💬 NLP

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

Original authors: Dianjing Fan, Yao Li, Kyaw Hpone Myint, Dwipam Katariya, Alexandre G. R. Day, Pranab Mohanty, Giri Iyengar

Published 2026-06-26

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Dianjing Fan, Yao Li, Kyaw Hpone Myint, Dwipam Katariya, Alexandre G. R. Day, Pranab Mohanty, Giri Iyengar

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are walking into a massive bank.

The Problem: The "Amnesia" at the Door
In the old days, if you walked up to the bank's website (the "pre-login" area) to look at credit cards or compare loan rates, the bank's computer system would see you as a stranger. You'd browse, click around, and then log in to check your actual account. The moment you logged in, the system would say, "Hello! I see you have a checking account," but it would completely forget that you just spent 20 minutes looking at mortgage rates. It's like walking into a store, browsing the shelves for an hour, and then the cashier greets you as if you've never been there before.

The bank wanted to remember what you were looking at before you logged in to give you better recommendations after you logged in. But there were two big hurdles:

The Identity Gap: It's hard to prove that the "stranger" browsing the website is the same person who just logged into the app.
The "Black Box" Issue: Even if they could track your clicks, computers are great at counting clicks but terrible at explaining why you clicked them. They could say, "This person clicked 5 times," but they couldn't say, "This person is specifically looking for a low-interest auto loan."

The Solution: A Two-Part Super-System
The researchers built a system that acts like a smart, memory-keeping concierge. It has two main jobs, working together like a brain and a translator.

Part 1: The "Brain" (The Session Embedding)

Think of your browsing history as a long, messy stream of actions: Clicked "Credit Cards," scrolled down, hovered over "Interest Rates," clicked "Apply," then left.

The system uses a special type of AI (a Transformer) to read this entire stream at once. Instead of just counting clicks, it turns your whole browsing session into a single, compact "digital fingerprint" (called a session embedding).

The Analogy: Imagine a librarian who reads your entire shopping list and then hands you a single, perfect index card that summarizes exactly what you need. This card is so detailed that if you show it to the bank's recommendation engine later, the engine knows exactly what to show you.
The Result: When you log in, the bank uses this "fingerprint" to rearrange the tiles on your mobile app homepage. The paper found this made the bank's recommendations much more accurate (improving their ability to guess what you'd click by nearly 2%) and reduced the "noise" in their predictions.

Part 2: The "Translator" (The LLM-Distilled Taxonomy)

The "fingerprint" from Part 1 is great for computers, but it's useless for humans. A bank manager can't look at a string of numbers and say, "Ah, this customer wants a car loan." They need words.

This is where the second part comes in:

The Teacher (The LLM): They used a powerful Large Language Model (like the one you are talking to now) to read thousands of these "fingerprints" and write a dictionary of human-readable labels. It looked at the patterns and said, "Okay, this group of clicks means 'Looking for Credit Card Pre-approval,' and that group means 'Checking Travel Rewards.'"
The Student (The Distilled Classifier): The powerful LLM is slow and expensive to run on millions of people every day. So, they trained a tiny, super-fast "student" model to mimic the teacher.

The Analogy: Imagine a brilliant professor (the LLM) teaching a class of 10,000 students. The professor is too busy to grade every single paper personally. So, the professor writes a perfect grading rubric and trains a fast, efficient teaching assistant (the distilled model) to grade the papers using that rubric. The assistant is 93% as accurate as the professor but works thousands of times faster.

Why This Matters (The Results)
The paper shows that this two-part system is a win-win:

For the Computer (Quantitative): The "fingerprint" (embedding) is the best predictor. It helped the bank predict which products users would buy better than any previous method.
For the Humans (Qualitative): The "translator" (distilled labels) gave the bank managers clear, readable reasons why a user was recommended something. They could now say, "We showed this user a travel deal because our system detected they were 'Researching Vacation Packages' before they logged in."
Speed: The old way of using the powerful LLM directly would have taken hours to process a day's worth of data. The new "student" model does it in seconds, making it possible to use in real-time.

In a Nutshell
The researchers built a bridge between "anonymous web browsing" and "authenticated app usage." They created a system that remembers what you looked at before you logged in, translates that behavior into a math-heavy "fingerprint" for the computer to use, and simultaneously translates that same fingerprint into simple English labels for humans to understand. This allows financial institutions to give you personalized, relevant suggestions the moment you log in, based on your actual intent, not just your account balance.

Technical Summary: From Clicks to Intent

Problem Statement

In financial services, a critical gap exists between pre-login web interactions and post-login authenticated experiences. Unlike e-commerce or social platforms where the richest behavioral signals occur after login, financial product exploration, comparison, and application initiation frequently happen on public websites before authentication. However, current recommender systems often discard these pre-login signals because:

Context Loss: Identifying users pre-login requires complex entity resolution, and connecting this data to post-login accounts is difficult, leading to the loss of expressed intent.
Limitations of Existing Representations: Current intent representations are often ad-hoc and narrow (e.g., simple page visit logs). They fail to capture nuanced behavioral signals and lack the interpretability required for qualitative analysis, user segmentation, and regulatory compliance in the financial sector.

The core challenge is to transform raw, multi-modal web clickstreams into a representation that is both quantitatively powerful for downstream ranking tasks and qualitatively interpretable for human understanding, all while solving the cross-platform entity resolution problem.

