LLM Prompting for Cold-Start Targeted Recommendations
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing recommender systems face inefficiencies due to the cold start problem, requiring extensive tuning and long learning times to provide relevant recommendations, leading to inconveniences for users.
Innovation Solution
A method utilizing machine learning models to identify user attributes and topics, generate prompts, and leverage large language models to provide targeted recommendations, ensuring relevance from the outset.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If existing recommender systems use collaborative filtering or content-based filtering, then they can provide recommendations based on user interactions, but they require long learning times and extensive model tuning before providing relevant recommendations
Solution Approach 1:
The patent applies preliminary action by pre-training the LLM on diverse datasets and pre-computing embeddings for content items before actual recommendation needs arise. This allows the system to quickly generate relevant recommendations without requiring extensive real-time learning, thus reducing the time loss while maintaining high recommendation relevance.
Solution Approach 2:
The patent uses an intermediary approach by introducing a LLM as a mediator between raw user interactions and final recommendations. The LLM processes user data and activity data to generate prompts, which then produce targeted recommendations. This intermediary layer enables the system to achieve high relevance quickly without requiring direct extensive tuning of traditional recommender models.
2Reliability
If existing recommender systems use collaborative filtering or content-based filtering, then they can generate recommendations, but they require extensive tuning of the model before recommending the most relevant and helpful information
Solution Approach 1:
The patent applies the disposable principle by using a flexible LLM framework that can be quickly re-prompted for different recommendation scenarios without requiring extensive re-tuning of the underlying model. The LLM serves as a reusable component that adapts to different tasks through prompt engineering rather than requiring complex model retraining or extensive parameter tuning for each specific application.
Solution Approach 2:
The patent utilizes parameter changes by modifying the prompt inputs to the LLM rather than changing the model's underlying parameters. By adjusting prompt templates, input data formats, and retrieval parameters, the system can adapt to different recommendation scenarios without extensive model tuning, thus reducing device complexity while maintaining high recommendation relevance.
3Productivity
If the system provides targeted recommendations quickly, then user convenience is improved, but the system must accurately identify user attributes and topics from limited data
Solution Approach 1:
The patent applies preliminary action by pre-computing embeddings for content items and pre-processing user data into structured formats before recommendation generation. This preparation allows the system to quickly and accurately identify user attributes and topics when receiving new user data, as the LLM can process pre-organized information rather than raw data, thus improving both speed and accuracy.
Solution Approach 2:
The patent uses segmentation by dividing the recommendation process into distinct stages: data processing and feature extraction, topic identification, and recommendation generation. Each stage handles specific tasks independently, allowing the system to accurately identify user attributes in one pass and then quickly generate recommendations in subsequent passes, thus improving both measurement precision and productivity.
Data Source
AI summary
The present disclosure provides techniques for recommending targeted information. One example method includes receiving user data indicative of one or more attributes of one or more users and activity data indicating past actions by the one or more users in association with particular content items, identifying, using a first machine learning model, a topic based on the activity data, identifying, using a second machine learning model, a subset of attributes of the one or more attributes of the one or more users that are associated with the topic, generating a prompt based on the topic and the subset of attributes associated with the topic, and generating, based on the prompt using a large language model (LLM), content to provide to a user having the subset of attributes associated with the topic.