Source-linked AI summary

Maximum-distance nonnegative matrix factorization for unmixing highly mixed grain-size distribution data: A generalization of AnalySize

Qianqian Qi, Zhongming Chen, Peter G. M. van der Heijden

arXiv:2608.22681v1cs.LG

TL;DR

Highly mixed grain-size data can prevent AnalySize from recovering true end members because no observed specimens are close to them. The paper introduces maximum-distance NMF optimized with HALS, and reports effective recovery of true end members and abundances on highly mixed datasets.

  • Problem

    AnalySize struggles to recover true end members from highly mixed grain-size datasets, where no observed specimen is close to a true end member.

  • Method

    The paper introduces maximum-distance NMF, optimized with a hierarchical alternating least squares algorithm, to encourage greater separation among estimated end members.

  • Results

    MAD-NMF effectively recovers the true end members and abundances from highly mixed datasets, whereas AnalySize fails to recover them.

  • Takeaways & Limitations

    The proposed formulation extends AnalySize by maximizing rather than minimizing distances among end members for highly mixed grain-size data.

  • Takeaways & Limitations

    The experiments are limited to simulated datasets.

Abstract

from arXiv · show

Nonnegative matrix factorization (NMF) decomposes a nonnegative matrix into the product of two nonnegative matrices. This property makes NMF well suited for unmixing grain-size distribution data, which are inherently nonnegative and have row sums equal to one. Previous studies have shown that AnalySize, an NMF-based method, performs well on poorly mixed grain-size distribution data but struggles when the data is highly mixed, where no observed samples are close to the true end members. To overcome this limitation, we introduce a maximum-distance NMF that encourages the estimated end members to be as distinct as possible and develop a hierarchical alternating least squares algorithm for optimization. The proposed formulation can be regarded as a generalization of AnalySize, where AnalySize minimizes the distance among end members while the proposed method maximizes it. Experimental results demonstrate that the method effectively decomposes highly mixed grain-size distribution data.

1 Introduction

NMF is used to unmix compositional grain-size distributions, but AnalySize can fail when highly mixed data contain no observations near true end members. This study introduces maximum-distance NMF with HALS optimization to separate and recover end members in such cases.

  • NMF-based unmixing: NMF factorizes nonnegative grain-size distribution data into end members and specimen abundances under row-sum-to-one constraints.The data and both factors reflect compositional structure.
  • Existing methods: AnalySize combines reconstruction-error minimization with a minimum-distance constraint and performed best on synthetic and noisy coversand data when the end-member count was specified correctly.Its optimization uses HALS while enforcing row-sum-to-one constraints during iteration.
  • Existing methods: AnalySize nevertheless failed to recover true end members from a synthetic highly mixed coversand dataset.Highly mixed data contain no single observed specimen close to a true end member, so estimated end members may remain mixtures.
  • Motivation: Maximum-distance approaches seek distinct end members, but earlier simulated-annealing and genetic-algorithm methods are heuristic, while research in this direction remains limited.These methods maximize distances such as chi-squared or Manhattan distance while also considering reconstruction error.
  • Contribution: The study introduces maximum-distance NMF as a generalization of AnalySize, develops a HALS estimation algorithm, and reports effective decomposition of highly mixed data.The formulation changes the distance objective from minimizing separation to maximizing it.

2 Maximum-distance nonnegative matrix factorization

The proposed maximum-distance NMF augments reconstruction-error minimization with a term that separates estimated end members. It uses HALS optimization to address highly mixed data, where observations are not close to true end members.

  • Formulation: Maximum-distance NMF combines Frobenius reconstruction-error minimization with a distance term designed to drive end members away from observation specimens.The formulation retains nonnegativity and row-sum-to-one constraints on the factors.
  • Motivation: The separation objective is intended for highly mixed grain-size data in which no observation specimen is close to a true end member.Greater separation allows estimated end members to lie farther from observed specimens.
  • Relation to prior methods: Unlike AnalySize and MDC-NMF, which use a positive distance penalty that pulls end members toward their centroid, MAD-NMF maximizes separation among end members.The contrasting sign of the distance term changes the direction of the separation objective.

3 Algorithm

The algorithm solves the non-convex maximum-distance NMF problem with HALS block updates for S and A, followed by constrained least-squares refinement. Under stated assumptions and parameter bounds, the iterates converge to a stationary point of the penalized optimization problem.

