Learning: Shared-Bottom, MMoE & PLE Networks) - 🗺️ 知识图谱模块：
工业级系统设计导图
🔬 算法科学家与机器学习深度考察全量题库 (Science Depth)
本题收录于 TalentMe 算法科学家深度考察真题库 (Science Depth)。全库共 856 道硬核考点，深度覆盖数学统计、经典ML、深度学习、Transformer、大语言模型、多模态、推荐系统与 MLOps。支持 Jev 面经智能匹配、一键离线单文件 HTML 手册导出并直连 Obsidian 本地记忆。