Temporal Graph Neural Networks (TGNNs) are increasingly applied in dynamic scenarios, however, their limited explainability hinders their adoption in high-stakes domains. Existing methods tend to conflate causality with temporal proximity, leading to ambiguous explanations that mix impactful and irrelevant events. Moreover, they lack counterfactual reasoning to assess whether altering specific temporal events would change TGNN predictions. To overcome these challenges, we propose CTM-Explainer, which identifies critical temporal dependencies through iterative "what-if" perturbation analysis. To the best of our knowledge, this is the first post-hoc counterfactual explanation framework for TGNN. It enables precise attribution of how specific timestamped events influence TGNN predictions. By embedding causal analysis into a reinforcement learning framework, CTM-Explainer constructs Counterfactual Temporal Motifs (CTMs) that are causally grounded in model outcome shifts via interventional probability estimation. This design eliminates temporally correlated but non-essential events, while preserving those with verified causal influence. Extensive experiments on real-world and synthetic datasets confirm that CTM-Explainer generates more faithful and concise explanations than existing methods, at significantly lower computational cost.

While user preferences are important to cross-domain recommendation (CDR), existing methods primarily discover preferences under specific, yet possibly redundant, item features. To this end, we first propose a novel Preference Prototype-Aware (PPA) learning method to quantitatively learn user preferences while minimizing disturbances from the source domain. It introduces a mix-encoder and a proto-decoder. On the one hand, the mix-encoder learns better general representations of interacted items and captures the intrinsic relationships between items across different domains. On the other hand, the proto-decoder implements a learnable prototype matching mechanism to quantitatively perceive user preferences, avoiding disturbances caused by item features from the source domain. Moreover, through experiments on PPA, we observe another two issues that affect existing CDR methods' performance, i.e., the semantic deficiency caused by sparse item categories and the imbalance weights caused by different user-item distributions. Thus, we further propose a LoRA-based extractor and a domain cross-attention module to alleviate the two issues, respectively. The PPA incorporating with new extractor and attention module is called PPA++. Extensive experiments show that PPA++ outperforms the other state-of-the-art counterparts in four different CDR scenarios.

Recommender systems are widely applied in numerous online platforms such as shopping and social media platforms. They typically utilize large embedding tables that map users and items to dense vectors of uniform sizes. As the number of users and items continues to grow, this design leads to significant memory consumption and computational inefficiencies. This challenge is particularly pronounced in scenarios such as federated learning, where model parameters are updated locally on edge devices with limited computational resources before being transmitted to a central server for aggregation. Numerous approaches have been proposed to address this issue, among which embedding pruning methods have emerged as a compelling solution. Compared to parameter-sharing and variable-size embedding techniques, embedding pruning methods offer lower training costs and leverage sparse embeddings for improved efficiency. Notably, embedding pruning methods based on the Dynamic Sparse Training (DST) paradigm maintain consistent sparsity throughout training and provide a controllable memory budget, establishing them as state-of-the-art lightweight embedding solutions for resource-constrained environments. However, embedding pruning methods are not without limitations. First, despite the use of sparse embeddings during forward passes, dense gradients are still computed in backward passes, introducing inefficiencies. Second, DST's weight exploration mechanism tends to prioritize users or items from the most recent batch, reactivating pruned parameters that do not necessarily enhance overall performance. In this work, we introduce SparseRec, a lightweight embedding method designed to overcome these obstacles. SparseRec accumulates gradients to better identify inactive parameters that, when reactivated, contribute more meaningfully to model performance. Additionally, SparseRec avoids dense gradient computation during backpropagation by selectively sampling key vectors. Gradients are calculated only for parameters in this subset, ensuring sparsity throughout both forward and backward passes. Experiments on three benchmark datasets show that SparseRec achieves up to 11.79% performance gains across three base recommenders and multiple density configurations, highlighting its effectiveness in optimizing memory-constrained recommendation systems.

The high sampling frequency of highway structural health monitoring systems brings a heavy burden on data storage. However, existing dynamic response identification approaches can guarantee either reduced data volume after identification or high accuracy of dynamic response identification. Motivated by this, we propose a real-time dynamic response identification method to filter meaningless data. Our method not only selects effective features from highway structural health monitoring data, but also designs a training data generation strategy for machine learning models within the dynamic response identification framework. Experimental results on real highway structural health monitoring data demonstrate that our proposed approach spends 0.4 ms to process the monitoring data generated in 1 s and saves around 91.63% storage space. Also, the recall value of our method achieves 0.91 on average.

Dynamic time warping (DTW), a typical elastic similarity measure that compares one-to-many points, has been proven effective for various time-series data mining tasks. However, it requires a quadratic time complexity O(n²) proportional to the length of time-series data, which undermines its applications involving long time series. In this paper, a representation-based similarity measure called Dynamic Sub-Sequence Warping (DSSW) is proposed. Instead of working on the raw data directly, we perform data representation to extract the distributional features of time series. Then, the similarity between two time series is measured by aligning the corresponding sub-sequences composed of the extracted features. We evaluate the proposed method through a supervised learning task on extensive real-world datasets. The results show that DSSW outperforms the prevalent DTW-based methods in terms of precision, and achieves one order of magnitude faster execution time on average compared with DTW.