AgentSkillsCN

em-mta-methodology

精通期望最大化算法在多触点归因分析、营销组合模型以及营销分析中应用贝叶斯推断领域的专业知识。适用于讨论 EM 算法、MTA 方法论、MMM 归因模型、潜变量模型，或各类营销归因相关问题时使用。

SKILL.md
--- frontmatter
name: em-mta-methodology
description: "Expert knowledge in Expectation-Maximization algorithms for Multi-Touch Attribution, Marketing Mix Models, and applied Bayesian inference in marketing analytics. Use when discussing EM algorithms, MTA methodology, MMM attribution, latent variable models, or marketing attribution problems."

EM/MTA Methodology Expertise

Core Concepts

Expectation-Maximization for Attribution

  • •E-step: Compute expected latent allocations $z_{b,h}$ under current parameters
  • •M-step: Update productivity parameters $\lambda_h$ using allocation weights
  • •Mass conservation: Ensure $\sum_h z_{b,h} = y_b$ (total incremental impact)
  • •Convergence: Monitor likelihood changes and parameter stability

Key Notation

  • •$\omega_{b,h}$: Exposure for entity $h$ in brick $b$ (adstocked/saturated)
  • •$\lambda_h$: Productivity parameter for entity $h$
  • •$z_{b,h}$: Latent contribution of entity $h$ to brick $b$
  • •$y_b$: Total MMM incremental impact for brick $b$

Attribution Framework

MMM → MTA Methodology

  1. •Top-down approach: Start with MMM channel totals
  2. •Exposure design: Transform raw impressions (adstock + saturation)
  3. •Allocation model: Use EM to distribute MMM totals across entities
  4. •Validation: Ensure allocations sum to MMM incrementality

Common Challenges

  • •Identifiability: Collinear exposures → arbitrary parameter splits
  • •Cold start: New entities with minimal exposure history
  • •Stability: Model performance across data updates
  • •Causality confusion: Attribution ≠ causal impact measurement

Business Communication

Stakeholder Messaging

  • •CMOs: "Preserves MMM incrementality while providing campaign-level insights"
  • •Data Scientists: "Probabilistic allocation using EM with mass conservation constraints"
  • •Finance: "Allocates exact MMM totals—no attribution inflation"
  • •Legal: "Correlation-based allocation, not causal claims about individual touchpoints"

HCP-Brick Analogy

Explain complex methodology using familiar healthcare concepts:

  • •HCP (Healthcare Provider) → Entity (campaign/creative)
  • •Brick (geographic territory) → Time/geo unit
  • •Total prescriptions → MMM incremental contribution
  • •Doctor visits → Media exposure/opportunity

Problem-Solution Patterns

When to Apply

Input signals that trigger this skill:

  • •"How to allocate channel impact to campaigns?"
  • •"EM algorithm convergence issues"
  • •"MTA vs MMM causality questions"
  • •"Attribution methodology selection"
  • •"Mass conservation in attribution"

Solution Templates

For identifiability issues:

code
Problem: Identical exposure patterns → arbitrary allocations
Solutions: 
1. Hierarchical priors with partial pooling
2. Regularization (L1/L2) on productivity parameters
3. External constraints from incrementality tests

For cold start entities:

code
Problem: New campaigns with limited history
Solutions:
1. Bayesian hierarchical model with category priors
2. Shrinkage toward category/channel means
3. Bootstrap from similar historical entities

For convergence problems:

code
Problem: EM not converging or oscillating
Solutions:
1. Check exposure collinearity (condition number)
2. Add small regularization term
3. Verify mass conservation constraints
4. Monitor likelihood and parameter changes