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.

As user behavior data becomes increasingly scattered across different platforms, achieving cross-domain knowledge fusion while preserving privacy has become a critical issue in recommender systems. Existing Privacy-Preserving Cross-Domain Recommendation (PPCDR) methods usually rely on overlapping users or items as a bridge, making them inapplicable to non-overlapping scenarios. They also suffer from limitations in the collaborative modeling of global and local semantics. To this end, this paper proposes a Federated Cross-domain Recommendation method with deep knowledge Fusion (FedCRF). Using textual semantics as the cross-domain bridge, FedCRF achieves cross-domain knowledge transfer via federated semantic learning under the non-overlapping scenario. Specifically, FedCRF constructs global semantic clusters on the server side to extract shared semantic information, and designs a Fine-Grained Semantic Adaptation and Transfer (FGSAT) module on the client side to dynamically adapt to local data distributions and alleviate cross-domain distribution shift. Meanwhile, it builds a semantic graph based on textual features to learn representations that integrate both structural and semantic information, and introduces contrastive learning constraints between global and local semantic representations to enhance semantic consistency and promote deep knowledge fusion. In this framework, only item semantic representations are shared, while user interaction data remains locally stored, effectively mitigating privacy leakage risks. Experimental results on multiple real-world datasets show that FedCRF significantly outperforms existing methods in terms of Recall@20 and NDCG@20, validating its effectiveness and superiority in non-overlapping cross-domain recommendation scenarios.

Mining high utility repeated negative sequential patterns (HURNSPs) from data streams is an important method for data stream analysis. However, the existing methods in this topic don’t consider negative events and repeated events, which results in weak decision-making effectiveness. So in this paper, we propose an effective algorithm DS_HURNSP for mining HURNSPs from data streams based on a sliding window model. First, we propose an effective utility-list prefix tree structure to store high utility repeated positive sequential patterns (HURPSPs). Second, we construct a utility mapping set based on a hash-table structure to enable rapid querying of HURPSPs information. Finally, we propose a two-stage computation method to compute the utility of high utility repeated negative sequential candidates (HURNSCs) by mapping them to the set of HURPSPs, avoiding rescanning database. Extensive experiments on six datasets show that the DS_HURNSP algorithm generates tens to thousands of times as many HURNSPs as the baseline method, and reduces the average runtime by more than half.

Multimedia online platforms such as Amazon and TikTok have significantly enhanced their personalized Multi-Modal Recommendation (MMRec) systems by incorporating multi-modal content. These various modalities provide intuitive semantics that improve the accuracy of user preference modeling. Despite these advancements, two major challenges remain in MMRec: First, the inclusion of multi-modal side information often introduces biases and redundancies, distorting modality-interaction dependencies and failing to accurately reflect genuine user preferences. Second, existing graph-based methods are still affected by over-smoothing and inefficiency problems. To address these challenges, we propose an approach called Prompt-wise Multi-Modal Knowledge Distillation (PMMKD) for MMRec. Specifically, we introduce continuous soft prompts to bridge the semantic gap and account for cross-modal noise. We then design multiple knowledge distillation strategies to distill implicit knowledge from GNNs to MLP, alleviating over-smoothing and scalability issues. Experiments with real-world data demonstrate that PMMKD surpasses current methods. Ablation studies validate the significance of the crucial components, and supplementary tests underscore both the efficiency and effectiveness of PMMKD.

Cognitive diagnosis serves as a core method in intelligent education, aiming to learn students’ proficiency in knowledge concepts by analyzing the records of their exercises. Existing approaches perform well in individual-level cognitive diagnosis, they face limitations in group-level cognitive diagnosis tasks and are not suitable for collaborative group scenarios. However, existing group-level cognitive diagnosis methods still have the following limitations: 1) They fail to fully model the complex relationships among groups, students, and exercises, resulting in the learning of group representations lacking comprehensive capture of multi-dimensional interaction information; 2) they neglect the complementary information between student-level interactions and exercise-level group preferences, leading to shallow group representations. To overcome the previously discussed limitations, we propose a Graph Learning with Bidirectional Distillation framework for Group Cognitive Diagnosis (GDGCD). Specifically, to address Limitation 1, we use hypergraphs to model groups and connect students and exercises and design a novel student-level hypergraph neural network to aggregate node information, forming fine-grained student-level group representations. Meanwhile, in the exercise-level bipartite graph, groups are connected to exercises. A distinctive exercise-level representation of the group is constructed via the graph neural network. To handle Limitation 2, we introduce a mutual distillation model to integrate interaction features from both perspectives, enabling the generation of rich group representations. We perform extensive experiments on four real-world datasets, and the experimental results demonstrate the effectiveness of our GDGCD method over several recent SOTA approaches on RMSE and MAE metrics.

