Structure·Centric ML
Field Note No. 07  ·  Structure-Centric ML
arXiv: 2603.13339  ·  A. Elmahdi, PhD
Priority establishment, IP protection, and the published record.

Publications & patents.

Five patent applications. One arXiv preprint. Three papers in preparation. The IP foundation under the structure-centric paradigm is being built deliberately, with priority dates established before public disclosure of implementation details.

Five filings, six independent claims.

TitleStatusFiling
AdaBox — structure-centric clustering algorithm and scale-invariant parameter formulationFiledJanuary 2026
SCOPE — decomposable evaluation framework for density-based clusteringFiledJanuary 2026
AdaGraph — native high-dimensional clustering, including SLCD pipeline, Density-Aware Sampler, two-pass prototype deployment, and subspace probe (six independent claims plus 22 dependent)Prepared, with counselApril 2026
Graph-SCOPE — unsupervised decomposable validity metric for high-dimensional clusteringPrepared, with counselApril–May 2026

The AdaGraph filing alone protects six distinct inventive concepts as independent claims: the core AdaGraph algorithm, the scale-invariant parameter formulation, the Density-Aware Sampler, the two-pass prototype deployment method, the subspace probe for automatic dimensionality assessment, and the complete SLCD pipeline. Each is independently licensable.

Public record on arXiv.

arXiv:2603.13339

Structure-Centric Density-Based Clustering with Scale-Invariant Parameters

A reduced preprint establishing priority for the structure-centric paradigm and the scale-invariant parameter formulation. Includes the foundational AdaBox algorithm and the SCOPE evaluation framework. Implementation details are deliberately abbreviated to establish priority without enabling immediate replication.

View on arXiv →

Three papers on track.

PAPER 01

AdaGraph & the Structure-Centric Approach to Native High-Dimensional Clustering

Full algorithmic specification, complexity analysis, theoretical guarantees, and empirical validation across text, genomics, and materials science. Target venue: KDD 2026.

PAPER 02

Graph-SCOPE: A Decomposable Unsupervised Validity Metric for High-Dimensional Clustering

Mathematical formulation of the five-component decomposition; benchmark against Silhouette, Davies-Bouldin, Calinski-Harabasz across 10 synthetic and 5 real-world datasets; dimensionality scaling proof. Target venue: NeurIPS 2026.

PAPER 03

Smoking-Specific Gene Co-Expression Modules in Lung Adenocarcinoma via Native-Dimension Clustering

The C6 module discovery on GSE31210, literature validation of all 44 genes, and laboratory validation of the highest-priority candidates. Co-authored with cancer genomics laboratory. Target venue: a leading genomics or oncology journal.

Why publication and protection happen together.

Algorithmic IP is notoriously difficult to enforce after public disclosure. The structure-centric portfolio addresses this by separating the priority-establishment phase (arXiv preprints with reduced implementation detail) from the full publication phase (peer-reviewed papers after patent filing is complete). Patent priority dates precede every full disclosure.

This sequencing is also why the research roadmap matters: a published paper demonstrating AdaGraph's discovery of cancer-relevant gene modules is worth more — both academically and commercially — than a paper announcing the algorithm in isolation. Each domain validation strengthens both the academic record and the commercial case.