Abstract
Technology of data visualization and data modeling is suggested. The basic of the technology is original idea of elastic net and methods of its construction and application. A short review of relevant methods has been made. The methods proposed are illustrated by applying them to the real economical, sociological and biological datasets and to some model data distributions. The basic of the technology is original idea of elastic net - regular point approximation of some manifold that is put into the multidimensional space and has in a certain sense minimal energy. This manifold is an analogue of principal surface and serves as non-linear screen on what multidimensional data are projected. Remarkable feature of the technology is its ability to work with and to fill gaps in data tables. Gaps are unknown or unreliable values of some features. It gives a possibility to predict plausibly values of unknown features by values of other ones. So it provides technology of constructing different prognosis systems and non-linear regressions.
Introduction
In the modern era of "information tombs," massive databases define objects by hundreds or thousands of features. While standard tools like graphs and diagrams work for two or three variables, they fail to capture the "multidimensionality" of real-world data. Traditional linear methods, such as Principal Component Analysis (PCA), often fall short because real data structures are frequently curved or non-linear. Furthermore, real-world datasets are notoriously messy, often containing "gaps" (missing or unreliable values) that break standard statistical models. The authors propose "Elastic Maps," a method that visualizes these complex structures while remaining "regular" enough to be useful for prediction and gap-filling.
The Elastic Net Methodology
The core innovation is the Elastic Net. Imagine a grid or a line placed inside a high-dimensional cloud of data points. The authors define an energy function $U$ that governs how this grid moves, balancing approximation with physical constraints:
- Approximation Error ($U^{(Y)}$): Pulls the nodes toward the data points.
- Stretching Energy ($U^{(E)}$): Acts like springs on the edges, preventing excessive expansion.
- Bending Energy ($U^{(R)}$): Acts on "ribs" (adjacent edges) to keep the manifold smooth and prevent folding.
By minimizing this total energy, the grid settles into a "principal surface"—a non-linear screen onto which multidimensional data can be projected. Unlike Kohonen's Self-Organizing Maps (SOM), which project points to the nearest node, Elastic Maps project points onto the actual surface of the manifold, providing a continuous visual representation.
Adaptive Topologies and Training
The net can take various shapes, such as spheres or tori, depending on the data's underlying structure. This flexibility allows the map to adapt to the intrinsic geometry of the dataset.
To find the global energy minimum without getting stuck in local "twists," the authors employ an annealing process. The map starts very rigid and gradually becomes "softer" during training. Additionally, the method handles "gaps" by treating missing features as geometric objects (lines or planes) and finding the closest point on the manifold surface to recover the missing value.
Applications and Results
The versatility of Elastic Maps was demonstrated across several domains:
- Synthetic Models: In a five-cluster model, Elastic Maps clearly distinguished all five clusters, whereas linear PCA merged two small clusters into one.
- Economics: Analysis of the 200 largest Russian companies separated the oil-gas industry into distinct subclasses based on profitability and identified that high output does not always correlate with high growth.
- Politics: Using a spherical topology for US Presidential election data (1860–2000), the method showed that "ruling party wins" and "opposition wins" naturally clustered on opposite hemispheres.
- Genomics: Visualizing DNA triplet frequencies in the Prototheca wickerhamii mitochondrial genome revealed a "bullet-like" structure with three distinct "wings" representing genes in different reading phases.
Conclusion
Elastic Maps provide a robust framework for non-linear dimensionality reduction, balancing the trade-off between accuracy and regularity. While boundary effects and parameter sensitivity remain areas for refinement, the method's ability to handle missing data and reveal complex structures makes it a powerful tool for multidisciplinary data analysis.
Summary updated