Single AI View of Customer: A retailers guide to Know Your Customer Better using Customer Embeddings and Customer Persona

Amlan Jyoti Das6 Mar 2025

Problem

A retailer holds a lot of customer data but mostly describes customers by demographics, which are generic and frozen in time. Traits such as "vegan" or "dog owner" appear in no column, yet they decide which products a customer wants. Building a separate model and feature set for each trait does not scale, and unsupervised clustering followed by profiling never guarantees that a specific persona is found.

Example from the article: a customer who buys almond milk is a "milk" shopper at a coarse level. Knowing she is vegan shows that soy, cashew and oat milk are relevant too.

Solution in brief

Turn each customer's activity into a "customer sentence" and encode it with a BERT-like model into one general-purpose vector. Train a classifier per persona on that vector, with labels from an automated annotation module (heuristics plus a generative model) on a customer sample. A new persona then needs only new labels, not new features.

Questions this case answers
  • How to learn one task-agnostic representation of entities from many heterogeneous data sources so that new downstream labels need no new feature engineering
  • How to infer latent entity traits at scale when ground truth has to be generated automatically and positive examples are rare
  • How to build one general customer representation that many different customer-trait models can reuse?
  • How to tag customers with latent traits (pet owner, vegan, sports fan) when no field in the data states them?

Tools and methods

  • Customer sentence — text rendering of the customer's ranked activity, made by a fine-tuned LLM
  • BERT / Sentence Transformers with a sequential auto-encoder — one 768-dimensional customer vector
  • Persona annotation module — heuristics plus a generative model create training labels
  • CatBoost, LightGBM, XGBoost, neural network — L1 binary and L2 multi-label persona classifiers
  • Dask, GPU, TensorRT, ONNX, Apache Spark — scaling embedding and persona inference
  • t-SNE — 2-D view of customer groups
Continue reading — it's free

Google or a one-time email code. No password.