  • HALS updates one row of S or one column of A at a time while holding the remaining blocks fixed.
  • The S subproblem uses Euclidean projection onto the nonnegative probability simplex, while the A subproblem enforces row sums through a quadratic regularization term.
  • α > 0 makes each A(:, k) subproblem strictly convex with a unique minimizer; the S subproblem is likewise strictly convex with a unique minimizer.
  • Under Assumption 1, α > 0, and 0 < λ ≤ c/(1 − 1/K)^2, every limit point of Algorithm 1 is a stationary point of the optimization problem.
  • Each HALS block update exactly minimizes its subproblem, so the objective value is monotonically nonincreasing.
  • The convergence analysis applies to the penalized objective, because the equality constraint on A is incorporated through a quadratic penalty during optimization.

4 Experiments settings

The experiments evaluate the method on synthetic highly mixed two-end-member and coversand datasets, including noisy variants. Performance is assessed by angular errors for estimated end members and abundances, supplemented by visual inspection.

  • 4.1.1 Highly mixed two-end-member dataset: The highly mixed two-end-member dataset contains 99 specimens generated from two lognormal end members over 100 grain-size classes.
  • 4.1.1 Highly mixed two-end-member dataset: The first abundance is uniformly distributed between 0.13 and 0.87, the second equals one minus the first, and no specimen is close to a pure end member.
  • 4.1.1 Highly mixed two-end-member dataset: A noisy two-end-member variant is created by element-wise multiplication with random numbers having mean 1 and standard deviation 0.01, followed by zeroing negatives and renormalization.
  • 4.1.2 Highly mixed coversand dataset: The synthetic highly mixed coversand dataset contains 200 specimens formed by linearly mixing four reported coversand end members.
  • 4.1.2 Highly mixed coversand dataset: Coversand abundances are accepted only when they sum to one, with every abundance constrained between 0.08 and 0.49.
  • 4.2 Evaluation: Performance is evaluated using mean angles between estimated and true end members and abundances, together with visual inspection.
  • 4.2 Evaluation: Smaller MAEM and MAAB values indicate better performance; R2 is excluded because it measures reconstruction accuracy rather than end-member and abundance accuracy.

5 Experimental results

Experiments on highly mixed two-end-member and coversand datasets show that MAD-NMF recovers end members and abundances more effectively than AnalySize. Its maximum-distance objective produces more separated estimated end members, while noisy coversand data yield more scattered abundance estimates.

  • Highly mixed two-end-member dataset: MAD-NMF effectively recovers end members and abundances on highly mixed two-end-member datasets, whereas AnalySize fails to recover the true components.The comparison includes both noise-free and noisy datasets.
  • Highly mixed two-end-member dataset: MAD-NMF achieves lower MAEM and MAAB values than AnalySize for the highly mixed two-end-member dataset.These quantitative comparisons are reported in Table 1 and are consistent with the visual results.
  • Highly mixed two-end-member dataset: MAD-NMF produces larger distances among estimated end members than AnalySize, matching their opposing objectives of maximizing versus minimizing those distances.The distance comparison is reported for the highly mixed two-end-member dataset.
  • Highly mixed coversand dataset: MAD-NMF effectively recovers the true end members and abundances for the highly mixed coversand dataset, while AnalySize does not.The same pattern is observed for the noisy highly mixed coversand dataset.
  • Highly mixed coversand dataset: Abundance estimates from MAD-NMF are more scattered for noisy coversand data than for the corresponding noise-free data.This comparison is made between Figure 5d and Figure 4d.

6 Conclusion and discussion

The paper proposes MAD-NMF for highly mixed grain-size distribution datasets and optimizes it with HALS. Experiments show effective recovery in highly mixed settings, while real-data, higher-noise, and parameter-selection robustness remain open questions.

  • MAD-NMF is proposed for highly mixed grain-size distribution datasets.
  • HALS optimizes the proposed MAD-NMF formulation.
  • A negative distance regularization term encourages greater separation among estimated end members, allowing them to lie farther from observed specimens.
  • MAD-NMF was evaluated on noisy and non-noisy highly mixed two-end-member and coversand datasets.
  • MAD-NMF effectively recovers true end members and abundances in highly mixed datasets, whereas AnalySize fails to do so.
  • Experiments are limited to simulated datasets, low Gaussian noise with standard deviation 0.01, and unresolved choices of λ and α.

Statements and Declarations

The authors report no competing interests.

  • Qianqian Qi declares no competing interests.
  • Zhongming Chen declares no competing interests.
  • Peter G. M. van der Heijden declares no competing interests.
Loading 2608.22681v1…