LLM Page Sequence Prediction for Flexible Digital Navigation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional data analytics systems for predicting client behavior are inflexible, inaccurate, and computationally inefficient, often requiring predefined input and output formats, limited contextual information, and extensive retraining for different use cases.
Innovation Solution
A digital page sequence machine learning system that utilizes large language models with customizable input prompts and a page order agnostic measure of loss to generate variable-length page sequence predictions, leveraging tokenized user navigation data for proactive client behavior prediction.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If conventional machine learning models with predefined input formats are used, then the system can generate predictions for specific tasks, but the system becomes inflexible and requires retraining for different downstream tasks
Solution Approach 1:
The patent applies universality by designing a single machine learning model that can perform multiple downstream tasks through prompt engineering. The model is trained once on a universal task and can be adapted to different specific tasks (e.g., conversion probability prediction, search recommendations, interaction prediction) by changing the prompts, eliminating the need for separate models for each task.
Solution Approach 2:
The patent utilizes parameter changes by modifying the prompt input format and parameters rather than retraining the model. The same trained model can adapt to different downstream tasks by changing the textual prompts and input parameter formats, allowing flexible adaptation without computational retraining overhead.
2Measurement precision
If conventional machine learning models with limited contextual information are used, then the system can process data efficiently, but the predictions become inaccurate due to insufficient context
Solution Approach 1:
The patent applies segmentation by dividing the contextual information into structured segments that can be selectively included in prompts. The system can segment user data, session history, and contextual metadata into manageable portions that feed into the model, allowing comprehensive context utilization without overwhelming the model with all available data at once.
Solution Approach 2:
The patent leverages dimensionality change by transitioning from traditional tabular data representations to text-based prompt representations. This allows the model to process contextual information in a higher-dimensional semantic space, capturing nuanced relationships and patterns that are difficult to represent in traditional tabular formats.
3Reliability
If extensive training data and model retraining are used for different downstream tasks, then the system can achieve task-specific accuracy, but the computational efficiency decreases significantly
Solution Approach 1:
The patent applies preliminary action by performing the heavy computational task of model training only once during an initial phase. The trained model is then deployed with pre-computed embeddings and parameters, allowing rapid adaptation to different downstream tasks through simple prompt changes without requiring repeated training computations.
Solution Approach 2:
The patent utilizes copying by creating prompt templates and input format copies that can be reused across different tasks. Instead of training new models for each task, the system copies the same trained model and adapts it through prompt variations, significantly reducing computational requirements while maintaining task-specific accuracy.
4Adaptability or versatility
If conventional systems with predefined output formats are used, then the system can generate specific predictions, but the system cannot scale to variable-length outputs and diverse use cases
Solution Approach 1:
The patent applies dynamics by enabling the model to generate variable-length outputs adaptively based on the input prompts and tasks. The system can dynamically adjust the number, type, and structure of output elements (e.g., single prediction, multiple recommendations, detailed analysis) without predefined format constraints, allowing flexible scaling to different use case requirements.
Data Source
AI summary
This disclosure describes one or more implementations of systems, non-transitory computer-readable media, and methods that utilizes digital page sequence data with large language models (LLMs) to generate digital page navigation predictions for users. In some implementations, the disclosed systems leverage large language models with page sequence input prompts to predict page sequences in additional digital navigation sessions. Indeed, in one or more implementations, the disclosed systems tokenize page sequences from user navigation data and utilize the tokenized page sequences to generate input prompts to utilize with an LLM to generate page sequence predictions. Furthermore, in some instances, the disclosed systems train an LLM to predict page sequences using a page order agnostic loss. Indeed, in one or more implementations, the disclosed systems utilize the LLM to execute a wide variety of use cases by utilizing the predicted page sequences to select (or generate) digital content for client devices of users.