High-utility repeated negative sequential patterns (HURNSPs) mining plays a key role in behavioral analysis and user preference mining. However, existing HUSPM mining methods do not consider the importance of repeated negative sequential patterns (RNSPs) or high-utility negative sequential patterns (HUNSPs), which pose the following challenges for HURNSPs mining: (1) Lack of an effective method for calculating the utility of high-utility repeated positive sequential patterns (HURPSPs), (2) Lack of an effective method for calculating the utility value of high-utility repeated negative sequential candidate patterns (HURNSCs). To solve the above challenges, this paper proposes an effective algorithm, HU-RNSP, for mining HURNSPs. First, an algorithm, called HURSpan, is proposed to mine HURPSPs by integrating RNSP and HUSPM into the mining of HURNSPs. Second, an algorithm, NSPGwl, is proposed, which converts HURPSPs into HURNSCs, effectively calculates the utility of HURNSCs, and compares the utility of HURNSCs with a minimum utility threshold to obtain HURNSPs. Experimental results on nine datasets demonstrate that HU-RNSP is more effective than baseline methods in discovering HURNSPs. Additionally, we analyze the impact of data features on HURNSP mining. The results indicate that HU-RNSP demonstrates strong adaptability and computational efficiency across experiments on datasets with varying data factors.

Large Language Models (LLMs) have shown great success in recommender systems. However, the limited and sparse nature of user data often restricts the LLM's ability to effectively model behavior patterns. To address this, existing studies have explored cross-domain solutions by conducting Cross-Domain Recommendation (CDR) tasks. But previous methods typically assume domains are overlapped and can be accessed readily. None of the LLM methods address the privacy-preserving issues in the CDR settings, that is, Privacy-Preserving Cross-Domain Recommendation (PPCDR). Conducting non-overlapping PPCDR with LLM is challenging since: 1) The inability to share user identity or behavioral data across domains impedes effective cross-domain alignment. 2) The heterogeneity of data modalities (e.g., textual vs. ID-based features) across domains complicates knowledge integration. 3) Fusing collaborative filtering signals from traditional recommendation models with LLMs is difficult, as they operate within distinct feature spaces. To address the above issues, we propose SF-UBM, a Semantic-enhanced Federated User Behavior Modeling method as our solution. Specifically, to deal with Challenge 1, we leverage natural language as a universal bridge to connect disjoint domains via a semantic-enhanced federated architecture. Here, text-based item representations are encrypted and shared, while user-specific data remains local. To handle Challenge 2, we design a Fact-counter Knowledge Distillation (FKD) module to integrate domain-agnostic knowledge with domain-specific knowledge, across different data modalities.To tackle Challenge 3, we project pre-learned user preferences and cross-domain item representations into the soft prompt space, aligning behavioral and semantic spaces for effective LLM learning. We conduct extensive experiments on three pairs of real-world domains, and the experimental results demonstrate the effectiveness of SF-UBM compared to the recent SOTA methods. Our code will be publicly available at: https://github.com/Nexus-Yang/SF-UBM\_master.

