Continual learning (CL) aims to enable networks to learn continuously from sequentially arriving task streams while avoiding catastrophic forgetting (CF) of previously learned tasks. In recent years, Orthogonal gradient projection (OGP)-based CL methods have garnered significant attention from the research community due to their remarkable performance. However, existing OGP approaches overlook two critical issues: (1) representation matrices are typically constructed via random sampling, which introduces misclassified and class-imbalanced samples into the projection basis, contaminating important gradient directions and degrading stability; and (2) task-specific output scale variations induce domain drift, resulting in projection bias that weakens orthogonal constraints across tasks. To address these limitations, we propose Stable Gradient Projection for Continual Learning (STEP), a plug-and-play enhancement framework for OGP-based CL that integrates Correctness-aware Balanced Sampling (CBS) to construct purified and class-balanced projection subspaces using only correctly classified samples, and Sigmoid Attention Constraint (SAC) to enforce consistent output scaling via a sigmoid-based gating mechanism, thereby mitigating scale-induced projection bias. Extensive experiments on Split CIFAR-100, CIFAR-100 Superclass, and 5-Datasets demonstrate that STEP consistently improves state-of-the-art OGP methods, achieving up to +1.4% average accuracy (ACC) gains on Split CIFAR-100, improving backward transfer (BWT) from − 0.37 to − 0.09 for GPM and from − 1.06 to − 0.73 for SGP, and attaining 93.28% ACC with positive BWT (0.17) on 5-Datasets. These results validate STEP as a simple yet effective strategy for enhancing stability–plasticity balance in OGP-based CL.

In this article, we introduce a novel, general methodology, called integrate and conquer, for simultaneously accomplishing the tasks of feature extraction, manifold construction, and clustering, which is taken to be superior to building a clustering method as a single task. When the proposed novel methodology is used on two-dimensional (2D) data, it naturally induces a new clustering method highly effective on 2D data. Existing clustering algorithms usually need to convert 2D data to vectors in a preprocessing step, which, unfortunately, severely damages 2D spatial information and omits inherent structures and correlations in the original data. The induced new clustering method can overcome the matrix-vectorization-related issues to enhance the clustering performance on 2D matrices. More specifically, the proposed methodology mutually enhances three tasks of finding subspaces, learning manifolds, and constructing data representation in a seamlessly integrated fashion. When used on 2D data, we seek two projection matrices with optimal numbers of directions to project the data into low-rank, noise-mitigated, and the most expressive subspaces, in which manifolds are adaptively updated according to the projections, and new data representation is built with respect to the projected data by accounting for nonlinearity via adaptive manifolds. Consequently, the learned subspaces and manifolds are clean and intrinsic, and the new data representation is discriminative and robust. Extensive experiments have been conducted and the results confirm the effectiveness of the proposed methodology and algorithm.

Construction of a reliable similarity matrix is fundamental for graph-based clustering methods. However, most of the current work is built upon some simple manifold structure, whereas limited work has been conducted on nonlinear data sets where data reside in a union of manifolds rather than a union of subspaces. Therefore, we construct a similarity graph to capture both global and local manifold structures of the input data set. The global structure is exploited based on the self-expressive property of data in an implicit feature space using kernel methods. Since the similarity graph computation is independent of the subsequent clustering, the final results may be far from optimal. To overcome this limitation, we simultaneously learn similarity graph and clustering structure in a principled way. Experimental studies demonstrate that our proposed algorithms deliver consistently superior results to other state-of-the-art algorithms.

We introduce a new supervised learning model using a discriminative regression approach. This new model estimates a regression vector to represent the similarity between a test example and training examples while seamlessly integrating the class information in the similarity estimation. This distinguishes our model from usual regression models and locally linear embedding approaches, rendering our method suitable for supervised learning problems in high-dimensional settings. Our model is easily extensible to account for nonlinear relationship and applicable to general data, including both high- and low-dimensional data. The objective function of the model is convex, for which two optimization algorithms are provided. These two optimization approaches induce two scalable solvers that are of mathematically provable, linear time complexity. Experimental results verify the effectiveness of the proposed method on various kinds of data. For example, our method shows comparable performance on low-dimensional data and superior performance on high-dimensional data to several widely used classifiers; also, the linear solvers obtain promising performance on large-scale classification.

Matrix factorization is a useful technique for data representation in many data mining and machine learning tasks. Particularly, for data sets with all nonnegative entries, matrix factorization often requires that factor matrices be nonnegative, leading to nonnegative matrix factorization (NMF). One important application of NMF is for clustering with reduced dimensions of the data represented in the new feature space. In this paper, we propose a new graph regularized NMF method capable of feature learning and apply it to clustering. Unlike existing NMF methods that treat all features in the original feature space equally, our method distinguishes features by incorporating a feature-wise sparse approximation error matrix in the formulation. It enables important features to be more closely approximated by the factor matrices. Meanwhile, the graph of the data is constructed using cleaner features in the feature learning process, which integrates feature learning and manifold learning procedures into a unified NMF model. This distinctly differs from applying the existing graph-based NMF models after feature selection in that, when these two procedures are independently used, they often fail to align themselves toward obtaining a compact and most expressive data representation. Comprehensive experimental results demonstrate the effectiveness of the proposed method, which outperforms state-of-the-art algorithms when applied to clustering.

