Latent Customer Segmentation and Value-Based Recommendation Leveraging a Two-Stage Model with Missing Labels

2025-05-08

The Web Conference

Technology Category

• Machine Learning: Active Learning
• Search and Optimization: Algorithm Configuration
• Data Mining & Knowledge Management: Conversational Systems for Recommendation & Retrieval

Application Category

• User Modeling, Personalization and Recommendation: User modeling for targeted and personalized online advertising
• Economics, Online Markets and Human Computation: Humans versus LLMs for data annotation and labeling
• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking

Abstract

The success of businesses relies heavily on their ability to convert consumers into loyal customers. The value proposition of a customer is considered the primary determinant in this conversion process, and establishing intrinsic product worth at a specified value point is essential. While affordability remains a significant factor for online shoppers, brands are transitioning away from broad-scale promotions and campaigns that reduce the perceived brand worth. Implementing such widespread marketing campaigns can inadvertently erode brand equity and diminish the probability of achieving the desired return on marketing investment. Existing solutions have fallen short in addressing the twin challenges of finding the optimal balance between value proposition and long-term brand value. At the same time, existing dynamic economic algorithms often misidentify highly engaging customers as ideal campaign targets, leading to sub-optimal engagement and conversion rates, thereby diminishing customer loyalty. This article introduces a two-stage multi-model architecture that employs Self-Paced Loss to enhance customer categorization. The first layer is a Multi-Class Neural Network that differentiates high engaging customers stimulated by a campaign, high engaging customers not influenced by campaign nudges, and low engaging customers. The Binary Label Correction Model forms the second layer, further refining the classification of highly engaged customers by differentiating between those responding to a campaign and those engaging organically. The actual customer's intent can be determined using a missing label framework. This level gathers data from customers displaying variable value proposition behavior, rectifying their true state labels during training for more efficient customer segmentation. By distinguishing customer engagement intent (prompted vs organic), the suggested solution allows businesses to enhance their marketing campaign strategies and target prompted engaged segment, reducing exposure rates while boosting conversion rates. When testing this solution using an A/B test framework, we noticed an increase of over 100 basis points in the success metric.

Problem

• customer segmentation
• missing labels
• value-based recommendation
• campaign targeting
• conversion efficiency

Innovation

• two-stage model
• missing labels
• customer segmentation
• Self-Paced Loss
• value-based recommendation