An Exact Reformulation of Feature-Vector-Based Radial-Basis-Function Networks for Graph-Based Observations.

Area of Science:

• Machine Learning
• Graph Theory
• Network Analysis

Background:

• Radial basis function (RBF) networks traditionally process vector data.
• Representing relationships between objects often involves adjacency matrices for weighted, directed graphs.
• Bridging these representations is crucial for applying advanced network analysis techniques.

Purpose of the Study:

• To reformulate radial basis function (RBF) networks for direct application to adjacency matrices.
• To develop a method that avoids the need for explicit vectorization of graph data.
• To ensure the preservation of underlying relational information encoded in graph structures.

Main Methods:

• Restating the sum-of-squares objective function to depend solely on adjacency matrix entries.
• Deriving gradient descent updates for network weights based on the reformulated objective.
• Developing gradient updates for radial basis prototype repositioning and parameter adjustments.

Main Results:

• The proposed RBF networks are guaranteed to produce identical outputs to conventional RBF networks trained on equivalent vector data.
• This equivalence holds when a vector realization provably exists, such as with metric distances on latent vectors.
• The method successfully avoids constructing explicit vector realizations via multidimensional scaling, preserving all relational information.

Conclusions:

• The reformulated RBF networks offer a powerful new tool for analyzing graph-structured data.
• Directly processing adjacency matrices preserves the integrity of complex relationships.
• This approach enhances the applicability of RBF networks in fields dealing with relational data.