Consistency regularization and pseudo-labeling have significantly advanced semi-supervised learning (SSL). Prior works have effectively employed Mixup for consistency regularization in SSL. However, our findings indicate that applying Mixup for consistency regularization may degrade SSL performance by compromising the purity of artificial labels. Moreover, most pseudo-labeling based methods utilize thresholding strategy to exclude low-confidence data, aiming to mitigate confirmation bias; however, this approach limits the utility of unlabeled samples. To address these challenges, we propose RegMixMatch, a novel framework that optimizes the use of Mixup with both high- and low-confidence samples in SSL. First, we introduce semi-supervised RegMixup, which effectively addresses reduced artificial labels purity by using both mixed samples and clean samples for training. Second, we develop a class-aware Mixup technique that integrates information from the top-2 predicted classes into low-confidence samples and their artificial labels, reducing the confirmation bias associated with these samples and enhancing their effective utilization. Experimental results demonstrate that RegMixMatch achieves state-of-the-art performance across various SSL benchmarks.

Federated learning is vulnerable to backdoor attacks due to its distributed nature and the inability to access local datasets. Meanwhile, the heterogeneity of distributed data further complicates the detection of such attacks. However, existing defense strategies often overlook the presence of non-stationary objectives and noisy gradients across multiple clients, making it challenging to accurately and efficiently identify malicious participants. To address these challenges, we propose a backdoor defense method for Federated Learning with Adam optimizer and multi-granularity Clustering (FLAC), incorporating both coarse-grained and fine-grained clustering mechanisms to neutralize backdoor attacks. First, the Adam optimizer accelerates the learning process by mitigating the impact of noisy gradients and addressing the non-stationary objectives posed by different clients under attack. Second, a multi-granularity clustering process is considered to differentiate between benign clients and potential attackers. This is followed by an adaptive clipping strategy to further alleviate the influence of malicious attackers. Our theoretical analysis demonstrates the consistent convergence of Adam in a federated backdoor defense environment. Extensive experimental results validate the effectiveness of our defense approach.

Link prediction serves as a fundamental task in graph-based applications, where graph neural networks (GNNs) are extensively applied to estimate node connectivity likelihood. However, GNN-based methods for homogeneous graphs encounter the semantic mixing issue in heterogeneous graphs. Previous works leverage the disentangled-based model to separate the semantic information into different factors and conduct the message passing for link prediction. However, their models suffer from information loss and inadequate expression, which harms link prediction performance. To address these limitations, we propose a language-assisted semantic-aware disentangled method for link prediction on heterogeneous graphs. First, we employ a factor-wise attention mechanism to reduce the information loss caused by the disentangled model. Specifically, we design a factor selection strategy to select disentangled factors and combine them to utilize more semantic information. Second, a language-graph learning method is developed to enhance contextual expression by fusing the features of nodes and edge textual information. Extensive experiments show that the proposed method outperforms existing state-of-the-art baselines.

Time series are often complex and rich in information but sparsely labeled and therefore challenging to model. Existing contrastive learning methods conduct augmentations and maximize their similarity. However, they ignore the similarity of adjacent timestamps and suffer from the problem of sampling bias. In this paper, we propose a self-supervised framework for learning generalizable representations of time series, called$\mathbf {R}$anking n$\mathbf {E}$ighborhood and cla$\mathbf {S}$s prototyp$\mathbf {E}$contr$\mathbf {A}$stive$\mathbf {L}$earning (RESEAL). It exploits information about similarity ranking to learn an embedding space, ensuring that positive samples are ranked according to their temporal order. Additionally, RESEAL introduces a class prototype contrastive learning module. It contrasts time series representations and their corresponding centroids as positives against truly negative pairs from different clusters, mitigating the sampling bias issue. Extensive experiments conducted on several multivariate and univariate time series tasks (i.e., classification, anomaly detection, and forecasting) demonstrate that our representation framework achieves significant improvement over existing baselines of self-supervised time series representation.

Accurate and robust time series forecasting is critical in numerous domains but remains challenging due to issues such as temporal misalignments, noise, and complex multivariate dependencies. Transformer-based models have demonstrated strong performance in sequential tasks; however, their reliance on dot-product attention renders them sensitive to noise and less effective for misaligned time series. To address these limitations, we propose a novel dynamic time warping (DTW)-based attention mechanism, leveraging a Sakoe-Chiba-constrained softDTW framework to replace the traditional dot-product similarity. This approach enables dynamic sequence alignment, enhancing robustness to temporal misalignments. Building on this innovation, we introduce DTWformer, a multi-scale Transformer model that integrates DTW-attention with adaptive patching to capture dependencies across varying temporal resolutions. DTWformer achieves superior forecasting performance and efficiency, addressing the limitations of existing approaches in handling misaligned and noisy time series. The implementation is publicly available at https://github.com/unihe/DTWformer.

