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

2026-06-24 · Source: Artificial Intelligence · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Advanced, quick

Summary

A new intent prediction framework addresses the challenge of integrating pre-login web interactions with authenticated in-app experiences for financial services recommendations. This approach tackles the underutilization of web-based intent signals for post-authentication personalization, a common issue due to cross-channel entity resolution difficulties. The framework transforms raw web clickstreams into two outputs: a self-supervised Transformer generates compact session embeddings from multi-modal clickstreams, while an LLM-based taxonomy pipeline distills interpretable intent labels. This dual-purpose system demonstrates significant performance gains in production. On mobile homepage tile ranking, the session embedding improves macro Recall@1 by 1.88% and reduces Log Loss by 13.38% compared to production baselines. For user conversion prediction, the embedding achieves a 4.3% higher micro F1 score than LLM labels, with the distillation layer providing interpretable labels at ultra-low latency with only a 7% performance drop.

Key takeaway

For Machine Learning Engineers building financial services recommender systems, integrating cross-platform intent signals is crucial for enhanced personalization. You should consider adopting a dual-purpose framework that combines self-supervised session embeddings with LLM-distilled taxonomies to bridge pre-login web and in-app experiences. This approach can significantly improve metrics like Recall@1 and micro F1, while also providing interpretable intent labels for better qualitative understanding of user behavior.

Key insights

The framework combines self-supervised session embeddings with LLM-distilled taxonomies for cross-platform financial service recommendations, improving both quantitative and qualitative understanding.

Principles

• Cross-channel intent signals enhance personalization.
• Dual-purpose models serve quantitative and qualitative tasks.
• LLM distillation provides interpretable labels efficiently.

Method

The method involves a self-supervised Transformer encoding multi-modal clickstreams into session embeddings, alongside an LLM-based pipeline for taxonomy generation and distillation to produce interpretable intent labels from raw web clickstreams.

In practice

• Improve mobile homepage tile ranking.
• Enhance user conversion prediction.
• Generate interpretable intent labels for analysis.

Topics

• Financial Services
• Recommender Systems
• Session Embeddings
• LLM Taxonomy Distillation
• Cross-Platform Personalization
• Clickstream Analysis

Best for: AI Engineer, AI Scientist, Research Scientist, Machine Learning Engineer, NLP Engineer, Data Scientist

Related on AIssential

• Semantic Search At LinkedIn, LLM-Driven Autonomous Optimization for Industrial-Scale Recommendation Systems, and More! — Recent advancements in information retrieval focus on LLM-driven optimization, efficient embeddings, and scalable recommendation systems.
• TokenMizer: Graph-Structured Session Memory for Long-Horizon LLM Context Management — LLM session history is a structured knowledge artifact, not flat text, enabling efficient context management.
• From Feelings to Metrics: Understanding and Formalizing How Users Vibe-Test LLMs — Formalized "vibe-testing" bridges the gap between LLM benchmark scores and real-world user utility by personalizing evaluation.
• SHARD: Safe and Helpful Alignment via Self-Reframing Distillation — Large language models can self-refactor sensitive prompt responses into safe, helpful outputs via distillation.
• Leveraging LLMs in your Obsidian Notes — LLMs can transform note-taking by enabling conceptual search and conversational interaction with personal knowledge bases.
• Brick-DICL: Dynamic In-Context Learning for Automated Brick Schema Classification — Brick-DICL automates Brick schema classification using dynamic in-context learning, RAG for domain knowledge, and multi-LLM filtering for accuracy.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Artificial Intelligence.