Matrix factorization is often used for data representation in many data mining and machine-learning problems. In particular, for a dataset without any negative entries, nonnegative matrix factorization (NMF) is often used to find a low-rank approximation by the product of two nonnegative matrices. With reduced dimensions, these matrices can be effectively used for many applications such as clustering. The existing methods of NMF are often afflicted with their sensitivity to outliers and noise in the data. To mitigate this drawback, in this paper, we consider integrating NMF into a robust principal component model, and design a robust formulation that effectively captures noise and outliers in the approximation while incorporating essential nonlinear structures. A set of comprehensive empirical evaluations in clustering applications demonstrates that the proposed method has strong robustness to gross errors and superior performance to current state-of-the-art methods.

Recommender systems play an increasingly important role in online applications to help users find what they need or prefer. Collaborative filtering algorithms that generate predictions by analyzing the user-item rating matrix perform poorly when the matrix is sparse. To alleviate this problem, this paper proposes a simple recommendation algorithm that fully exploits the similarity information among users and items and intrinsic structural information of the user-item matrix. The proposed method constructs a new representation which preserves affinity and structure information in the user-item rating matrix and then performs recommendation task. To capture proximity information about users and items, two graphs are constructed. Manifold learning idea is used to constrain the new representation to be smooth on these graphs, so as to enforce users and item proximities. Our model is formulated as a convex optimization problem, for which we need to solve the well known Sylvester equation only. We carry out extensive empirical evaluations on six benchmark datasets to show the effectiveness of this approach.

Recovering low-rank matrices is a problem common in many applications of data mining and machine learning, such as matrix completion and image denoising. Robust Principal Component Analysis (RPCA) has emerged for handling such kinds of problems; however, the existing RPCA approaches are usually computationally expensive, due to the fact that they need to obtain the singular value decomposition (SVD) of large matrices. In this paper, we propose a novel RPCA approach that eliminates the need for SVD of large matrices. Scalable algorithms are designed for several variants of our approach, which are crucial for real world applications on large scale data. Extensive experimental results confirm the effectiveness of our approach both quantitatively and visually.

Robust principal component analysis (RPCA) has been widely used for recovering low-rank matrices in many data mining and machine learning problems. It separates a data matrix into a low-rank part and a sparse part. The convex approach has been well studied in the literature. However, state-of-the-art algorithms for the convex approach usually have relatively high complexity due to the need of solving (partial) singular value decompositions of large matrices. A non-convex approach, AltProj, has also been proposed with lighter complexity and better scalability. Given the true rank r of the underlying low rank matrix, AltProj has a complexity of O(r2dn), where d × n is the size of data matrix. In this paper, we propose a novel factorization-based model of RPCA, which has a complexity of O(kdn), where k is an upper bound of the true rank. Our method does not need the precise value of the true rank. From extensive experiments, we observe that AltProj can work only when r is precisely known in advance, however, when the needed rank parameter r is specified to a value different from the true rank, AltProj cannot fully separate the two parts while our method succeeds. Even when both work, our method is about 4 times faster than AltProj. Our method can be used as a light-weight, scalable tool for RPCA in the absence of the precise value of the true rank.

Matrix rank minimization problem is in general NP-hard. The nuclear norm is used to substitute the rank function in many recent studies. Nevertheless, the nuclear norm approximation adds all singular values together and the approximation error may depend heavily on the magnitudes of singular values. This might restrict its capability in dealing with many practical problems. In this paper, an arctangent function is used as a tighter approximation to the rank function. We use it on the challenging subspace clustering problem. For this nonconvex minimization problem, we develop an effective optimization procedure based on a type of augmented Lagrange multipliers (ALM) method. Extensive experiments on face clustering and motion segmentation show that the proposed method is effective for rank approximation.

Numerous applications in data mining and machine learning require recovering a matrix of minimal rank. Robust principal component analysis (RPCA) is a general framework for handling this kind of problems. Nuclear norm based convex surrogate of the rank function in RPCA is widely investigated. Under certain assumptions, it can recover the underlying true low rank matrix with high probability. However, those assumptions may not hold in real-world applications. Since the nuclear norm approximates the rank by adding all singular values together, which is essentially a l1-norm of the singular values, the resulting approximation erroris not trivial and thus the resulting matrix estimator can be significantly biased. To seek a closer approximation and to alleviate the above-mentioned limitations of the nuclear norm, we propose a nonconvex rank approximation. This approximation to the matrix rank is tighter than the nuclear norm. To solve the associated nonconvex minimization problem, we develop an efficient augmented Lagrange multiplier based optimization algorithm. Experimental results demonstrate that our method outperforms current state-of-the-art algorithms in both accuracy and efficiency.

A number of machine learning and computer vision problems, such as matrix completion and subspace clustering, require a matrix to be of low-rank. To meet this requirement, most existing methods use the nuclear norm as a convex proxy of the rank function and minimize it. However, the nuclear norm simply adds all nonzero singular values together instead of treating them equally as the rank function does, which may not be a good rank approximation when some singular values are very large. To reduce this undesirable weighting effect, we use a log-determinant function as a non-convex rank approximation which reduces the contributions of large singular values while keeping those of small singular values close to zero. We apply the method of augmented Lagrangian multipliers to optimize this non-convex rank approximation-based objective function and obtain closed-form solutions for all subproblems of minimizing different variables alternatively. The log-determinant low-rank optimization method is used to solve subspace clustering problem, for which we construct an affinity matrix based on the angular information of the low-rank representation to enhance its separability property. Extensive experimental results on face clustering and motion segmentation data demonstrate the effectiveness of the proposed method.