Large Language Models have shown great success in recommender systems. However, the limited and sparse nature of user data often restricts the LLM's ability to effectively model behavior patterns. To address this, existing studies have explored cross-domain solutions by conducting Cross-Domain Recommendation tasks. But previous methods typically assume domains are overlapped and can be accessed readily. None of the LLM methods address the privacy-preserving issues in the CDR settings, that is, Privacy-Preserving Cross-Domain Recommendation. Conducting non-overlapping PPCDR with LLM is challenging since: 1)The inability to share user identity or behavioral data across domains impedes effective cross-domain alignment. 2)The heterogeneity of data modalities across domains complicates knowledge integration. 3)Fusing collaborative filtering signals from traditional recommendation models with LLMs is difficult, as they operate within distinct feature spaces. To address the above issues, we propose SF-UBM, a Semantic-enhanced Federated User Behavior Modeling method. Specifically, to deal with Challenge 1, we leverage natural language as a universal bridge to connect disjoint domains via a semantic-enhanced federated architecture. Here, text-based item representations are encrypted and shared, while user-specific data remains local. To handle Challenge 2, we design a Fact-counter Knowledge Distillation module to integrate domain-agnostic knowledge with domain-specific knowledge, across different data modalities. To tackle Challenge 3, we project pre-learned user preferences and cross-domain item representations into the soft prompt space, aligning behavioral and semantic spaces for effective LLM learning. We conduct extensive experiments on three pairs of real-world domains, and the experimental results demonstrate the effectiveness of SF-UBM compared to the recent SOTA methods.

Traditional recommendation models primarily rely on display feedback and typically utilize a single type of user-item interaction data, which often results in significant data sparsity issues. In contrast, multi-behavioral recommendation models leverage various behaviors such as browsing, favoriting, and other interactions. These additional behaviors help improve the prediction of user-item interactions. Existing multi-behavioral recommendation methods often overlook the potential factors influencing multi-behavioral interactions and the differences between various behavior types. In this study, we introduce a multi-behavioral recommendation algorithm utilizing decoupled graph convolution (MBR-DGC), which effectively mitigates the data sparsity of the target behaviors and improves recommender system performance by capturing the differences between the semantics of different behaviors. Specifically, we construct multiple non-overlapping independent isomorphic graphs and separate potential factors affecting the interactions among users, items, and behaviors using decoupled convolutional networks to reconstruct the node features of users in different behaviors. Afterwards, multi-behavioral features of users are aggregated using contrastive learning to achieve personalized multi-behavioral information aggregation. Experimental results on multiple datasets show that MBR-DGC effectively leverages multi-behavioral data, significantly enhancing recommendation performance compared to other state-of-the-art methods.

With the continuous development of recommender systems, leveraging users' future interactions to capture the evolution of preferences in sequential behaviors has become a key focus in Sequential Recommendation (SR). Although current approaches have made notable progress, several important challenges persist: 1) An excessive dependence on historical interactions limits the ability to effectively model the evolution of user preferences. 2) There is a lack of mechanisms to capture the dynamic transition path between historical and future behaviors. 3) Existing denoising techniques often fail to fully consider the semantic information in behavior denoising. To address these issues, we propose an SR framework-Future Distillation Recommendation (FDistRec), which introduces the following strategies: To tackle Limitation 1, we incorporate future interaction data to help alleviate the constraints posed by relying on past behaviors. To address Limitation 2, we adopt a distillation framework along with the Causal Mixture-of-Experts (CMoE) module, effectively capturing the evolution and shift path in user interests. To overcome Limitation 3, we present a Masked Denoising Variational Auto Encoder (MDVAE) model that applies semantic-aware masking to perturb user behaviors, facilitating generative preference modeling and noise-aware interest recovery. Experiments on five real-world datasets demonstrate the superior performance of FDistRec compared to the SOTA baselines on SR.

Unified Large Language Models (LLMs) have transformed diverse recommendation tasks into a shared “text-to-text” paradigm. However, existing methods primarily face two challenges: 1) joint fine-tuning across highly heterogeneous tasks (e.g., discriminative sequential recommendation versus generative explanation generation) often suffers from the “seesaw effect” due to severe gradient conflicts; and 2) parameter-efficient methods struggle to isolate these task-specific conflicts while preserving the collaborative sharing of underlying general recommendation knowledge. To address these limitations, we propose Rec-MoELoRA, a hybrid fine-tuning framework that seamlessly integrates the mixture-of-experts architecture with low-rank adaptation. Specifically, we devise a soft-decoupling strategy that maintains the updates of most backbone network layers to continuously absorb globally shared knowledge, while exclusively introducing task-motivated low-rank experts into the self-attention layers for physical task isolation. Furthermore, to effectively overcome the prevalent “expert collapse” dilemma, we incorporate a dual regularization strategy of diversity and entropy. Combined with an asymmetric structural prior (e.g., allocating three experts for two tasks), this forces the model to spontaneously evolve a “Shared-Specific” representation pattern. Extensive experiments on real-world e-commerce datasets (Amazon Sports and Beauty) demonstrate that Rec-MoELoRA significantly outperforms the full fine-tuning baseline across core ranking and text generation metrics while introducing only a marginal number of additional parameters (e.g., HR@10 and BLEU-4 improve by 4.3% and 3.7% on the Sports dataset, respectively).