Methodology

The authors propose a scalable, dual-purpose framework that generates two complementary outputs from pre-login web clickstream data: a dense session embedding for quantitative tasks and human-readable intent labels for qualitative understanding.

1. Multi-Modal Session Embedding (Quantitative Output)

The system encodes raw clickstream events into a compact session embedding using a self-supervised Transformer.

Input: Chronologically ordered sequences of events ( $x_1, \dots, x_T$ ) containing heterogeneous features: page information, dwell time, event type (click, scroll, view), and semantically enriched features (page HTML content).
Feature Encoding:
- Discrete features: Mapped via learned embedding tables.
- Page content: Pre-computed vectors derived from HTML text using a pretrained sentence transformer, projected to the embedding dimension.
- Position: Learned positional embeddings.
Architecture: Features are fused via an MLP, positional embeddings are added, and a Transformer encoder processes the sequence.
Training: The model is trained via self-supervised learning (predicting next or masked events/titles) without labeled data. The final model uses bidirectional attention with mean pooling to capture full context, as the complete session is available at inference time.

2. LLM-Distilled Taxonomy (Qualitative Output)

To bridge the gap between dense vectors and human understanding, the system generates an intent taxonomy and distills it into a lightweight classifier.

Clustering-Based Sampling: Session embeddings are clustered (using K-Means) to ensure stratified sampling. This guarantees the LLM sees the full diversity of user behavior rather than a random subset.
Iterative Taxonomy Generation: Inspired by TnT-LLM, an LLM generates, updates, and reviews an intent taxonomy through iterative minibatches. This process refines a fixed set of intent categories (e.g., "Credit Card Preapproval," "Fraud Protection") based on observed patterns, ensuring semantic coherence and domain relevance.
Label Assignment & Distillation:
- The LLM assigns multiple intent labels with confidence scores to sampled sessions (zero-shot prompting).
- A lightweight MLP student model is trained to predict these labels directly from the session embeddings (64-dimensional vector), using the LLM's confidence scores as soft targets.
- This allows the system to produce interpretable labels at scale without requiring LLM calls or separate text encoding during inference.

Key Contributions

Dual-Purpose Intent Prediction System: The framework produces a single shared representation with two outputs: a dense embedding for quantitative tasks (ranking, conversion prediction) and human-readable labels for qualitative understanding (segmentation, analytics).
Clustering-Enhanced LLM Taxonomy Generation: By clustering embeddings before sampling, the system ensures the LLM observes the full spectrum of user behavior during taxonomy generation, improving the robustness and coverage of the resulting intent categories.
Embedding-Space Intent Distillation: The system distills LLM-generated labels into a classifier that operates on fixed-dimensional vectors rather than raw text. This enables ultra-low latency inference at scale (10M+ daily sessions) while retaining the semantic richness of the LLM.

Experimental Results

The system was evaluated on two downstream tasks using data from a major financial services institution (approx. 10M daily sessions, filtered to 100K cross-platform sessions).

1. Mobile Homepage Tile Ranking

Task: Predicting which content tiles a user will click on the mobile app homepage after logging in.
Baseline: A LightGBM model using post-login features and coarse pre-login site retargeting features (page visit counts).
Performance: Adding the session embedding improved Macro Recall@1 by 1.88% and reduced Log Loss by 13.38% compared to the baseline.
Finding: The bidirectional attention with mean pooling variant performed best, confirming that pre-login sessions contain rich signals beyond simple page counts.

2. User Conversion Prediction

Task: Predicting post-login actions (e.g., credit card applications) based on pre-login signals.
Performance Comparison:
- Hand-crafted features (SRF): Performed significantly worse (35.8% lower F1) than LLM-based representations.
- Distilled Student (MLP): Achieved 93% of the LLM teacher's performance (7% drop in F1), validating effective knowledge transfer.
- Session Embedding: Outperformed the LLM teacher by 4.3% on Micro F1. This confirms the embedding captures the richest signal, while the distillation layer trades a small amount of accuracy for interpretability.
Label Quality: Manual review showed 97.0% strict agreement and 99.5% lenient agreement between human annotators and LLM-assigned labels.

3. Inference Efficiency

Direct LLM inference for labeling 100K sessions would take >80 hours serially.
The proposed pipeline (Transformer encoding + Distilled MLP) processes 100K sessions in under 61 seconds, representing a >3,000x throughput improvement with only a 7% performance drop.

Significance and Claims

The paper claims that this approach successfully bridges the gap between raw behavioral signals and semantic understanding in financial services.

Practical Utility: It demonstrates that self-supervised clickstream representations combined with LLM-distilled taxonomies can simultaneously serve high-stakes quantitative tasks (ranking, conversion prediction) and essential qualitative use cases (segmentation, explainability) in a production environment.
Scalability: The distillation layer allows the system to scale to millions of daily sessions without the prohibitive cost and latency of direct LLM inference.
Data Efficiency: The system leverages pre-login data, which is the only available signal for unauthenticated prospects, to drive personalization and conversion prediction, addressing a specific limitation in financial services where pre-login context is often discarded.

The authors acknowledge that the system is currently evaluated on data from a single institution and that generalizability to other domains remains to be validated. Future work includes online A/B testing, automated taxonomy refinement, and extending the model to user-level intent prediction across multiple sessions.