Computer Application Recommendations Using Middle-Layer Neural Embeddings
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing systems provide inaccurate and inefficient recommendations of computer applications based on similarities between online entities, leading to user disengagement due to irrelevant suggestions.
Innovation Solution
A neural network model is used to determine similarity between online entities by measuring distances in a continuous multidimensional space through a middle layer, allowing for accurate clustering and recommendation of relevant computer applications.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional similarity-based recommendation systems are used, then recommendations can be provided to users, but the accuracy and relevance of recommendations deteriorate leading to user disengagement
Solution Approach 1:
The patent transforms the similarity measurement process by changing parameters from traditional feature-based comparison to neural network embedding space coordinates. Entities are represented as vectors in a multidimensional space where similarity is determined by geometric distance metrics (e.g., cosine similarity, Euclidean distance) rather than rule-based feature matching. This parameter transformation enables more accurate and nuanced similarity assessment, directly improving recommendation relevance and preventing user disengagement.
2Measurement precision
If neural network models with multiple hidden layers are used, then similarity measurement accuracy is improved, but system complexity increases
Solution Approach 1:
The patent extracts and utilizes only the essential middle layer representations from the neural network model for similarity measurement, rather than employing the entire multi-layer network architecture for every comparison. By selecting specific intermediate layer outputs as the basis for embedding entities in multidimensional space, the system achieves high measurement precision while reducing computational complexity and avoiding the need to process through all hidden layers for each similarity query.
3Reliability
If dynamic analysis of similar entities is performed, then recommendation relevance is improved, but processing time increases
Solution Approach 1:
The patent performs preliminary action by pre-computing neural network embeddings for all entities and storing them in a database before actual recommendation queries. When a recommendation is needed, the system retrieves pre-computed embeddings and performs fast similarity calculations based on geometric distance metrics, rather than dynamically running the full neural network analysis at query time. This preliminary processing significantly reduces real-time processing time while maintaining high recommendation relevance through accurate similarity measurement.
Data Source
AI summary
Systems, methods and computer readable medium are provided for generating recommendations of computer applications for entities including: receiving, at an input of a neural network model, a set of input features for characteristics of a set of entities; automatically selecting a middle layer of the model from amongst one or more hidden layers; determining values at the middle layer for entities corresponding to output values of the nodes of the middle layer; measuring similarity distances for a first entity as compared to other entities of the set of entities based on distance between the corresponding values at the middle layer; identifying a similar entity to the first entity based on the measured similarity distance for the first entity to the similar entity being the shortest of the measured similarity distances; and providing a recommendation for a computer application for the first entity based on the similar entity.