Sequential recommendation aims to predict a user's next interaction based on their historical behavior. While diffusion models have advanced sequential recommendation by capturing user uncertainty, they suffer from a fundamental challenge: distribution misalignment. This issue arises when models compress a user's multi-dimensional interests into a single representation for conditional guidance, causing the generated distribution to deviate from true user preferences. We propose DSCDO, a framework that resolves this challenge through a Dual-correction of Semantic Control and Distribution Optimization. It synergistically integrates two components: (1) an orthogonal semantic decomposition module that projects the user's interaction history into non-interfering semantic subspaces, thereby preventing the collapse of multifaceted user preferences into a single, over-generalized representation, and (2) an gradient enhance denoising network that employs a score-matching objective to explicitly model the gradient of the true preference distribution, thereby applying a precise, single-step adjustment to the initial prediction to directly fix the distribution misalignment. Experiments on four real-world datasets show DSCDO achieves 2.3%-16.75% improvements in HR@20 and 1.9%-9.14% gains in NDCG@20 over state-of-the-art baselines.1

Cross-domain Recommendation (CR) has been extensively studied in recent years to alleviate the data sparsity issue in recommender systems by utilizing different domain information. In this work, we focus on the more general Non-overlapping Cross-domain Sequential Recommendation (NCSR) scenario. NCSR is challenging because there are no overlapped entities (e.g., users and items) between domains, and there is only users' implicit feedback and no content information. Previous CR methods cannot solve NCSR well, since (1) they either need extra content to align domains or need explicit domain alignment constraints to reduce the domain discrepancy from domain-invariant features, (2) they pay more attention to users' explicit feedback (i.e., users' rating data) and cannot well capture their sequential interaction patterns, (3) they usually do a single-target cross-domain recommendation task and seldom investigate the dual-target ones. Considering the above challenges, we propose Prompt Learning-based Cross-domain Recommender (PLCR), an automated prompting-based recommendation framework for the NCSR task. Specifically, to address the challenge (1), PLCR resorts to learning domain-invariant and domain-specific representations via its prompt learning component, where the domain alignment constraint is discarded. For challenges (2) and (3), PLCR introduces a pre-trained sequence encoder to learn users' sequential interaction patterns, and conducts a dual-learning target with a separation constraint to enhance recommendations in both domains. Our empirical study on two sub-collections of Amazon demonstrates the advance of PLCR compared with some related SOTA methods.

In the evolving landscape of recommender systems, the challenge of effectively conducting privacy-preserving Cross-Domain Recommendation (CDR), especially under strict non-overlapping constraints, has emerged as a key focus. Despite extensive research has made significant progress, several limitations still exist: 1) Previous semantic-based methods fail to deeply exploit rich textual information, since they quantize the text into codes, losing its original rich semantics. 2) The current solution solely relies on the text-modality, while the synergistic effects with the ID-modality are ignored. 3) Existing studies do not consider the impact of irrelevant semantic features, leading to inaccurate semantic representation. To address these challenges, we introduce federated semantic learning and devise FFMSR as our solution. For Limitation 1, we locally learn items'semantic encodings from their original texts by a multi-layer semantic encoder, and then cluster them on the server to facilitate the transfer of semantic knowledge between domains. To tackle Limitation 2, we integrate both ID and Text modalities on the clients, and utilize them to learn different aspects of items. To handle Limitation 3, a Fast Fourier Transform (FFT)-based filter and a gating mechanism are developed to alleviate the impact of irrelevant semantic information in the local model. We conduct extensive experiments on two real-world datasets, and the results demonstrate the superiority of our FFMSR method over other SOTA methods. Our source codes are publicly available at: https://github.com/Sapphire-star/FFMSR.

