Data quality for the inverse Ising problem

Aurélien Decelle, Federico Ricci‐Tersenghi, Pan Zhang

Abstract

There are many methods proposed for inferring parameters of the Ising model from given data, that is a set of configurations generated according to the model itself. However little attention has been paid until now to the data, e.g. how the data is generated, whether the inference error using one set of data could be smaller than using another set of data, etc. In this paper we address the data quality problem in the kinetic inverse Ising problem. We quantify the quality of data using effective rank of the correlation matrix, and show that data gathered in a out-of-equilibrium regime has a better quality than data gathered in equilibrium for coupling reconstruction. We also propose a matrix-perturbation based method for tuning the quality of given data and for removing bad-quality (i.e. redundant) configurations from data.

Introduction

The inverse Ising problem, often called Boltzmann machine learning, is a foundational challenge in statistical mechanics and machine learning. The goal is to reconstruct interaction strengths (couplings) between variables in a network based on observed configurations. While most research focuses on developing better inference algorithms, this paper shifts the spotlight to the data itself. In many physical systems, especially at low temperatures, the system can get "stuck" in local energy minima (equilibrium), producing highly redundant configurations that are less informative for reconstruction.

Quantifying Data Quality

The authors propose using the effective rank of the correlation matrix to quantify data quality. Unlike standard rank, effective rank measures how "spread out" the data is across the configuration space. As a practical proxy, they also use standard lossless compression (gzip) on the data matrices. If a dataset is highly compressible, it contains more redundancy and is likely of lower quality for inference.

Correlation between reconstruction error and data compressibility.

Results: Equilibrium vs. Out-of-Equilibrium

Using the Sherrington-Kirkpatrick (SK) model as a benchmark, the study compares data gathered at different time steps T after a random start. Small T represents the "out-of-equilibrium" regime, while large T represents the "equilibrium" regime.

In the low-temperature (glassy) phase (beta > 1), the system tends to settle into a few states. Data gathered here is highly correlated and redundant. Conversely, data gathered immediately after a "quench" (early in the dynamics, small T) is much higher quality. Because the system hasn't settled yet, the configurations span a broader region of the state space, revealing more about the underlying coupling structure.

Reconstruction error vs. inverse temperature for different sampling times.

As shown in the figures, the reconstruction error grows monotonically with time T in the glassy phase. Out-of-equilibrium data consistently yields lower error than equilibrium data.

Smart Data Pruning

The researchers developed a matrix-perturbation decimation algorithm to identify and remove redundant data. By treating the removal of a single configuration as a small perturbation to the correlation matrix, they can efficiently estimate which data points most increase the effective rank of the remaining set. They found that throwing away 90% of a redundant dataset using this method still achieves better inference results than random removal.

Conclusion

Data quality is a critical factor in the inverse Ising problem. Information for coupling reconstruction is maximized when the system is actively fluctuating out-of-equilibrium. These findings suggest that experimentalists should intentionally perturb systems—for example, providing a "kick" to a neural circuit—to gather higher-quality data for network reconstruction rather than simply collecting more stationary data.

Summary updated

Abstract

There are many methods proposed for inferring parameters of the Ising model from given data, that is a set of configurations generated according to the model itself. However little attention has been paid until now to the data, e.g. how the data is generated, whether the inference error using one set of data could be smaller than using another set of data, etc. In this paper we address the data quality problem in the kinetic inverse Ising problem. We quantify the quality of data using effective rank of the correlation matrix, and show that data gathered in a out of-equilibrium regime has a better quality than data gathered in equilibrium for coupling reconstruction. We also propose a matrix-perturbation based method for tuning the quality of given data and for removing bad-quality (i.e. redundant) configurations from data.
BibTeX
@article{decelle2016data,
  author = {Aurélien Decelle and Federico Ricci‐Tersenghi and Pan Zhang},
  title = {{Data quality for the inverse Ising problem}},
  journal = {Journal of Physics A Mathematical and Theoretical},
  year = {2016},
  doi = {10.1088/1751-8113/49/38/384001},
}
Figure