Data augmentation plays a crucial role in enhancing sequential recommendation (SR) by providing richer training signals. Recently, diffusion models (DMs) have been introduced into SR to generate realistic interaction sequences. However, existing DM-based methods face three key limitations: (1) semantic deviation. The rounding procedure, which maps continuous embeddings to discrete item sequences, may introduce semantic deviation; (2) preference misalignment. The explicit preference guidance is neglected during generation, resulting in synthetic sequences that misalign with users' actual interests; (3) suboptimal training strategy. The two-stage methods, which train the SR model and the DM separately, overlook their potential complementarity. To address these challenges, we propose Continuous Data Augmentation via Condition-Tokenized Diffusion Transformer for Sequential Recommendation (CATDiT). Specifically, CATDiT discards the rounding operation and leverages continuous embeddings as augmented data to preserve semantic integrity. Then, we guide the generation process with user intent via a condition-tokenized Diffusion Transformer, aligning synthetic sequences with users' real preferences. Finally, we propose an alternating optimization strategy to enable mutual learning between the SR model and the DM. Extensive experiments on five real-world datasets demonstrate that CATDiT consistently outperforms state-of-the-art baselines, validating its effectiveness in generating high-quality sequences and improving SR performance.

Heterogeneous graph neural networks (HGNNs) have recently demonstrated significant advantages of capturing powerful structural and semantic information in heterogeneous graphs. Different from homogeneous graph neural networks directly aggregating information based on neighbors, HGNNs aggregate information based on complex metapaths. ReRAM-based processing-in-memory (PIM) architecture can reduce data movement and compute matrix-vector multiplication (MVM) in analog. It can be well used to accelerate HGNNs. However, the complex metapath-based aggregation of HGNNs makes it challenging to efficiently utilize the parallelism of ReRAM and vertices data reuse. To this end, we propose ChainPIM, the first ReRAM-based processing-in-memory accelerator for HGNNs featuring high-computing parallelism and vertices data reuse. Specifically, we introduce R-chain, which is based on a chain structure to build related metapath instances together. We can efficiently reuse vertices through R-chain and process different R-chains in parallel. Then, we further design an efficient storage format for storing R-chains, which reduces a lot of repeated vertices storage. Finally, a specialized ReRAM-based architecture is developed to pipeline different types of aggregations in HGNNs, fully exploiting the huge potential of multilevel parallelism in HGNNs. Our experiments show that ChainPIM achieves an average memory space reduction of 47.86% and performance improvement by$128.29\times $compared to NVIDIA Tesla V100 GPU.

Fake news detection on social media is crucial to purifying the online environment and protecting public safety. Many existing methods explore the news propagation structures through graph neural networks (GNNs) to determine the truthfulness of news. End-to-end supervised GNNs notoriously depend on large amounts of labels. Recently, self-supervised graph pretraining has been a promising solution to alleviate the dependence on labels. However, the application of graph pretraining in fake news detection still suffers from two challenges: 1) the missing and unreliable interactions intrinsic in the news propagation structures seriously damage the pretraining performance. 2) There is an inherent gap between pretraining and downstream fake news detection tasks due to inconsistency in optimization objectives, which hinders the efficient transfer of pretrained prior knowledge and causes suboptimal detection results. To address the above two challenges, we propose RGCP, a structure redefined graph pretraining with contrastive prompting for fake news detection. Specifically, we design a propagation structure refinement module that adds potential implicit interactions and removes noisy interactions according to the connection probabilities between posts estimated under the guidance of self-supervised contrastive learning. Thereby, the redefined structures provide reliable news propagation patterns to generate robust pretrained news representations. Moreover, we propose a novel prompt tuning based on the contrastive learning module that reformulates the downstream fake news detection task in a similar form as the graph contrastive pretraining, bridging the optimization objective gap. The extensive experiments on benchmark datasets demonstrate the superiority of RGCP, achieving an average improvement of 10.15% in few-shot classification.

Self-supervised learning (SSL) has become one of the most popular learning paradigms and has achieved remarkable success in the graph field. Recently, a series of pre-training studies on heterogeneous graphs (HGs) using SSL have been proposed considering the heterogeneity of real-world graph data. However, verification of the robustness of heterogeneous graph pre-training is still a research gap. Most existing researches focus on supervised attacks on graphs, which are limited to a specific scenario and will not work when labels are not available. In this paper, we propose a novel unsupervised heterogeneous graph rewriting attack via node clustering (HGAC) that can effectively attack HG pre-training models without using labels. Specifically, a heterogeneous edge rewriting strategy is designed to ensure the rationality and concealment of the attacks. Then, a tailored heterogeneous graph contrastive learning (HGCL) is used as a surrogate model. Moreover, we leverage node clustering results of the clean HGs as the pseudo-labels to guide the optimization of structural attacks. Extensive experiments exhibit powerful attack performances of our HGAC on various downstream tasks (i.e., node classification, node clustering, metapath prediction, and visualization) under poisoning attack and evasion attack.