Music recommendation systems aim to suggest tracks that users may enjoy. However, the accuracy of recommendation results is affected by popularity bias. Previous studies have focused on mitigating the direct effect of single-item popularity in video, news, or e-commerce recommendations, but have overlooked the multisource popularity biases in music recommendations. This article proposes a causal inference-based method to reduce the influence of both track and artist popularity. First, we construct a causal graph that encompasses users, tracks, and artists within the context of music recommendations. Next, we employ matrix factorization in conjunction with counterfactual inference theory to mitigate the popularity effects of artists and tracks, taking into account both the natural direct and indirect effects of these entities on music recommendations. Experimental results evaluated on four music recommendation datasets indicate that our method outperforms other baselines and effectively alleviates the popularity bias of both tracks and artists.

As a well-known graph embedding method, Graph Convolutional Networks (GCNs) have been widely applied to recommendation systems and social media analysis, in which privacy concerns regarding sensitive data have emerged in the public view due to the collection of personal preferences. Although the regularization methods are introduced to improve the network's security, the GCN tends to memorize individual user information in latent representations susceptible to the Membership Inference Attack (MIA). In addition, the previous works focus on improving the security while hurting the utility, or vice versa, which induces “negative transfer”. In this paper, we propose a novel disentangled-based framework to defend MIA and alleviate the issue of negative transfer in multi-task learning. First, we divide the sensitive and practical channels from the latent representations of graph nodes to minimize their linear dependency. Then, to effectively train our model, we employ the sub-computational graphs to generate local gradients for different tasks and allocate losses to them. Finally, we propose a novel mixed updating strategy to accumulate the updating information of sub-computational graphs. Extensive experiments show that the proposed method can mitigate the risk of membership inference while ensuring model accuracy.

Semisupervised human activity recognition (SemiHAR) has attracted attention in recent years from various domains, such as digital health and ambient intelligence. Currently, it still faces two challenges. For one thing, discriminative features may exist among multiple sequences rather than a single sequence since activities are combinations of motions involving several body parts. For another thing, labeled data and unlabeled data suffer from distribution discrepancies due to the different behavior patterns or biological conditions of users. For that, we propose a novel SemiHAR method based on multitask learning. First, a dimension-based Markov transition field (DMTF) technique is designed to generate 2-D activity data for capturing the interactions among different dimensions. Second, we jointly consider the user recognition (UR) task and the activity recognition (AR) task to reduce the underlying discrepancy. In addition, a task relation learner (TRL) is introduced to dynamically learn task relations, which enables the primary AR task to exploit preferred knowledge from other secondary tasks. We theoretically analyze the proposed SemiHAR and provide a novel generalization result. Extensive experiments conducted on four real-world datasets demonstrate that SemiHAR outperforms other state-of-the-art methods.

Sequential recommendation aims to predict users’ future activities based on their historical interaction sequences. Various neural network architectures, such as Recurrent Neural Networks (RNN), Graph Neural Networks (GNN), and self-attention mechanisms, have been employed in the tasks, exploring multiple aspects of user preferences, including general interests, short-term interests, long-term interests, and item co-occurrence patterns. Despite achieving good performance, there are still limitations in capturing complex user preferences. Specifically, the current structures of RNN, GNN, etc., only capture item-level transition relations while neglecting attribute-level transition relations. Additionally, the explicit item relations are studied using item co-occurrence modules, but they cannot capture the implicit item-item relations. To address these issues, we propose a knowledge-augmented Gated Recurrent Unit (GRU) to improve the short-term user interest module and adopt a collaborative item aggregation method to enhance the item co-occurrence module. Additionally, our long-term interest module utilizes a bitwise gating mechanism to select historical item features significant to users’ current preferences. We extensively evaluate our model on three real-world datasets alongside competitive methods, demonstrating its effectiveness in top K sequential recommendation.

