Client Interest Embeddings for Research Personalization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Sell-side firms face challenges in providing personalized research recommendations to clients based on their interests, as existing methods lack effective mechanisms to digitalize client interest profiles and research content entities, leading to inefficient distribution of research work products.
Innovation Solution
A computer system employing deep learning and embedding techniques computes client interest scores and research topic embeddings using machine learning models, such as random forests and artificial neural networks, to digitalize client interest profiles and research content entities, allowing for personalized recommendations and targeted distribution of research work products.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of operation
If traditional methods are used to distribute research work products, then distribution coverage is broad, but personalization and client engagement are insufficient
Solution Approach 1:
The patent transforms client interest profiles and research content into numerical embeddings vectors, changing the parameter representation from categorical to continuous numerical space. This enables mathematical operations and similarity calculations, allowing personalized recommendations while maintaining systematic processing capability.
Solution Approach 2:
The patent creates digital copies of client interest profiles and research content in the form of embeddings. These embeddings are stored and reused for multiple recommendation queries, enabling personalized service without repeatedly analyzing raw data, thus reducing computational complexity while improving personalization.
2Measurement precision
If manual analysis of client interests is performed, then accuracy of interest profiling is high, but time consumption and efficiency are poor
Solution Approach 1:
The patent replaces manual mechanical analysis with automated machine learning models. The random forest model and embedding generation algorithms automatically compute client interest scores and embeddings, achieving both high accuracy in interest profiling and high efficiency in processing large numbers of clients simultaneously.
Solution Approach 2:
The patent performs preliminary computation of client interest embeddings and research content embeddings in advance. These pre-computed embeddings are stored and can be quickly retrieved and matched when generating recommendations, eliminating the need for repeated complex analysis and significantly improving distribution efficiency.
3Quantity of substance
If research work products are distributed to all clients, then distribution volume is high, but resource utilization and relevance are low
Solution Approach 1:
The patent uses client engagement data as feedback to continuously refine interest profiles. By monitoring which research products clients actually engage with, the system adjusts and updates client embeddings, improving the accuracy of future distribution decisions and ensuring higher relevance while maintaining appropriate distribution volume.
4Measurement precision
If deep learning models are trained with extensive data, then model accuracy improves, but training time and computational resources increase
Solution Approach 1:
The patent uses a two-stage approach where a random forest model provides initial interest scores, and then embedding models are trained only on the most relevant client-content pairs. This partial training approach achieves sufficient accuracy without requiring exhaustive training on all possible combinations, reducing training time while maintaining prediction quality.
Data Source
AI summary
Computer systems and methods for a research organization compute interests of clients in research work product, and, using deep learning and embedding techniques, digitalize client interest profiles and research content entities. A first machine learning model is trained to compute client-interest scores from, at least in part, client engagement data. The computer system is also configured to compute embeddings for each of the clients and embeddings for each of certain research topics. The embeddings are computed using a second machine learning model, such as deep artificial neural network, such that embeddings for clients with similar interests are close, distance-wise, in a client embedding space, and such that embeddings for certain research topics with similar client engagements are close, distance-wise, in a research topic embedding space. The second machine learning model can be trained with both positive and negative training samples generated from the client-interest scores.