M7-049M7: Retrieval, Ranking & RecSys学习排序 (LTR)Hard
⚡Flashcard Drill➔Mastery:

Explain Multi-Objective Ranking Formulation, Loss Functions, and Pareto Optimization

🎯 Active Recall Self-Check (Draft First, Verify After)
Login-free practice tracker: recall core answer, audit pitfalls, and log review status.
Step 1 · Draft your concise answer & assumptions
Step 2 · Check Pitfalls & Follow-up Inquiries
Next: M7-050 →
📐 Mathematical Definition
score=∑iwi y^i or constrained max⁡;tradeoff: CTR↔CVR↔long-term\text{score}=\sum_i w_i\,\hat y_i\ \text{or}\ \text{constrained}\ \max;\qquad \text{tradeoff}:\ \text{CTR}\leftrightarrow\text{CVR}\leftrightarrow\text{long-term}
⚡ One-Sentence Core Answer
Multi-objective ranking models simultaneously predict diverse user engagement signals (CTR, CVR, dwell time, long-term retention) and combine them via parametric weighting, Pareto efficiency, or constrained optimization to balance short-term monetization against ecosystem health.

📌 Key Takeaways

  • •
    Multi-objective spectrum: Spans short-term engagement (CTR, dwell time), economic monetization (CVR, GMV), and long-term ecosystem health (retention, diversity, churn).
  • •
    Architectural multi-task learning (MTL): Employs Shared-Bottom, MMoE (Multi-gate Mixture-of-Experts), or PLE (Progressive Layered Extraction) to resolve task conflicts.
  • •
    Fusion mechanisms: Combines predicted probabilities via expected revenue formulas (pCTR * pCVR * bid), linear scalarization, or Constrained Optimization (Lagrangian multipliers).

📐 Mathematical Derivations

Mathematical & Optimization Formulation: Multi-Objective Mechanics.

(1) Multi-Task Loss Formulation:
Let candidate (u,d)(u, d) have multiple binary and continuous feedback labels: yclick∈{0,1}y_{\text{click}} \in \{0, 1\}, yconv∈{0,1}y_{\text{conv}} \in \{0, 1\}, and dwell time tdwell∈R+t_{\text{dwell}} \in \mathbb{R}^+. The joint loss across TT tasks is:
Ltotal=∑t=1TwtLt(y^t,yt)\mathcal{L}_{\text{total}} = \sum_{t=1}^T w_t \mathcal{L}_t(\hat{y}_t, y_t)
- For classification tasks: Binary Cross-Entropy: LBCE=−yln⁡y^−(1−y)ln⁡(1−y^)\mathcal{L}_{\text{BCE}} = - y \ln \hat{y} - (1 - y) \ln(1 - \hat{y}).
- For dwell time: Log-loss regression: Lreg=12(ln⁡(1+tdwell)−t^)2\mathcal{L}_{\text{reg}} = \frac{1}{2} (\ln(1 + t_{\text{dwell}}) - \hat{t})^2.

(2) Mitigating Task Conflicts (MMoE & PLE):
When tasks have negative transfer (e.g., high CTR clickbait having low conversion and high return rates), shared bottom layers fail. MMoE deploys EE shared expert networks {f1,…,fE}\{f_1, \dots, f_E\} routed via task-specific gating networks:
ht(x)=∑i=1Egt(x)i⋅fi(x),gt(x)=Softmax(Wtx)h_t(x) = \sum_{i=1}^E g_t(x)_i \cdot f_i(x), \quad g_t(x) = \text{Softmax}(W_t x)
Progressive Layered Extraction (PLE) further isolates task-specific experts from shared experts to eliminate 'seesaw' performance degradation.

(3) Production Value Fusion Formulas:
- E-Commerce Ad Value (eCPM):
Score=pCTR⋅pCVR⋅Price+α⋅QualityScore\text{Score} = p\text{CTR} \cdot p\text{CVR} \cdot \text{Price} + \alpha \cdot \text{QualityScore}
- Constrained Optimization via Lagrangian Relaxation:
Maximize revenue subject to maintaining minimum user satisfaction:
max⁡θE[Revenue]s.t.E[DwellTime]≥Tmin,E[Churn]≤ϵmax\max_{\theta} \mathbb{E}[\text{Revenue}] \quad \text{s.t.} \quad \mathbb{E}[\text{DwellTime}] \ge T_{\text{min}}, \quad \mathbb{E}[\text{Churn}] \le \epsilon_{\text{max}}

🏭 Production Trade-offs

In-Depth Analysis & Engineering Trade-offs: ① Short-term vs. long-term objective friction—optimizing purely for click-through rate (CTR) promotes clickbait, low-quality tabloids, and aggressive ads; within 3 months, user retention and daily active users (DAU) collapse. Modern platforms force negative constraints or penalty multipliers for early exits (tdwell<5 st_{\text{dwell}} < 5\text{ s}). ② The seesaw phenomenon in multi-task learning—increasing the weight on conversion rate (CVR) often depresses CTR; MMoE and PLE allow gradients to flow into dedicated expert pathways, preventing gradient cancellation. ③ Dynamic weight tuning in value fusion—hardcoding linear weights (w1⋅pCTR+w2⋅pCVRw_1 \cdot p\text{CTR} + w_2 \cdot p\text{CVR}) requires tedious manual retuning as market conditions change; deploying automated Bayesian optimization or online multi-objective evolutionary algorithms tunes weights continuously. ④ Calibration across disparate loss distributions—combining uncalibrated classification logits with continuous regression values corrupts ranking; all task outputs must be converted to calibrated probabilities or standardized percentiles before evaluation in scoring equations. ⑤ Gradient normalization (GradNorm)—dynamically balances the gradient magnitudes across tasks during backpropagation: wt(k)∝Lt(k)/Lt(0)mean loss ratiow_t(k) \propto \frac{\mathcal{L}_t(k) / \mathcal{L}_t(0)}{\text{mean loss ratio}}, preventing high-loss regression tasks from overwhelming classification heads. ⑥ Interview takeaway—formulate multi-task learning objectives, explain task conflict resolution via MMoE and PLE, write down the production eCPM value fusion formula, and discuss short-term vs. long-term metric governance.
⚠️ Common Interview Pitfalls
  • ✕
    Optimizing solely for short-term CTR, creating severe platform degradation through clickbait proliferation and long-term user churn.
  • ✕
    Deploying naive Shared-Bottom multi-task architectures on conflicting objectives (e.g., click vs. purchase), suffering severe negative transfer.
  • ✕
    Combining uncalibrated raw model logits directly in value fusion formulas without transforming them into calibrated event probabilities.
🎯 Interviewer Follow-up Inquiries
  • ?
    How does Progressive Layered Extraction (PLE) resolve the 'seesaw phenomenon' in multi-task learning compared to MMoE?
  • ?
    How does GradNorm automatically adjust task loss weights w_t during multi-task neural network training?
  • ?
    What mathematical formulations enforce long-term user retention constraints in real-time e-commerce ad ranking auctions?
⚡
Test muscle memory on this topic via 3-Layer Flashcards
Active Recall: Mentally formulate Prompt ➔ Math Derivation ➔ Benchmark Implementation in 3 minutes.
🎯 Drill This Card Now➔
📚

Associated Knowledge Base Guides & Mindmaps

← PreviousM7-048: Explain the Taxonomy and Classification of Ranking Features in Search and Recommendation📋 Back to BankNext →M7-050: Explain the Sources of Offline-Online Metric Discrepancies in Search and Recommendation