From Clicks to Intent: Cross-Platform Session Embeddings with LLM-Distilled Taxonomy for Financial Services Recommendations

2026-06-24

AI Summary

This study addresses the challenge in financial recommendation systems where user behavior is fragmented across pre- and post-login states and difficult to reconcile across channels, hindering effective utilization of web-based intent signals. To bridge this gap, the authors propose a dual-purpose intent prediction framework: a self-supervised Transformer models multimodal clickstream data to generate compact session embeddings for recommendation, while a lightweight, interpretable intent classifier is constructed via knowledge distillation from a large language model (LLM). This approach uniquely unifies high-performance recommendation with semantic intent analysis in financial applications. Experimental results on a mobile homepage card ranking task show a 1.88% improvement in Recall@1 and a 13.38% reduction in Log Loss. For intent prediction, session embeddings achieve a 4.3% higher micro F1 score than LLM-generated labels, and the distilled classifier incurs only a 7% performance drop while substantially reducing inference latency.

Abstract

Sequential user behavior modeling is widely adopted in industrial recommender systems; however, significant gaps remain in financial services, where pre-login web interactions and authenticated in-app experiences differ drastically. Specifically, pre-login web users typically explore new products, whereas logged-in app users focus on account servicing. Due to the challenge of cross-channel entity resolution (e.g., matching anonymous web sessions to authenticated mobile accounts), web-based intent signals remain underutilized for post-authentication personalization. Existing methods for capturing web-based intent are often ad-hoc and narrow, lacking the flexibility to support both quantitative downstream recommendations and qualitative understanding at scale. In this work, we propose a scalable and dual-purpose intent prediction framework for web-based interactions and demonstrate its applicability for personalization. Our approach transforms raw web clickstreams into two outputs: a self-supervised Transformer encodes multi-modal clickstreams into a compact session embedding, while an LLM-based taxonomy generation and distillation pipeline produces interpretable intent labels. Our system demonstrates that self-supervised clickstream representations combined with LLM-distilled taxonomies can jointly serve quantitative tasks and qualitative understanding in production: on the mobile homepage tile ranking task, the session embedding improves macro Recall@1 by 1.88% and reduces Log Loss by 13.38% over production baselines. On the user conversion prediction task, the embedding outperforms the LLM labels by 4.3% on micro F1, while the distillation layer delivers interpretable labels at ultra-low latency with only a 7% performance drop.

Problem

• cross-platform session
• intent prediction
• financial services recommendations
• clickstream modeling
• entity resolution

Innovation

• cross-platform session embedding
• LLM-distilled taxonomy
• self-supervised clickstream modeling
• intent prediction
• financial recommender systems