The Shared-Account Recommendation (SAR) aims to accurately identify and accommodate the varied preferences of multiple users sharing a single account by analyzing their aggregated interactions. SAR faces challenges in preference identification when multiple users share an account. Existing Shared-Account Modeling (SAM) methods assume overly simplistic conditions and overlook the robustness of representations, leading to inaccurate embeddings that are susceptible to disturbances. To address limitations in existing SAR methods, we introduce the Contrastive Clustering User Identification Network (CCUI-Net) framework to enhance SAR. This framework employs graph-based transformations and node representation learning to refine user embeddings, utilizes hierarchical contrastive clustering for improved user identification and robustness against data noise, and leverages an attention mechanism to dynamically balance contributions from various users. These innovations significantly boost the precision and reliability of recommendations. Experimental results across four domains from the HVIDEO and HAMAZON datasets (E-domain and V-domain in HVIDEO, M-domain and B-domain in HAMAZON) demonstrate that CCUI-Net exceeds the performance of many existing available methods on the metrics MRR@5, MRR@20, Recall@5, and Recall@20. Specifically, the improvements in the M-domain and B-domain for Recall@5 and Recall@20 are 14.64%, 8.55%, 18.67%, and 9.59% respectively.

Non-overlapping Cross-domain Sequential Recommendation (NCSR) is the task that focuses on domain knowledge transfer without overlapping entities. Compared with traditional Cross-domain Sequential Recommendation (CSR), NCSR poses several challenges: 1) NCSR methods often rely on explicit item IDs, overlooking semantic information among entities. 2) Existing CSR mainly relies on domain alignment for knowledge transfer, risking semantic loss during alignment. 3) Most previous studies do not consider the many-to-one characteristic, which is challenging because of the utilization of multiple source domains. Given the above challenges, we introduce the prompt learning technique for Many-to-one Non-overlapping Cross-domain Sequential Recommendation (MNCSR) and propose a Text-enhanced Co-attention Prompt Learning Paradigm (TCPLP). Specifically, we capture semantic meanings by representing items through text rather than IDs, leveraging natural language universality to facilitate cross-domain knowledge transfer. Unlike prior works that need to conduct domain alignment, we directly learn transferable domain information, where two types of prompts, i.e., domain-shared and domain-specific prompts, are devised, with a co-attention-based network for prompt encoding. Then, we develop a two-stage learning strategy, i.e., pre-train prompt-tuning paradigm, for domain knowledge pre-learning and transferring, respectively. We conduct extensive experiments on three datasets and the experimental results demonstrate the superiority of our TCPLP. Our source codes have been publicly released.

Few-shot learning can construct neural architectures endowed with rapid task adaptation capabilities under limited labeled data regimes while preserving the generalization efficacy across distribution shifts. Meta-learning frameworks that adopt bi-level optimization paradigms have emerged as predominant solutions for few-shot learning problems owing to their architectural parsimony and parameter efficiency. However, inherent hierarchical optimization dynamics (outer-loop meta-optimization over task-specific loss landscapes and inner-loop, gradient-based task adaptation) cause computational pathology via second-order gradient backpropagation across inner-loop trajectories. This induces sensitivity degradation in parameter initialization and gradient propagation instability, particularly under cross-task distributional disparities. To address these limitations, we propose Meta-DDA, a novel meta-learning framework that substitutes conventional gradient descent in the inner-loop with diffusion-based denoising trajectories. Meta-DDA effectively circumvents the numerical instability of traditional meta-learning, second-order gradient backpropagation by reconstructing inner-loop gradient optimization as a denoising trajectory of the task conditions. This significantly reduces the sensitivity to initialization by utilizing noise scheduling with progressive parameter updating of Gaussian prior to achieve more stable and robust optimization in scenarios with few samples. Furthermore, we develop dual data augmentation strategies that are compatible with the bi-level architecture: (1) task-level augmentor at the meta-level stage mitigates excessive parameter updates caused by task difficulty variance; (2) sample-level augmentor at the base learner stage augments task-specific feature learning. Extensive experiments on four text classification datasets and four intent recognition datasets demonstrate the superior performance of Meta-DDA with markedly improved cross-domain generalization.

