Neural Network Social Network Completion
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for network completion in social networks are limited by the inability to accurately infer unrevealed friend relationships, leading to inaccuracies in identifying structural characteristics, especially when the network does not follow a power law structure.
Innovation Solution
A method and device using an artificial neural network to infer missing nodes and connections by learning graph structures from reference networks with similar attributes, generating candidate complete networks and selecting the most probable one based on connection probability vectors.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If Kronecker based network completion method is used, then network completion can be performed under power law assumption, but accuracy deteriorates when network does not follow power law structure
Solution Approach 1:
The patent changes the fundamental parameters and assumptions of network completion by transitioning from Kronecker product-based power law assumptions to neural network-based learning. The neural network learns connection probabilities directly from data without assuming specific network structures, allowing accurate completion of networks with various topologies including but not limited to power law structures.
Solution Approach 2:
The patent replaces the mechanical/mathematical Kronecker product system with a neural network system. Instead of using fixed mathematical transformations based on power law assumptions, the system uses learned neural network models that can adapt to different network structures, substituting rigid mathematical mechanics with flexible machine learning approaches.
2Measurement precision
If traditional algorithms are used for network completion, then computational simplicity is maintained, but accuracy of inferring unrevealed friend relationships deteriorates
Solution Approach 1:
The patent replaces traditional algorithmic approaches with neural network-based inference. The neural network learns complex patterns of friend relationships from observed data and uses this learned knowledge to infer unrevealed connections, achieving higher accuracy than traditional algorithms while managing computational complexity through efficient neural network architectures.
Solution Approach 2:
The patent performs preliminary learning of network patterns and connection probabilities before actual network completion. The neural network is trained on observed network data to learn underlying structures and relationship patterns, which are then applied to infer missing connections. This preliminary learning phase enables more accurate inference without requiring complex real-time computation during completion.
Data Source
AI summary
A device for completing a social network using an artificial neural network includes: a neural network unit configured to receive a target network having unrevealed missing nodes as input, infer the connections of the missing nodes with a neural network, and output multiple candidate complete networks according to various node sequences; and a selection unit configured to select one of the candidate complete networks outputted by the neural network unit, where the neural network unit outputs the candidate complete networks by using weights of a graph-generating neural network that has learned graph structures of reference networks having attributes similar to those of the target network, and the selection unit uses connection probability vectors obtained from the learned graph-generating neural network to select the candidate complete network probabilistically having a structure closest to that of the target network based on the connection probability vectors.