Targeting & Profiling Overview

Overview

Targeting & Profiling = the industrial face of personalization. The same methodological core (heterogeneous effect estimation → individual-level optimal policy; MOC-Personalization) appears in clinical settings as “optimal treatment assignment per patient,” and in industry as “optimal campaign/exposure assignment per customer.” This domain solves whom to send what in three stages — profiling/segmentation (customer representation) → uplift/CATE (individual incremental effect) → optimal targeting policy (value-maximizing rule).

flowchart LR
    X[Customer covariates] --> SEG[Customer Segmentation]
    X --> PROF[User Profiling]
    SEG --> UP[Uplift / CATE]
    PROF --> UP
    UP --> POL[Optimal Targeting Policy]
    POL --> V[Off-policy value]

Key Concepts

1. Customer Segmentation

Cluster customers by behavioral latent factors (NMF + K-Means) to form segments, then use them as HTE moderators.

2. User Profiling

Infer individual preference profiles (multi-layer L1/L2/L3) from behavioral history — the input to segmentation and targeting. Corresponds to patient covariate representation in clinical settings.

3. Uplift Modeling

The individual incremental effect of treatment (campaign exposure) = CATE under binary treatment. Selecting only the persuadables is the heart of targeting.

4. Optimal Targeting Policy

Policy learning that determines the treatment set via the $\hat\tau(x) > \text{breakeven}$ rule (Policy Learning); value is validated with OPE.

Comparison of Methods

Stage	Concept	Clinical Counterpart	Industrial Output
Representation	Customer Segmentation · User Profiling	Patient covariate / multimodal	Segments · preference profiles
Effect	Uplift Modeling (CATE)	Per-patient treatment effect	Per-customer uplift
Decision	Optimal Targeting Policy	Optimal treatment assignment (OTR)	Optimal targeting/recommendation/pricing policy

Related Concepts

• CATE · HTE · Policy Learning · Off-Policy Evaluation — shared method core (P1·P2)
• Selection Bias · Positivity — identification constraints of observational targeting data

Consuming Projects (First-Class Evidence)

• Segmentation & Causal Targeting — NMF segmentation + HTE targeting (Dunnhumby)
• LLM Factor Rec — LLM multi-layer attribute–based profiling and targeting

References

• MOC-Targeting ← domain hub