The rapid increase in encrypted network traffic has made detecting malicious activities a critical challenge in network management, attracting significant research attention. However, most existing methods focus primarily on flow-based features, often neglecting the inherent structural heterogeneity and dynamic temporal variations in encrypted traffic, which limits their effectiveness in capturing the evolving nature of malicious activities. To address the aforementioned issue, we present a novel Dynamic Heterogeneous Information Network framework, DMT-DHIN, for Detecting Malicious Encrypted Traffic. DMT-DHIN constructs dynamic heterogeneous graphs by partitioning traffic into time slices, with nodes representing network entities (e.g., packets, protocols, IPs, and ports) and edges capturing their interactions. A heterogeneous graph attention mechanism effectively captures spatial dependencies among different node types. At the same time, the Transformer captures the temporal evolution of node features, enabling DMT-DHIN to identify dynamic patterns in encrypted traffic. We validate the effectiveness of DMT-DHIN by conducting evaluations on three datasets: CICIDS2017, USTC-TFC2016, and CICIoT2023. The results demonstrate that DMT-DHIN surpasses existing state-of-the-art methods, delivering marked enhancements in detection accuracy and F1-score, highlighting its superior ability to identify malicious encrypted traffic.

In previous collaboration studies, a majority of them concentrate on examining cooperation models, often overlooking the pivotal role played by a Top Scientist (TS) in scientific advancements. As far as my knowledge extends, only one relevant work delves into the correlation between innovation and collaboration with TSs, and no research has explored this relationship from a causal perspective. More precisely, previous studies suffer from several limitations in their examination of this topic: 1) Existing studies on Papers' Novelty (PN) primarily focus on calculating methods, with limited exploration of its relationship with scientific cooperation. 2) Research that has explored the link between collaboration with TSs and output innovation often adopts a correlational perspective, lacking a causal analysis that could correct for potential confounding factors. 3) Previous methodologies overlook the attributes of citation networks as potential confounding factors, a crucial consideration in identifying identical papers in causal analyses. 4) The impact disciplinary diversity of papers on the innovation output when collaborating with TSs is often overlooked in prior research. To address these limitations, we conduct a causal analysis of publications in three subfields of computer science from the Web of Science (WoS) database to demonstrate the impact of collaborating with TSs on PN. Specifically, to tackle Limitations 1) and 2), we employ PN as a metric to assess the quality of academic output and explore its causal relationship with collaborating with TSs using the Propensity Score Matching (PSM) method. To address Limitation 3), we comprehensively consider potential confounding factors influencing PSM matching by further incorporating the attributes of citation networks, thereby minimizing selection bias. To deal with Limitation 4), we not only focus on the overall treatment effect but also delve into the treatment effect of intra-disciplinary and interdisciplinary collaboration modes. The research findings indicate that the papers collaborating with TSs exhibit lower PN compared to those without the participation of TSs. This suggests that collaboration with TSs may come at the cost of reduced novelty. This discovery prompts profound reflections on scientific collaboration, emphasizing the challenges and trade-offs that may exist in collaboration.

The expressiveness of historical reviews in capturing user preferences has garnered significant attention in recommender systems. However, this technology still has certain limitations. Firstly, irrelevant reviews can introduce noise that may adversely affect the performance of the model. Secondly, existing approaches often assume a flat structure for review features, thus failing to capture the intricate and hierarchical nature of user-item interactions. Thirdly, it is challenging for review-based recommendation models to effectively assess the usefulness of reviews due to sparse supervision signals. To address these challenges, we propose a novel Hierarchical Gating and Discriminative model for rating prediction. Specifically, we introduce a local gating module that utilizes personalized end-to-end differential thresholds to select reviews in a relatively “hard” manner, thereby minimizing the impact of noisy reviews while facilitating model training. Additionally, we incorporate a global gating module to assess the overall usefulness of review signals by estimating the uncertainties inherent in historical reviews. Moreover, we propose a hierarchical discriminative network to develop self-supervision signals at both global and local levels to guide the learning of the hierarchical gating network. Extensive experiments on public datasets have demonstrated the effectiveness of the proposed model, and further investigations provide deep insight into its superiority.