A Weak Signal Learning Dataset and Its Baseline Method

TL;DR

This work tackles the challenge of weak signal learning by introducing the WSLD benchmark, a high-dimensional spectral dataset with dominant low-SNR samples and extreme class imbalance drawn from LAMOST data. It proposes the Parallel Dual-View Fusion Network (PDVFN), which jointly exploits a vector representation and a time-frequency representation via two specialized modules, ACR and PMTF, to capture local sequential features and global frequency-domain structures. The model is trained with a Gaussian-likelihood parameter loss ($L_{param}$) and a focal loss ($L_{focal}$) to handle uncertainty and imbalance, achieving state-of-the-art MAEs for multiple stellar-parameter regressions and superior classification metrics (AUC, F1, G-mean, MCC) on the WSLD. Overall, the dataset and PDVFN baseline provide a robust framework for advancing weak-signal perception in high-noise, imbalanced, and high-dimensional settings with practical impact in astronomical spectroscopy and beyond.

Abstract

Weak signal learning (WSL) is a common challenge in many fields like fault diagnosis, medical imaging, and autonomous driving, where critical information is often masked by noise and interference, making feature identification difficult. Even in tasks with abundant strong signals, the key to improving model performance often lies in effectively extracting weak signals. However, the lack of dedicated datasets has long constrained research. To address this, we construct the first specialized dataset for weak signal feature learning, containing 13,158 spectral samples. It features low SNR dominance (over 55% samples with SNR below 50) and extreme class imbalance (class ratio up to 29:1), providing a challenging benchmark for classification and regression in weak signal scenarios. We also propose a dual-view representation (vector + time-frequency map) and a PDVFN model tailored to low SNR, distribution skew, and dual imbalance. PDVFN extracts local sequential features and global frequency-domain structures in parallel, following principles of local enhancement, sequential modeling, noise suppression, multi-scale capture, frequency extraction, and global perception. This multi-source complementarity enhances representation for low-SNR and imbalanced data, offering a novel solution for WSL tasks like astronomical spectroscopy. Experiments show our method achieves higher accuracy and robustness in handling weak signals, high noise, and extreme class imbalance, especially in low SNR and imbalanced scenarios. This study provides a dedicated dataset, a baseline model, and establishes a foundation for future WSL research.

Figures (3)

• Figure 1: The dependencies of learning performance on data quality SNR and the distribution of the data quality. The estimation performance decreases dramatically with the decrease of SNR, and the appearence probability of data increase with the decrease of SNR. The performance patterns of $\log~g$, [Fe/H], [C/H] and classification are similar with the $T_\texttt{eff}$.
• Figure 2: Sample distribution and model estimation performance in [Fe/H]-[C/Fe] spaces. Panels (a) show the non-uniform sample distribution, while (b) demonstrate that parameter estimation errors increase as sample density decreases. Similar phenomena are observed in other parameter space and in the scenario of classification. The estimation errors are quantified as the average normalized absolute error sum.
• Figure 3: Framework of the Frequency-aware Feature Compression (FFC) Unit, a core component within the PMTF module of PDVFN. The core innovation is a frequency attention mechanism that generates frequency weights through adaptive average pooling along the temporal axis, followed by 1×1 convolutions with a sigmoid activation. These weights recalibrate the fused multi-scale and reverse convolution features, selectively emphasizing informative frequency components. The compressed output is then resized via 3×3 convolution and adaptive pooling for subsequent state-space modeling.