Blind sparse-spike deconvolution with thin layers and structure

Blind sparse-spike deconvolution is a widely used method to estimate seismic wavelets and sparse reflectivity in the shape of spikes based on the convolution model. To increase the vertical resolution and lateral continuity of the estimated reflectivity, we further improve the sparse-spike deconvolution by introducing the atomic norm minimization and structural regularization, respectively. Specifically, we use the atomic norm minimization to estimate the reflector locations, which are further used as position constraints in the sparse-spike deconvolution. By doing this, we can vertically separate highly thin layers through the sparse deconvolution. In addition, the seismic structural orientations are estimated from the seismic image to construct a structure-guided regularization in the deconvolution to preserve the lateral continuity of reflectivities. Our improvements are suitable for most types of sparse-spike deconvolution approaches. The sparse-spike deconvolution method with Toeplitz-sparse matrix factorization (TSMF) is used as an example to demonstrate the effectiveness of our improvements. Synthetic and real examples show that our methods perform better than TSMF in estimating the reflectivity of thin layers and preserving the lateral continuities.

Signal Deconvolution and Noise Factor Analysis Based on a Combination of Time–Frequency Analysis and Probabilistic Sparse Matrix Factorization

Nuclear magnetic resonance (NMR) spectroscopy is commonly used to characterize molecular complexity because it produces informative atomic-resolution data on the chemical structure and molecular mobility of samples non-invasively by means of various acquisition parameters and pulse programs. However, analyzing the accumulated NMR data of mixtures is challenging due to noise and signal overlap. Therefore, data-cleansing steps, such as quality checking, noise reduction, and signal deconvolution, are important processes before spectrum analysis. Here, we have developed an NMR measurement informatics tool for data cleansing that combines short-time Fourier transform (STFT; a time–frequency analytical method) and probabilistic sparse matrix factorization (PSMF) for signal deconvolution and noise factor analysis. Our tool can be applied to the original free induction decay (FID) signals of a one-dimensional NMR spectrum. We show that the signal deconvolution method reduces the noise of FID signals, increasing the signal-to-noise ratio (SNR) about tenfold, and its application to diffusion-edited spectra allows signals of macromolecules and unsuppressed small molecules to be separated by the length of the T2* relaxation time. Noise factor analysis of NMR datasets identified correlations between SNR and acquisition parameters, identifying major experimental factors that can lower SNR.

ProNE: Fast and Scalable Network Representation Learning

Recent advances in network embedding has revolutionized the field of graph and network mining. However, (pre-)training embeddings for very large-scale networks is computationally challenging for most existing methods. In this work, we present ProNE---a fast, scalable, and effective model, whose single-thread version is 10--400x faster than efficient network embedding benchmarks with 20 threads, including LINE, DeepWalk, node2vec, GraRep, and HOPE. As a concrete example, the single-version ProNE requires only 29 hours to embed a network of hundreds of millions of nodes while it takes LINE weeks and DeepWalk months by using 20 threads. To achieve this, ProNE first initializes network embeddings efficiently by formulating the task as sparse matrix factorization. The second step of ProNE is to enhance the embeddings by propagating them in the spectrally modulated space. Extensive experiments on networks of various scales and types demonstrate that ProNE achieves both effectiveness and significant efficiency superiority when compared to the aforementioned baselines. In addition, ProNE's embedding enhancement step can be also generalized for improving other models at speed, e.g., offering >10% relative gains for the used baselines.