From CTR to satisfaction and uplift: define the supervision before sharing parameters¶
简体中文 · English
Updated 2026-10-29 · 36 topic entries / 6 representative comparisons
ESMM constrains click-to-conversion funnel probabilities. MMoE and PLE decide how tasks share experts. FLVM distinguishes consumption, active engagement, and positive versus negative direction. Incrementality compares potential outcomes with and without an action. More prediction heads, more shared parameters, or extra latent factors do not establish causal identification.
Mechanisms · Trade-offs · Cross-paper questions · Updates
How to read this card
The table compares representative methods. It is not a shared-benchmark leaderboard. Teaching examples are not measured results.
Mechanism comparison¶
Axes: Conditional events, feedback measurements, or potential-outcome differences as labels · Parameter sharing and routing across tasks · Missing observations, confounding, and counterfactual assumptions · Mechanism evidence versus real policy incrementality
| Paper and source | Mechanism | Benefits and trade-offs |
|---|---|---|
| Entire Space Multi-Task Model: An Effective Approach for Estimating Post-Click Conversion Rate · Original source Sections in the Chinese source: 背景与主要改动, 核心公式, 本地复现 |
Supervises CTR and CTCVR over all impressions. Connects the conditional conversion tower with p(click and conversion) = pCTR × pCVR. | The funnel addresses selection and sparsity from click-only training. A non-click is not a conditional CVR negative, and the probability identity does not establish uplift under randomized intervention. |
| Modeling Task Relationships in Multi-task Learning with Multi-gate Mixture-of-Experts · Original source Sections in the Chinese source: 背景与主要改动, 核心公式 |
All tasks share a set of experts; each task has a distinct softmax gate that combines experts per example. | More flexible sharing than fully separate models, but experts still receive gradients from multiple tasks. Gates alone do not eliminate conflicts or exposure bias. |
| Progressive Layered Extraction (PLE): A Novel Multi-Task Learning Model for Personalized Recommendations · Original source Sections in the Chinese source: 背景与主要改动, 核心公式 |
Groups shared and task-specific experts. Layerwise CGC extracts common and task-specific information. | Private paths can express conflicting tasks but add capacity and routing complexity. Evaluate seesaw effects per task; an overall mean does not establish improvement everywhere. |
| Mend the Measurement Gap: Latent User Preference Modeling for Short-Form Video Recommendation · Original source Sections in the Chinese source: 背景与主要改动, 核心公式, 本地复现 |
Three latent factors represent consumption, active engagement, and positive/negative direction. Heads use fixed sparse semantic routing and a constrained stop-gradient baseline. | Separates measurement factors such as duration from preference channels, but does not identify true causal preferences. The original mask is not public; local routing is an explicit testable choice. Missing feedback is not a zero label. |
| From Prediction to Incrementality: Causal Optimization for Large-Scale Targeting and Recommendation · Original source Sections in the Chinese source: 背景与主要改动, 核心公式, 复现边界 |
Estimates potential-outcome differences with and without treatment and allocates actions by uplift and uncertainty under a fixed contact budget. | Measures extra value caused by action, not expected outcomes of naturally responsive users. Credible treatment/control and identification are required. Local proxies do not establish real-user incrementality. |
| KL-MOPD: Discrepancy-Aware Multi-Objective On-Policy Distillation for Generative Recommendation · Original source Sections in the Chinese source: 背景与主要改动, 核心公式 |
Calibrates position-level teacher discrepancy once and minimizes two student-to-teacher KL terms with fixed alpha. | Does not include initial/current student samplers, the recommendation backbone, or Semantic ID evaluation. No online gains are inherited. |
Trade-off analysis¶
Sharing does not replace label semantics¶
ESMM defines CTR, conditional CVR, and joint CTCVR. MMoE and PLE determine how those or other tasks share parameters. Labeling every non-click a conditional CVR negative is wrong even with PLE, and a correct probability chain does not remove gradient conflicts. Audit label definitions, observability, and sharing separately.
