Feature Screening Techniques for Ultrahigh-Dimensional Data

Ultrahigh-dimensional data arise in domains such as genomics, image analysis and network science, where the number of candidate features can vastly exceed the number of observations. Feature screening serves as a preliminary step to filter out irrelevant or redundant variables, reducing dimensionality to a scale manageable for subsequent modelling. Techniques range from simple ranking by marginal correlation to more sophisticated model-free and structure-aware methods. Key objectives include retaining all truly relevant features (the sure screening property), enabling the detection of complex feature interactions and accommodating diverse data types. Advances have addressed challenges such as grouping of variables, hierarchical structures in multi-omics studies and computational efficiency for data with hundreds of thousands of features. The global significance of these methods lies in their capacity to accelerate discovery in fields from personalised medicine to natural language processing by streamlining predictive model construction without sacrificing important signals.

Research from Nature Portfolio

No recent Nature Portfolio content available.

Research from all publishers

A recent method introduces a random forest-based multiround screening (RFMS) approach for ultrahigh-dimensional multiclass problems. The algorithm partitions the full feature space into subsets, builds partial forest models and applies tournament-style sorting to select features by importance. This iterative strategy uncovers both individual predictors and higher-order interactions, demonstrating competitive performance on synthetic biometric datasets and practical advantages over single-round or purely univariate screening techniques.

In the context of multi-omics, a joint screening framework has been proposed to tackle three-level hierarchical structures: clusters of correlated genes, subgroups of omics within each gene and individual omic measurements. By combining gene clustering with marginal screening, this method secures the sure screening property at each level and outperforms competing procedures in simulations. Application to breast cancer datasets from large-scale consortia has yielded concise panels of genes and omics biomarkers linked to prognosis.

As a foundational tool for generalised ultrahigh-dimensional models, an R package implementing sure independence screening (SIS) provides efficient routines for linear, logistic and survival settings. The package delivers rapid ranking of features by marginal utility, incorporates stability assessments and interfaces seamlessly with subsequent penalised regression methods. Its theoretical underpinnings ensure that, under mild conditions, the reduced feature set retains the true predictors with high probability, facilitating accurate model fitting in high-throughput scenarios.

Feature Screening Techniques for Ultrahigh-Dimensional Data publication trend

The graph below shows the total number of articles in feature screening techniques for ultrahigh-dimensional data across all publications each year (not limited to Nature Index journals).

Technical terms

Ultrahigh-dimensional data: Datasets in which the number of features greatly exceeds the number of samples, often by several orders of magnitude.

Sure screening property: A guarantee that the screening procedure retains all truly relevant features with high probability, even while discarding many irrelevant ones.

Penalised regression: A modelling approach that adds a penalty term to the loss function to enforce sparsity, control overfitting and select important predictors.

Feature interaction: A situation in which two or more features jointly influence the response variable in a non-additive manner, requiring specialised methods to detect combined effects.

References

1. Feature space reduction method for ultrahigh-dimensional, multiclass data: random forest-based multiround screening (RFMS). Machine Learning: Science and Technology (2023).
2. Joint Screening for Ultra-High Dimensional Multi-Omics Data. Bioengineering (2024).
3. SIS : An R Package for Sure Independence Screening in Ultrahigh-Dimensional Statistical Models. Journal of Statistical Software (2018).