冯阳

Semiparametric Modeling and Analysis for Longitudinal Network Data

A semiparametric framework for networks observed over time that lets connection patterns evolve smoothly rather than forcing every edge to follow a fixed parametric law.

Neyman-Pearson Multi-Class Classification via Cost-Sensitive Learning

Extends the Neyman-Pearson paradigm — controlling the specific error you care about most — from binary to many classes, by recasting it as a cost-sensitive learning problem with guarantees on the user-specified error budget.

Transfer Learning under High-Dimensional Generalized Linear Models

Borrow strength from related high-dimensional generalized linear models to estimate a target model with provably minimax-optimal rates, together with a data-driven test that filters out source datasets that would actually hurt.

PCABM: Pairwise Covariates-Adjusted Block Model for Community Detection

A community-detection model that accounts for node-pair covariates partly explaining who connects to whom, blending block structure with regression-style adjustment for the part of connectivity the covariates already explain.

Testing Community Structure for Hypergraphs

A likelihood-ratio test for whether a hypergraph contains more than one community, with sharp boundaries on when communities are statistically detectable that improve over methods designed only for ordinary graphs.

RaSE: Random Subspace Ensemble Classification

A flexible high-dimensional classification framework that aggregates many randomly sampled low-dimensional subspace classifiers, with consistency theory and a CRAN package.

Neyman-Pearson Classification Algorithms and NP Receiver Operating Characteristics

A practical algorithm that controls the Type-I error of any classifier with prescribed high probability, plus an NP-ROC curve for comparing classifiers under asymmetric error costs.

Model Selection for High-Dimensional Quadratic Regression via Regularization

A regularized model-selection method for high-dimensional regression with interaction terms that respects the strong-heredity principle: an interaction is only chosen when both its main effects are also present. A weak-heredity version is also proposed that only requires one of the main effects to be present.

Nonparametric Independence Screening in Sparse Ultra-High-Dimensional Additive Models

Extends sure-independence screening to additive nonparametric models, drastically reducing ultra-high-dimensional feature pools while retaining all the truly relevant variables with high probability.

Network Exploration via the Adaptive LASSO and SCAD Penalties

Estimate the structure of a sparse Gaussian graphical model — the network of conditional dependencies among many variables — using adaptive LASSO and SCAD penalties, with oracle-property guarantees on variable selection.