Discriminative Lasso

Abstract

Lasso-type variable selection has been demonstrated to be effective in handling high-dimensional data. From the biological perspective, traditional Lassotype models are capable of learning which stimuli are valuable while ignoring the many that are not, and thus perform feature selection. Traditional Lasso has the tendency to over-emphasize sparsity and to overlook the correlations between features. These drawbacks have been demonstrated to be critical in limiting its performance on real-world feature selection problems. Although some work has considered the problem of correlation, the issue of discriminative ability resulting from sparsity has been overlooked. To overcome this shortcoming, we propose a discriminative Lasso (referred to as dLasso) in which sparsity and correlation are jointly considered. Specifically, the new method can select features (or stimuli) that are correlated more strongly with the response but are less correlated with each other. Moreover, an efficient alternating direction method of multipliers (ADMM) is presented to solve the resulting sparse non-convex optimization problem. Extensive experiments on different datasets show that although our proposed model is not a convex problem, it outperforms both its approximately convex counterparts and a number of state-of-the-art feature selection methods.

Diagnosis of Esophageal Squamous Cell Carcinoma by High‐Performance Serum Metabolic Fingerprints: A Retrospective Study

“…[15b,26] Typically, the LRScore-based method successfully selected features and yielded significantly improved diagnostic results compared to the control method (without any feature selection process, p < 0.05, by paired t-test), owing to its utilization of L1 regularization to recognize the predictive metabolic features, which was also promising for other metabolic fingerprints and biomedical applications. [27] Through metabolomics with advanced analytical tools and machine learning, we successfully built a diagnostic model toward ESCC, which held great promise as a valuable complement to current diagnostic methods. Particularly, classical diagnostic strategies for ESCC, such as endoscopy combined with biopsies for histopathological confirmation, are still hampered by its invasiveness and low compliance rates (down to 33.5% [5b] ).…”

“…[ 15b,26 ] Typically, the LRScore‐based method successfully selected features and yielded significantly improved diagnostic results compared to the control method (without any feature selection process, p < 0.05, by paired t ‐test), owing to its utilization of L1 regularization to recognize the predictive metabolic features, which was also promising for other metabolic fingerprints and biomedical applications. [ 27 ]…”

Learning Local Structured Correlation Filters for Visual Tracking via Spatial Joint Regularization

“…The composite absolute penalties (CAP) family is introduced by Zhao et al [42], which allows given grouping and hierarchical relationships between the variables to be expressed. After observing the limitation of traditional lasso, Zhang et al [43] proposed a discriminative lasso to select features that are strongly correlated with the response and less correlated with each other.…”

Representation learning schemes of multichannel encephalographic signals for Brain Computer Interfaces

“…In this way the MI-trials are embedded as points in a high-dimensional space, with dimensionality equal to the number of sensors. Considering that the abundance of training data is rarely encountered in BCI practice and that computational efficiency is usually sought by working within reduced (transformed) data-spaces, we employ a recently introduced feature selection technique, the discriminative Lasso (dLasso) [142] that defines a subset of non-redundant features with high discriminatory power. dLasso deviates from the lasso-type feature selection process as it introduces stricter nonconvex constrains, while considering both the response-variable and the variable-variable correlations.…”