Feature selection plays a vital role in recommender systems by identifying informative features for accurate prediction.While adaptive methods like AdaFS select instance-wise features based on sample variability, they often overlook globally important features and suffer from limited transferability.To address these limitations, we propose G2IFS (Global-to-Instance Feature Selection), a novel framework that integrates global distributional patterns with instance-level adaptation for more robust and generalizable feature selection.G2IFS consists of an online statistics module, a main scoring network, and a non-parametric Gaussian mixture module.The online statistics module maintains global estimates of class-wise statistics to compute Fisher scores, which guide the scoring network in learning instance-specific feature importance.The Gaussian module further mitigates co-adaptation and improves transferability.Extensive experiments across diverse recommendation models and real-world datasets show that G2IFS consistently outperforms state-of-the-art baselines in terms of accuracy, efficiency, and transferability.In-depth analysis further reveals that various global importance signals-when integrated into traditional methods like AdaFS-consistently lead to significant performance improvements, underscoring the general effectiveness of combining global and instance-level signals in recommender system feature selection.The code is available at https://github.com/youlj109/G2IFS.

Test-time adaptation (TTA) enhances generalization against out-of-distribution data during inference. Recent advances in TTA leverage some techniques such as contrastive learning and entropy minimization to improve the discriminability and robustness of models in target domains. However, existing methods often overlook simultaneous distribution shifts of sample and label, such as long-tail distributions, and contrastive learning approaches may require substantial storage for sample pairs. In this paper, we propose a novel Probabilistic Contrastive Test-time Adaptation (PCTA) method based on Expectation Maximization (EM), which is used to estimate the von Mises Fisher (vMF) distribution of test samples to capture both sample distribution and class proportions. The estimated distributions are used for probabilistic contrastive learning to adapt feature representations and optimize classification through class-weighted entropy minimization. Experimental results show that PCTA significantly enhances the performance across various distribution shifts and outperforms state-of-the-art methods in different scenarios involving both sample and label shifts. Code is available at https://github.com/youlj109/PCTA.

Test-Time Adaptation (TTA) aims to enhance the generalization of deep learning models when faced with test data that exhibits distribution shifts from the training data. In this context, only a pre-trained model and unlabeled test data are available, making it particularly relevant for privacy-sensitive applications. In practice, we observe that feature redundancy in embeddings tends to increase as domain shifts intensify in TTA. However, existing TTA methods often overlook this redundancy, which can hinder the model's adaptability to new data. To address this issue, we introduce Feature Redundancy Elimination for Test-time Adaptation (FRET), a novel perspective for TTA. A straightforward approach (S-FRET) is to directly minimize the feature redundancy score as an optimization objective to improve adaptation. Despite its simplicity and effectiveness, S-FRET struggles with label shifts, limiting its robustness in real-world scenarios. To mitigate this limitation, we further propose Graph-based FRET (G-FRET), which integrates a Graph Convolutional Network (GCN) with contrastive learning. This design not only reduces feature redundancy but also enhances feature discriminability in both the representation and prediction layers. Extensive experiments across multiple model architectures, tasks, and datasets demonstrate the effectiveness of S-FRET and show that G-FRET achieves state-of-the-art performance. Further analysis reveals that G-FRET enables the model to extract non-redundant and highly discriminative features during inference, thereby facilitating more robust test-time adaptation.

Deep neural networks often degrade under distribution shifts. Although domain adaptation offers a solution, privacy constraints often prevent access to source data, making Test-Time Adaptation (TTA)—which adapts using only unlabeled test data—increasingly attractive. However, current TTA methods still face practical challenges: (1) a primary focus on instance-wise alignment, overlooking CORrelation ALignment (CORAL) due to missing source correlations; (2) complex backpropagation operations for model updating, resulting in overhead computation and (3) domain forgetting. To address these challenges, we provide a theoretical analysis to investigate the feasibility of Test-time Correlation Alignment (TCA), demonstrating that correlation alignment between high-certainty instances and test instances can enhance test performances with a theoretical guarantee. Based on this, we propose two simple yet effective algorithms: LinearTCA and LinearTCA+. LinearTCA applies a simple linear transformation to achieve both instance and correlation alignment without additional model updates, while LinearTCA+ serves as a plug-and-play module that can easily boost existing TTA methods. Extensive experiments validate our theoretical insights and show that TCA methods significantly outperforms baselines across various tasks, benchmarks and backbones. Notably, LinearTCA achieves higher accuracy with only 4% GPU memory and 0.6% computation time compared to the best TTA baseline. It also outperforms existing methods on CLIP over 1.86%. Code: https://github.com/youlj109/TCA