Name-Based Demographic Predictor

Cold-start is the hardest problem in personalization. When a new user signs up with no click or purchase history, recommendation systems have nothing to work with. This CLI tool extracts demographic signals from the one piece of data every user provides at registration — their name.

Problem

Snapdeal’s recommendation and marketing pipelines rely on behavioral signals: clicks, searches, purchases. New users with no interaction history receive generic recommendations, leading to lower engagement and higher bounce rates. Demographic attributes (gender, cultural background) can bootstrap personalization until behavioral data accumulates.

Approach

Feature Engineering

Names are decomposed into character n-grams, phonetic encodings, and subword tokens. These features capture linguistic patterns that correlate with demographics — specific character sequences, name length distributions, and phonological structures vary systematically across demographic groups.

Model Architecture

An ensemble of gradient-boosted trees and a character-level neural network processes the feature set. The ensemble approach balances interpretability (tree models) with the neural network’s ability to capture complex character patterns. The model outputs probability distributions rather than hard labels, allowing downstream systems to calibrate confidence thresholds.

CLI Interface

The tool is packaged as a CLI that accepts batch input (CSV/JSON) and outputs predictions with confidence scores. This made integration straightforward — the marketing team pipes customer lists through the tool, and the recommendation service calls it during the onboarding flow.

Results

Metric	Value
Test accuracy	92%
Coverage of cold-start users	Previously 0%, now 100%
Integration points	Recommendation pipeline, marketing personalization

Predictions feed into the recommendation pipeline as prior signals, giving new users personalized results from their first session. The marketing team uses demographic predictions to segment campaigns for users without sufficient behavioral data.