LLM-distilled taxonomy boosts financial recommendations by 13% log loss reduction
New framework turns anonymous web clicks into intent signals for personalized banking apps.
A team of seven researchers from industry and academia has published a paper detailing a scalable framework that bridges the gap between anonymous web browsing and authenticated in-app behavior for financial services recommendations. The challenge is that pre-login web users explore new products while logged-in app users focus on account servicing, and matching sessions across channels is notoriously difficult. Their solution uses a self-supervised Transformer to encode multi-modal clickstreams into a compact session embedding, and an LLM-based taxonomy generation and distillation pipeline that produces interpretable intent labels.
The system is evaluated on two production tasks at a financial institution. 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 existing production baselines. For user conversion prediction, the embedding outperforms the LLM-provided labels by 4.3% on micro F1, while the distilled taxonomy layer delivers interpretable labels at ultra-low latency with only a 7% performance drop. This demonstrates that self-supervised clickstream representations combined with LLM-distilled taxonomies can jointly serve quantitative ranking tasks and qualitative understanding at scale, enabling better personalization across platforms.
- Self-supervised Transformer encodes raw web clickstreams into compact session embeddings, improving mobile homepage ranking Recall@1 by 1.88% and reducing Log Loss by 13.38%.
- LLM-based taxonomy generation and distillation pipeline produces interpretable intent labels at ultra-low latency with only a 7% drop in performance compared to full LLM outputs.
- On user conversion prediction, the session embedding outperforms LLM-provided intent labels by 4.3% micro F1, showing the strength of learned representations over LLM-only approaches.
Why It Matters
Bridges the gap between anonymous web behavior and authenticated app personalization, unlocking better financial product recommendations without compromising privacy.