Sources: Entire Space Multi-Task Model: An Effective Approach for Estimating Post-Click Conversion Rate · Modeling Task Relationships in Multi-task Learning with Multi-gate Mixture-of-Experts · Progressive Layered Extraction (PLE): A Novel Multi-Task Learning Model for Personalized Recommendations
Separating measurement factors does not prove uplift¶
FLVM separates watching, active engagement, and positive/negative direction using a constrained baseline and three-factor routing, avoiding the assumption that watching a long video means liking it. Incrementality compares potential outcomes with and without action. Better observational proxies for satisfaction cannot replace a credible control for counterfactual decisions.
Sources: Mend the Measurement Gap: Latent User Preference Modeling for Short-Form Video Recommendation · From Prediction to Incrementality: Causal Optimization for Large-Scale Targeting and Recommendation · Entire Space Multi-Task Model: An Effective Approach for Estimating Post-Click Conversion Rate
October 29: New method and source changes¶
KL-MOPD locates competition between clicks and value at SID positions rather than combining only item-level scores. Initial-student KL discrepancies to both teachers are calibrated once and frozen; the current student computes weighted reverse KL on its own prefixes. This is not adaptive reweighting by current KL or averaging teacher probabilities. The +1.63% online revenue belongs to the full system. Local tests check loss and gradients, not causal business gains.
Sources: Entire Space Multi-Task Model: An Effective Approach for Estimating Post-Click Conversion Rate · KL-MOPD: Discrepancy-Aware Multi-Objective On-Policy Distillation for Generative Recommendation
Cross-paper questions¶
Try answering before expanding the explanation. Each question relates at least two papers.
If ESMM uses MMoE or PLE instead of two towers, can we stop distinguishing CVR and CTCVR?¶
Show answer and example
Answer: No. MMoE and PLE change parameter sharing, not probability semantics. CVR is conditional on click; CTCVR is the joint post-impression event of click and conversion. A flexible network cannot mix these labels or turn non-clicks into conditional CVR negatives.
Illustrative example: From 1,000 impressions, 100 clicks, and 20 conversions, CTR is 10%, conditional CVR is 20%, and CTCVR is 2%. Expert gating does not equalize the denominators.
Sources: Entire Space Multi-Task Model: An Effective Approach for Estimating Post-Click Conversion Rate · Modeling Task Relationships in Multi-task Learning with Multi-gate Mixture-of-Experts · Progressive Layered Extraction (PLE): A Novel Multi-Task Learning Model for Personalized Recommendations
If MMoE already has per-task gates, why add task-specific experts in PLE?¶
Show answer and example
Answer: MMoE changes gates but still mixes shared experts. Gradients can conflict in the same parameters. PLE adds private transforms while preserving a shared path. This changes the sharing boundary; it does not guarantee better scores on every task.
Illustrative example: Click prediction may favor attention-grabbing titles while satisfaction penalizes clickbait. Two gates using the same expert can still compete; private paths preserve some separate processing.
Sources: Modeling Task Relationships in Multi-task Learning with Multi-gate Mixture-of-Experts · Progressive Layered Extraction (PLE): A Novel Multi-Task Learning Model for Personalized Recommendations
If FLVM predicts a user likes content, why might Incrementality decide not to contact them?¶
Show answer and example
Answer: FLVM estimates a latent value direction behind feedback. An intervention must consider whether the user would consume anyway. Incrementality compares potential-outcome differences and uncertainty under a budget. Users with high natural preference can have little uplift.
Illustrative example: A user's return rate rises from 90% to 92% with contact, another from 10% to 20%. The first has greater absolute return, the second greater uplift. These are illustrative numbers.
Sources: Mend the Measurement Gap: Latent User Preference Modeling for Short-Form Video Recommendation · From Prediction to Incrementality: Causal Optimization for Large-Scale Targeting and Recommendation
Why not recalculate the two teacher weights from the current student at each step?¶
Show answer and example
Answer: Alpha is calibrated from the fixed initial policy, making it part of the fixed training objective. Dynamic recalculation changes both objective and gradient semantics. A dynamic variant may work but needs separate validation.
Illustrative example: An initial 1:3 gap gives weights 1/4 and 3/4. Keep them even if later gaps change.
Sources: Entire Space Multi-Task Model: An Effective Approach for Estimating Post-Click Conversion Rate · KL-MOPD: Discrepancy-Aware Multi-Objective On-Policy Distillation for Generative Recommendation
Update log¶
Updates record new approaches, changed trade-offs, and follow-up evidence.