Abstract The deep forest presents a novel approach that yields competitive performance when compared to deep neural networks. Nevertheless, there are limited studies on the application of deep forest to time series classification (TSC) tasks, and the direct use of deep forest cannot effectively capture the relevant characteristics of time series. For that, this paper proposes time series cascade forest (TSCF), a model specifically designed for TSC tasks. TSCF relies on four base classifiers, i.e., random forest, completely random forest, random shapelet forest, and diverse representation canonical interval forest, allowing for feature learning on the original data from three granularities: point, subsequence, and summary statistics calculated based on intervals. The major contribution of this work, is to define an ensemble and deep classifier that significantly outperforms the individual classifiers and the original deep forest. Experimental results show that TSCF outperforms other forest-based algorithms for solving TSC problems.

The key problem of time series classification is the similarity measure between time series. In recent years, efficient and accurate similarity measurement methods of time series have attracted extensive attention from researchers. According to the different similarity measure strategies, the existing time series classification methods can be roughly divided into shape-based (original value) methods and structure-based (symbol transformation) methods. Shape-based methods usually use Euclidean distance (ED), dynamic time warping (DTW), or other methods to measure the global similarity between sequences. The disadvantage of these methods is that their measurement process does not necessarily achieve local sensible matchings of time series, which leads to a decrease in their accuracy and interpretability. To better capture the local information of the sequence, the structure-based methods discretize or symbolize the local value of the time sequence, which leads to the loss of the original information of the sequence. To address these problems, this paper proposes a novel similarity measurement method named dynamic time warping based on the local morphological pattern (MPDTW), which first decomposes the local subsequences of time series using discrete wavelet transforms for extracting the local structure information. Then, the decomposed subsequence will be encoded by the morphological pattern. Finally, the ED between points and their local structure difference based on morphological pattern will be weighted and applied to the DTW algorithm to measure the similarity between sequences. Experiments have been carried out on the classification tasks of the UCR datasets and the results show that our method outperforms the existing baselines.

Dynamic time warping (DTW) has been widely used in various domains of daily life. Essentially, DTW is a non-linear point-to-point matching method under time consistency constraints to find the optimal path between two temporal sequences. Although DTW achieves a globally optimal solution, it does not naturally capture locally reasonable alignments. Concretely, two points with entirely dissimilar local shape may be aligned. To solve this problem, we propose a novel weighted DTW based on local slope feature (LSDTW), which enhances DTW by taking regional information into consideration. LSDTW is inherently a DTW algorithm. However, it additionally attempts to pair locally similar shapes, and to avoid matching points with distinct neighborhood slopes. Furthermore, when LSDTW is used as a similarity measure in the popular nearest neighbor classifier, it beats other distance-based methods on the vast majority of public datasets, with significantly improved classification accuracies. In addition, case studies establish the interpretability of the proposed method.

Association rule mining that mainly focuses on symbolic items presented in transactions has attracted considerable interest since a rule provides a concise and intuitive description of knowledge. However, a time series is a sequence of data that is typically recorded in temporal order at fixed inte rvals of time. In order to mining rules in the context of time series data, a symbolic aggregate approximation (SAX) representation that could discretize the real-valued and high-dimensional time series data into segments and convert each segment to a symbol is applied in this paper. On this basis, a modified CBA algorithm is proposed to discover Class Sequential Rules (CSRs) and make the final prediction at first. Then we propose a new lazy associative classification method, in which the computation is performed on a demand driven basis. This is in contrast to rule-based classification methods like CBA which generate excessive number of rules, but is still unable to cover some test data with the discovered rules. Various experimental results show that our lazy associative classification for time series can be interpretable and competitive with the current state-of-the-art algorithm. In addition, four different methods that select the mined CSR(s) are proposed for carrying out associative classification.

Time series shapelets are subsequences of time series that could be representative of a class. Shapelets-based time series classification methods can be divided into two large categories. The first category integrates shapelets selection within the process of constructing classifier; while the second category disconnects the process of finding shapelets from the classification algorithm by adopting a shapelet transformation. However, there are two important limitations of shapelet transformation. First, the number of shapelets selected for transformation has great influence on classification result, but it is difficult to decide the quantity of shapelets which yields the best data for classification. Second, similar shapelets always exist among the selected shapelets in previous algorithms. In our work, the latter problem is addressed by introducing an efficient and effective pruning technique, it filters similar shapelets and decreases the number of candidate shapelets at the same time. Then, we propose a novel shapelet coverage method to select shapelets for a given dataset. The final selected shapelets are named after Discriminative Shapelets. Our experimental results demonstrate that, on the classic benchmark datasets used for time series classification, shapelet pruning and coverage method outperforms ShapeletFilter.