Vectorizing Topological Signatures for Machine Learning

365 concepts206 papers1 proposal

Topological Data Analysis (TDA) provides a powerful lens for capturing the shape, connectivity, and global structure of data. By tracking how geometric features (such as connected components, loops, and cavities) appear and disappear across multiple spatial scales, a process known as persistent homology, TDA generates rich structural summaries.

However, a fundamental mathematical mismatch prevents these summaries from being easily integrated into modern machine learning pipelines. The primary output of persistent homology is a persistence diagram (or barcode)—an unordered multiset of 2D points $(birth, death)$ representing the lifespan of topological features. Because different datasets or shapes yield diagrams with varying numbers of points, they lack a uniform dimensionality. Standard machine learning algorithms, from Support Vector Machines to Deep Neural Networks, require fixed-length vector or tensor inputs.

To bridge this gap, researchers have developed topological vectorization: a suite of methods designed to transform variable-sized multisets into stable, fixed-dimensional representations without sacrificing the underlying topological signal.


Functional and Grid-Based Embeddings

The most intuitive way to vectorize a persistence diagram is to map its discrete points into a continuous functional space or a fixed grid. Instead of treating birth-death pairs as isolated coordinates, these methods spread the topological information across a continuous domain.

One prominent approach is the Persistence Image, demonstrated by Juda in A study on topological descriptors for the analysis of 3D surface texture. This technique places a 2D Gaussian kernel over each birth-death point, scales its intensity by the feature's lifetime (ensuring that short-lived "topological noise" is de-emphasized), and integrates the resulting continuous density function over a fixed 2D grid. This transforms an abstract multiset into a standard matrix or image that convolutional neural networks (CNNs) can process natively.

[Persistence Diagram] ──(Apply Gaussian Kernels)──> [Continuous Surface] ──(Discretize)──> [Fixed Grid / Image]

Alternatively, Persistence Landscapes map diagrams into a sequence of piecewise-linear functions. As explained by Kim et al. in PLLay: Efficient Topological Layer based on Persistent Landscapes, this transformation embeds topological data into a well-behaved Hilbert space. Because these functions can be sampled at uniform intervals to form fixed-length vectors, they allow practitioners to compute averages, variances, and linear layers directly on the topological summaries while maintaining strong mathematical stability guarantees.


Quantization and Topological Codebooks

Borrowing concepts from classical computer vision and natural language processing, another family of vectorization methods treats birth-death points as "topological words" and builds a structured vocabulary.

In Persistence Bag-of-Words for Topological Data Analysis, Zieliński et al. propose quantizing the birth-death space into a learned dictionary of codewords. A variable-sized persistence diagram is vectorized by counting how many of its points fall closest to each codeword. To capture finer geometric nuances, Zieliński et al. extended this in Persistence codebooks for topological data analysis by calculating the exact residuals between persistence points and cluster centers, incorporating first- and second-order statistics akin to Fisher Vectors.

Similarly, the ATOL framework introduced by Royer et al. in ATOL: Measure Vectorization for Automatic Topologically-Oriented Learning uses unsupervised spatial quantization to place localized sensors across the diagram space. By aggregating the activations of these sensors across all points in a diagram, the method creates a dense, low-dimensional vector. This quantization strategy bypasses the computationally expensive quadratic-time distance calculations common in traditional kernel methods, allowing TDA to scale efficiently to massive datasets.


Differentiable Layers and End-to-End Learning

While hand-crafted vectorizations (like images and landscapes) are highly effective, they rely on fixed hyperparameters (such as grid resolution or kernel bandwidth) that may not be optimal for every task. To overcome this, researchers designed differentiable layers that allow neural networks to learn the best topological representation.

In the pioneering work Deep Learning with Topological Signatures, Hofer et al. introduced a neural network layer that maps persistence diagrams to vectors by calculating the sum of contributions from birth-death points to a set of Gaussian-like structure elements. Crucially, the location and scale parameters of these structural elements are learnable. By integrating this mapping directly into the backpropagation process, the network adaptively shifts its focus to the specific regions of the topological space that matter most for the target task.

This concept has been extended to various architectures:

  • Graph Neural Networks: In Graph Filtration Learning, Hofer et al. utilize differentiable coordinate functions to map graph barcodes into fixed-size vectors, allowing the model to optimize its own graph filtration functions end-to-end.
  • Direct Extraction: Rather than computing the diagram explicitly and then vectorizing it, frameworks like PI-Net by Som et al. (PI-Net: A Deep Learning Approach to Extract Topological Persistence Images) use neural networks to map raw signals directly to their corresponding persistence images, speeding up the pipeline by orders of magnitude.

Algebraic and Tropical Mappings

A key requirement for any vectorization method is stability: a small perturbation or noise injection in the input data should only cause a minor change in the resulting vector. Algebraic approaches achieve this by mapping diagrams into polynomial spaces.

In Tropical Sufficient Statistics for Persistent Homology, Monod et al. leverage tropical geometry—a framework operating over a max-plus semiring where addition represents taking a maximum and multiplication represents standard addition. By defining polynomials over this semiring, the authors map barcodes into a Euclidean space while preserving all relevant statistical properties. This establishes the vectors as sufficient statistics, meaning no information is lost, allowing standard parametric statistical models (like likelihood functions) to be rigorously applied to topological data.

Similarly, Wang et al. in Polynomial Representation for Persistence Diagram utilize bounded polynomial functions weighted by feature lifespans. The weight function ensures that noisy, short-lived features near the diagonal ($birth \approx death$) are automatically suppressed, guaranteeing that the final finite-dimensional vector remains robust against data perturbations.


Advanced Frontiers: Multiparameter Persistence

While single-parameter persistence tracks data changes across a single scale (e.g., varying a proximity radius), real-world data often requires evaluating multiple scales simultaneously (e.g., density and radius). This is known as multiparameter persistent homology.

Multiparameter persistence is significantly more expressive, but it lacks a canonical, simple representation like the 2D birth-death diagram. To vectorize these complex algebraic structures, researchers have introduced specialized decomposition techniques.

For instance, Xin et al. developed GRIL (GRIL: A $2$-parameter Persistence Based Vectorization for Machine Learning), which slices 2-parameter topological spaces along off-diagonal paths and uses unions of squares to capture features that persist across complex 2D trajectories. Further accelerating this domain, Loiseaux introduced the S-CDR framework (A Framework for Fast and Stable Representations of Multiparameter Persistent Homology Decompositions), which decomposes multiparameter modules into simpler, interval-based components. By representing these intervals via their corner points, the framework uses closed-form formulas to bin the algebraic structures into grid-based images, achieving massive computational speedups while maintaining strict mathematical stability.

History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 234 papers
  2. Version 1 · 2026-10-04 · 383 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_topological_vectorization_4b59ad8f,
  title = {topological vectorization},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/topological-vectorization/rdr_4b59ad8ff322488dbff05d0a2557370d},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-22},
}

This exact version (version 2, frozen text):

@misc{lacuna_topological_vectorization_4b59ad8f_v2,
  title = {topological vectorization},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/topological-vectorization/rdr_4b59ad8ff322488dbff05d0a2557370d/v/2},
  note = {Version 2, frozen 2026-10-04. Accessed 2026-11-22},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake