Tae Hyun Kim (Lowell)

User Profiling

1 min read #targeting#profiling

Definition

User Profiling is the task of inferring a personal preference profile (taste, context, latent patterns) from a customer’s behavioral history and representing it as a vector. It is the shared input layer for targeting, segmentation, and recommendation — the industry-side representation corresponding to patient covariate/multimodal representations (Multimodal Clinical Data) in the clinical domain.

Intuitive Understanding

It reverse-infers “what this person consumes, why, and through which hidden patterns” from behavioral traces. The key is behavior-based latent preference, not mere demographics.

Multi-Layer Profile

The 3-Layer design from LLM Factor Rec:

LayerQuestionExample (Music)
L1 ProductWhat is consumed?Genre, BPM, price range (objective)
L2 PerceivedWhat mood, what context?Mood, energy, listening context
L3 TheoreticalUnconscious patterns?Tonality, chord progression, rhythmic complexity

It mirrors item attributes to construct a user preference vector, and gains efficiency through incremental updates (summary of the existing profile + new behavior).

Role in Targeting

  • The atomic input of Customer Segmentation (profile → clusters).
  • An Affinity Matrix (user preference × product attribute similarity) yields latent affinity → combinations that are “high in latent affinity but low in realization” are targeting opportunities.
  • Part of the uplift covariate XX.

References

  • MOC-Targeting

Local graph