C Hou, Z Zhou.

• This paper presents OPID, a one-pass streaming learning approach that compresses information from vanished features into surviving ones and integrates augmented features without storing all data, effectively handling evolving features in incremental-decremental scenarios.

C Hou, L Zeng, D Hu.

• This paper proposes SAfe Classification (SAC), a method ensuring non-degenerate classification accuracy when using augmented features by learning diverse classifiers with robust loss and integrating them for safe predictions, supported by theoretical guarantees and experimental validation.

C Hou, S Gu, C Xu, Y Qian.

• This paper proposes SAFC, a two-stage incremental learning method that handles simultaneous augmentation of features and classes by reusing prior model knowledge via regularization, providing theoretical generalization guarantees and demonstrating effectiveness in dynamic feature-class expansion scenarios.

C Hou, R Fan, L Zeng, D Hu.

• This paper proposes AFS, an adaptive feature selection method that reuses prior feature selection models and incorporates an ℓ0-norm sparse constraint to effectively select valuable features in incremental feature scenarios, supported by theoretical analysis and experimental validation.

S Gu, C Xu, D Hu, C Hou.

• This work proposed an adaptive learning method capable of handling dynamic feature addition and reduction with noisy labels. By leveraging optimal transport and incorporating consistency constraints, heterogeneous models facilitate the estimation of the noise transition matrix and the training of classification models, with theoretical guarantees provided.

S Gu, T Luo, M He, C Hou.

• This work introduces a set of algorithms for learning from data streams with incremental feature spaces and bandit feedback. The algorithms theoretically analyze and provide bounds on the cumulative loss.

J Tu, X Tang, S Gu, Y Dai, R Fan, C Hou.

• This work proposes the ALIU algorithm to handle imbalanced data streams with unpredictable feature evolution via gradient reweighting, robust loss design, model reuse, and ensemble methods, supported by theoretical regret analysis and experimental validation

R Fan, X Ouyang, H Tao, C Hou.

• This paper proposes a method called BMDM that corrects label shift by bidirectionally matching the marginal distributions of labels and features, ensuring stable estimation of target label distributions and improving model generalization, with theoretical proof and extensions to continuous label shift scenarios.

Y Dai, S Gu, R Fan, C Xu, C Hou.

• This work proposed a method to ensure consistent adaptation to online label shift, with the regret bound provided.

C Xu, H Tao, J Zhang, D Hu, C Hou.

• This work formulates a novel LDCL framework to solve the problem of Label Distribution Changing Learning with Sample Space Expanding, together with its theoretical guarantee with generalization error bound.

X Tang, C Xu, H Tao, X Ma, C Hou.

• This paper addresses the critical challenge of instance-dependent label noise in positive-unlabeled learning. A robust framework is proposed to quantify uncertain noise through confidence measures while establishing instance-dependent noise modeling.

J Guo, H Wang, C Hou.

• This work utilizes one of the gradient flow equations’ significant physical properties, the energy dissipation law, to propose an adaptive energy-based sequential training method for PINNs.

C Xu, X Tang. H Tao, C Hou.

• This work focuses on the problem of feature-incremental learning, we distill four fundamental strategies—feature tailoring, data adaptation, model reuse, and data reconstruction. Based on these, we provide a theoretical analysis of their generalization ability and conduct a systematic comparative study to derive rigorous, quantitative insights.

J Zhang, H Tao, C Hou.

• This work is motivated by the framework of excess clustering risk, we introduce an imbalanced clustering algorithm and establish a corresponding learning-theoretic bound.

H Tao, J Jiang, C Hou, T Luo, R Fan, J Zhang.

• This paper proposes the Compound Weakly Supervised Clustering (CSWC) method, which integrates label-level pairwise constraints and feature-level multi-perspective structural consistency via a unified graph model to achieve effective clustering without additional techniques, validated on benchmark datasets and face clustering in video applications.

H Tao, C Hou, Y Qian, J Zhu, D Yi.

• This paper proposes the Latent Complete Row Space Recovery (LCRSR) method to address the inefficiency and subspace structure representation issues in multi-view subspace clustering by directly recovering the latent row space of underlying representations, avoiding time-consuming affinity matrix construction and enabling scalable clustering for large datasets, validated in multi-view data clustering and background subtraction.

C Hou, F Nie, X Li, D Yi, Y Wu.

• This paper proposes a novel unsupervised feature selection framework called JELSR that jointly performs embedding learning and sparse regression to integrate their merits for more effective feature selection, validated by experiments on various datasets.

C Hou, C Zhang, Y Wu, F Nie.

• This paper proposes a multi-view semi-supervised dimensionality reduction method that uses pairwise constraints to derive view-specific embeddings, aligns them via linear transformations, and learns a consensus pattern in low-dimensional space, supported by an iterative algorithm and experimental validation.