User Profile Expansion via Word Embeddings for Infrequent Users
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Content-aggregation websites face challenges in creating personalized content streams for infrequent users, as their user-interest profiles are often nearly empty, leading to inadequate research and experimentation in this area.
Innovation Solution
A processor-executed method generates a first representation of interests for infrequent users using pivot and extended entities from a nearest-neighbor search of word embeddings, associating nearness scores, and creates rankings for ingested articles based on similarity measures, serving a content stream accordingly.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If collaborative filtering is used to generate personalized content streams, then recommendation accuracy improves for frequent users, but the system fails to provide meaningful recommendations for infrequent users with empty or nearly empty interest profiles
Solution Approach 1:
The patent introduces word embeddings as an intermediary mechanism that bridges the gap between user interactions and content recommendations. Instead of directly relying on user interaction history (which is empty for infrequent users), the system uses word embeddings to represent both user interactions and content items in a shared semantic space, enabling recommendations based on semantic similarity rather than direct collaborative filtering signals
Solution Approach 2:
The patent transforms the recommendation problem from one based on user interaction counts to one based on semantic similarity metrics. By changing the parameter space from discrete user-item interaction frequencies to continuous semantic embeddings, the system can provide meaningful recommendations for infrequent users while maintaining accuracy for frequent users
2Measurement precision
If the system waits for sufficient user interaction data before providing personalized recommendations, then recommendation quality improves, but user engagement and content stream personalization deteriorate for infrequent users
Solution Approach 1:
The patent performs preliminary action by pre-computing and storing word embeddings for all content items in advance. This allows the system to immediately generate personalized content streams for infrequent users using their limited interaction history mapped through the pre-computed embeddings, without waiting for additional interaction data to accumulate
Solution Approach 2:
The word embedding space serves as an intermediary that enables immediate personalization by translating sparse user interactions into meaningful recommendations through semantic similarity, eliminating the need to wait for sufficient interaction data
3Measurement precision
If the system uses only explicit user feedback for personalization, then profile accuracy improves for users who provide feedback, but the profile remains nearly empty for infrequent users who do not provide explicit feedback
Solution Approach 1:
The patent implements implicit feedback mechanisms by analyzing user interactions with content items and using these signals to update user profiles through the word embedding framework. This allows the system to accumulate profile data from infrequent users through their interaction patterns rather than requiring explicit feedback
Solution Approach 2:
The word embedding space acts as an intermediary that transforms implicit interaction signals into meaningful profile representations, enabling profile construction for infrequent users without requiring explicit feedback
Data Source
AI summary
Software for a website hosting a content-aggregation service generates a first representation of interests for a user. The first representation includes a plurality of entities including pivot entities and extended entities, where the extended entities result from a nearest-neighbor search of word embeddings. Each of the extended entities is associated with a nearness score that is weighted using a distance of the extended entity from one of the pivot entities. For each of a plurality of articles ingested by the content-aggregation service, the software generates a second representation that associates an aboutness score with each of the plurality of entities. The software uses the first representation, a similarity measure, and the second representations to create rankings of the plurality of articles. The software receives a request for access to the content-aggregation service from the user and serves the user a content stream based at least in part on the rankings.