2026-10-29 · KL-MOPD locates competition between clicks and value at SID positions rather than combining only item-level scores. Initial-student KL discrepancies to both teachers are calibrated once and frozen; the current student computes weighted reverse KL on its own prefixes. This is not adaptive reweighting by current KL or averaging teacher probabilities. The +1.63% online revenue belongs to the full system. Local tests check loss and gradients, not causal business gains.
Affected: KL-MOPD: Discrepancy-Aware Multi-Objective On-Policy Distillation for Generative Recommendation
2026-10-27 · Initial version: compared ESMM, MMoE, PLE, FLVM, and Incrementality, with analysis and questions on label semantics, shared parameters, and routing.
Covers 35 topic entries; see the full list below.
All 36 topic entries
The full catalog follows. The main comparison covers the representative papers listed above.
- Statistical Learning from Attribution Sets
- BAFF: Bid-Aware Filter Family for Mitigating Training Data Interference in RTB A/B Tests
- Baseline and Incremental: User, Item, and Request Debiasing in Instagram Feed Ranking
- CASI: Credit Assignment via Set Intervention for Agentic Retrieval
- Causal Representation Learning for Generalisable Recommendation
- Deep-learning Causal Retrieval Optimization for Efficient e-commerce Distribution in Pinterest
- Click A, Buy B: Rethinking Conversion Attribution in E-Commerce Recommendations
- Learned Cross-Task Relationships in Multi-Task Models
- Counteracting Duration Bias in Video Recommendation via Counterfactual Watch Time
- DADF: A Distribution-Aware Debiasing Framework for Watch-Time Regression in Recommender Systems
- Cheaper is Better: A Discount-Aware Network for Conversion Rate Prediction in E-commerce Recommendation System
- DCEO: Direct Causal Effect Optimization for Long-Term User Value Modeling in E-commerce Search
- Towards End-to-End Alignment of User Satisfaction via Questionnaire in Video Recommendation
- Effective Knowledge Transfer for Multi-Task Recommendation Models
- Entire Space Multi-Task Model: An Effective Approach for Estimating Post-Click Conversion Rate
- Mend the Measurement Gap: Latent User Preference Modeling for Short-Form Video Recommendation
- Generative Pseudo-Labeling for Pre-Ranking with LLMs
- Hierarchical GTV Estimation: Bridging the Gap between Container-Level Ranking and Item-Level Conversion
- From Prediction to Incrementality: Causal Optimization for Large-Scale Targeting and Recommendation
- Lightweight Ranking Heads: Accelerating Multi-Task Experimentation in Production Recommender Systems
- LLM-Derived Priors for Thompson Sampling in Cold-Start Comment Recommendation
- Macro Graph of Experts for Billion-Scale Multi-Task Recommendation
- Modeling Task Relationships in Multi-task Learning with Multi-gate Mixture-of-Experts
- Bridging Passive and Active: Enhancing Conversation Starter Recommendation via Active Expression Modeling
- PEARL: Unbiased Percentile Estimation via Contrastive Learning for Industrial-Scale Livestream Recommendation
- PinEqualizer: Full Funnel Content Exploration and Debiasing System at Pinterest
- Progressive Layered Extraction (PLE): A Novel Multi-Task Learning Model for Personalized Recommendations
- Cold-Starting Podcast Ads and Promotions with Multi-Task Learning on Spotify
- Measuring and Mitigating Identity-Cue Preference Drift in LLM-based Recommender Systems
- RAMP: Robust Ad Recommendation Under Limited Personalized-Feature Availability via Masking and Alignment Pathways
- Synthetic Data from Cross-Domain Events for Large-Scale Recommendation Systems
- SMES: Towards Scalable Multi-Task Recommendation via Expert Sparsity
- TWICE: Two-Clock, Two-Window Learning for Long-Horizon Conversion Prediction in Online Advertising
- Multi-Objective Ranking for Live-Streaming: Balancing Fresh and Delayed Signals with Segment-Aware Targeting
- Jointly Optimizing Debiased CTR and Uplift for Coupons Marketing: A Unified Causal Framework
- KL-MOPD: Discrepancy-Aware Multi-Objective On-Policy Distillation for Generative Recommendation