Document-level entity relation extraction is an important task in the field of natural language processing, which plays an important role in semantic understanding and knowledge graph construction. However, existing deep neural networks and graph neural networks models are limited by their performance and parameters number, which can not capture global semantics and have poor generalization ability. Furthermore, existing methods employing large language model for entity relation extraction do not establish good relationships among multi-tasks of entity relation extraction task, resulting in more information can not be effectively shared and transmitted between tasks. In addition, previous approaches can not effectively eliminate false entities and relationships. To solve these problems, we propose an entity relation extraction framework based on large language model and multi-tasks iterative prompt engineering. In our model, we design an iterative prompt engineering, which can better establish the relationship among multi-tasks, and ensure every task to obtain the optimal results. Moreover, we design semantic merging, group disambiguation and self-verification modules to eliminate the false entity relations and noise nodes. Additionally, we design summary prompts to provide sufficient global semantics for better text segmentation. Finally, we evaluated our model on wikiann, wikineural, ACE2005, CoNLL2003, CoNLL2004, and SciERC datasets and compared it with other baseline models.

Personalized news recommendation is the process of predicting the relevance of news to users and recommending news to user to fulfill their information needs. However, existing news recommendation methods extract semantic information from users and candidate news respectively, ignoring semantic interaction information between users and candidate news. Furthermore, previous models only use same node types for message passing, ignoring different characteristics and topology between different node types. In addition, existing methods learn news representations through text representations, ignoring semantic correlation information between entity relationships and texts. To solve these problems, we propose a personalized news recommendation model named CoHG. In our model, we design a collaborative fusion module to obtain semantic interaction information through interacting user history news with candidate news. Furthermore, we design a heterogeneous gated graph neural network that maps different node types into a same space to extract higher-order information in user graphs for message passing. Moreover, we design an enhanced relevant attention module to enhance semantic correlation information of text content by aggregating text representation and entity representation into a unified representation. Finally, we conducted experiments on MIND and Adressa datasets to compare with other baseline models.

Casting sequential recommendation (SR) as a reinforcement learning (RL) problem is promising and some RL-based methods have been proposed for SR. However, these models are sub-optimal due to the following limitations: (a) they fail to leverage the supervision signals in the RL training to capture users’ explicit preferences, leading to slow convergence; and (b) they do not utilize auxiliary information (e.g., knowledge graph) to avoid blindness when exploring users’ potential interests. To address the above-mentioned limitations, we propose a multiplex information-guided RL model (MELOD), which employs a novel RL training framework with Teach and Explore components for SR. We adopt a Teach component to accurately capture users’ explicit preferences and speed up RL convergence. Meanwhile, we design a dynamic intent induction network (DIIN) as a policy function to generate diverse predictions. We utilize the DIIN for the Explore component to mine users’ potential interests by conducting a sequential and knowledge information joint-guided exploration. Moreover, a sequential and knowledge-aware reward function is designed to achieve stable RL training. These components significantly improve MELOD’s performance and convergence against existing RL algorithms to achieve effectiveness and efficiency. Experimental results on seven real-world datasets show that our model significantly outperforms state-of-the-art methods.

In this paper, an image coding algorithm based on directional lifting wavelet transform (DLWT) and universal trellis coded quantization (UTCQ) is presented, and the coding performance is evaluated with multi-scale structural similarity index (MSSIM) and peak signal-to-noise ratio (PSNR). Compared with discrete wavelet transform (DWT), DLWT can provide an efficient representation of edges, but shows a similar ability in representing the smooth area. In order to alleviate blurring artifacts in the smooth area, UTCQ is adopted to quantizing the wavelet coefficients. Experimental results show that the proposed algorithm has the best MSSIM performance among the compared algorithms (including JPEG2000), and its decoded images at low bit-rate are visually more appealing in both edges and smooth areas. The experimental results also show that UTCQ does perform better than scalar quantization (SQ) in MSSIM and improves the subjective visual quality, although it is not necessary better than SQ in PSNR.