Neural network for graphs: a contextual constructive approach.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Graph Neural Networks

Background:

• Supervised neural networks traditionally struggle with complex, structured data.
• Existing models for graph learning often rely on recursive dynamics or strict hierarchical assumptions.

Purpose of the Study:

• To introduce a new constructive neural network for graphs (NN4G) capable of learning from general graph structures.
• To extend the applicability of supervised learning to diverse graph types, including cyclic and directed graphs.
• To enable adaptive contextual transductions for graph classification and regression.

Main Methods:

• A constructive feedforward neural network architecture (NN4G) with state variables and no feedback connections.
• A general graph traversal process that applies neurons to input graphs, relaxing causality constraints.
• An incremental learning approach that avoids cyclic dependencies in system state variables.
• Exploitation of local contextual information at graph vertices during traversal.

Main Results:

• The NN4G model successfully extends supervised learning to a general class of labeled graphs.
• The model demonstrates adaptive contextual transduction capabilities for both classification and regression.
• Compositionality of contextual information allows handling incrementally extended graph topology.
• Theoretical analysis and experimental results validate the approach's effectiveness and generality.

Conclusions:

• NN4G offers a powerful and generalizable method for learning in structured domains, particularly with graph-based data.
• The constructive, feedforward architecture provides an alternative to recursive models, simplifying learning dynamics.
• The approach effectively captures and utilizes contextual information within graph structures for various machine learning tasks.