Node-Adaptive Feature Smoothing for Graph-Based Market Models

Summary

This article explains Node-Adaptive Feature Smoothing (NAFS), a graph representation method that combines node features with information from neighborhoods at multiple smoothing scales. Rather than applying the same amount of smoothing to every node, it assigns weights based on cosine similarity between a node’s original features and its smoothed representations. The resulting weighted combination is intended to reflect differences in how quickly nodes reach a stable representation. Combining smoothing operators can capture information at different graph scales. The article also describes an MQL5 implementation that constructs neighborhood averages in parallel using OpenCL.

The method is presented as a non-parametric way to produce embeddings without training the smoothing stage, with the aim of reducing computational cost and avoiding the over-smoothing associated with deep graph networks. The article reports experiments using historical market data and out-of-sample testing, and says the approaches showed potential as components of other models. However, the provided text omits much of the implementation and performance discussion, so it does not establish specific trading gains. It also reports declining monthly profitability, suggesting the training data may become less representative over time.

Key ideas

• NAFS combines a node’s own features with smoothed features from multiple neighborhood scales.
• Cosine similarity is used to weight smoothed representations for each node.
• Different smoothing operators can capture distinct graph structures and information scales.
• The smoothing method is described as non-parametric, with parallel computation used in the implementation.
• The reported market experiments are incomplete in the provided text and show declining profitability over time.

Tags

This summary was written by Stratmill's research agent from the original; it is not a copy of the source.