Traditional Convolutional Neural Networks (CNNs) can efficiently acquire local features, while ViT uses a Transformer structure that captures global contextual information. In this paper, a new high-performance lightw...
详细信息
ISBN:
(纸本)9798350359329;9798350359312
Traditional Convolutional Neural Networks (CNNs) can efficiently acquire local features, while ViT uses a Transformer structure that captures global contextual information. In this paper, a new high-performance lightweight target detector is designed by combining the respective advantages of CNN and ViT: ShuffleViTNet. In this paper, we improve the VIT Block in MobileVIT, firstly, we use channel-by-channel convolution to replace the normal convolution to reduce the memory access cost (MAC), number of parameters (Parmas) and floating-point operations (FLOPs) in the model inference. The 1x1 convolution is then used to fuse the local features learned by the convolution with the global features learned by the Transofmer, which further lightens the model and allows for easier modification of the number of channels and migration to other lightweight neural networks. Finally a lightweight network friendly to mobile devices is constructed using the improved VIT Block. On the ImageNet1k dataset, ShuffleViTNet achieves 75.2% Top-1 accuracy, the model's floating-point operation is 536.5 MFLOPs, which is 44.0% less compared to MobileViT-XS, and the cost of memory accesses for model inference is saved by about 25.6%, and the required memory is saved by about 66.37%. For the object detection task on the MS-COCO dataset, ShuffleViTNet obtains higher mAP accuracy. The experimental results show that ShuffleViTNet is a high-performance lightweight network and can be well applied to downstream tasks.
In recent years, some deep neural network models like Recurrent Neural Network (RNN), transformer and Bert have been applied to sequence recommendation. It aims to obtain dynamic preference features from recorded user...
详细信息
ISBN:
(纸本)9798400718144
In recent years, some deep neural network models like Recurrent Neural Network (RNN), transformer and Bert have been applied to sequence recommendation. It aims to obtain dynamic preference features from recorded user behavior data, so as to achieve accurate recommendations. However, the user's interaction history inevitably contains a lot of noise information, which is easy to mislead the training of the model. In order to address this issue, we propose a novel simplified Bert filter denoising sequence recommendation model (LightBertR). LightBertR adds a filter layer Bertf with a learnable filter to Bert, which adaptively attenuates noise information by converting time domain and frequency domain information to each other. The incorporation of the Bert filter layer facilitates the model's capacity to discern the intrinsic characteristics of the denoised interaction sequence, thereby enhancing the efficacy of the sequence recommendation model. Furthermore, LightBertR reduces the number of transformer layers in Bert model, markedly diminishes the training parameters of the model, and expedites model training. Extensive experiments on two real datasets demonstrate that LightBertR outperforms other Bert-based methods and numerous baseline methods in the recommendation field.
Graph convolutional networks have gained traction in recommender systems recently, addressing issues likematrix sparsity. LightGCN simplifies models to avoid overfitting and improve generalization. However, it only co...
详细信息
ISBN:
(纸本)9789819755875;9789819755882
Graph convolutional networks have gained traction in recommender systems recently, addressing issues likematrix sparsity. LightGCN simplifies models to avoid overfitting and improve generalization. However, it only considers single behavior, neglecting the impact of multiple behaviors on user preferences. Hence, we propose a multi-behavior recommender based on lightweight graph convolution. We construct a heterogeneous graph capturing various user-item interactions and design a heterogeneous graph attention network. User embeddings from the graph neural network are mapped to different behaviors, enhancing user information mining. Multi-task training enhances model performance, as evidenced by superior results compared to LightGCN and NGCF across multiple datasets.
Sequence recommendation systems usually learn users personalized preferences based on historical behavior sequences of users. Previous methods often use rich item interaction informations combined with context to mine...
详细信息
ISBN:
(纸本)9789819757787;9789819757794
Sequence recommendation systems usually learn users personalized preferences based on historical behavior sequences of users. Previous methods often use rich item interaction informations combined with context to mine user sequential patterns. However, the time decay and dynamic evolution of users preferences are rarely taken into consideration. In this paper, we propose TDH4Rec(Time-aware Dual-kernel Hawkes process for sequential recommendation). First, we establish a time-aware attention embedding module. We utilize time sensitivity and importance of interactions to capture the time dependence and frequency dependence of items during the interaction process. Secondly, we design a Hawkes process based dual-kernel learning module. The Hawkes interaction kernel function with historical interaction effect and the Hawkes time kernel function with time decay effect are designed to capture the influence of historical interaction and the influence of time change of interacted items, respectively. Finally, extensive experiments on three real-world datasets show the efficacy of our model compared with conventional methods and state-of-the-art methods.
Previous studies on user-item interaction graphs have typically concentrated on simple interactions, often overlooking the significant role of user intent in shaping these interactions. While some recent research has ...
详细信息
Previous studies on user-item interaction graphs have typically concentrated on simple interactions, often overlooking the significant role of user intent in shaping these interactions. While some recent research has explored intent relationships to enhance modeling, these approaches mainly focus on user preferences derived from interactions, ignoring the knowledge information with good interpretation in knowledge graphs. In addition, some recent work usually utilize undenoised graph structure information to learn the node representations, which introduces plenty of noise and impedes the well learning of users' preference. In this paper, we utilize the rich interpretable knowledge information in the knowledge graph to design a novel knowledge-driven hierarchical intent modeling framework called KHIM. The focus is on designing a hierarchical user intent modeling process and an intent-based multi-view contrastive learning mechanism. The former extracts both the popular and personalized preferences of users from attribute tuples within the knowledge graph at the global-level and local-level, respectively. For global level, we capture the user's true intents from positive view and augment the positive intent representation with negative view. While the latter generates high-quality user and item representations through multi-level cross-view contrastive learning. Two data augmentation strategies are designed during the contrastive process to mitigate the effects of noise in the learning process. Additionally, we also designed a neighbor filtering strategy based on semantic view to obtain more neighbor semantic information of user and item nodes, so as to further improve the recommendation performance. Experimental results on three benchmark datasets demonstrate that KHIM significantly outperforms various state-of-the-art approaches, highlighting its effectiveness in leveraging knowledge graph information for better recommendations.
The paper proposes a multi-factor interlinked POI recommendation model called MFIP. Extracting user similarity for user-sensitive implicit modeling to enrich user representation. Using contextual information such as s...
详细信息
ISBN:
(数字)9781665480185
ISBN:
(纸本)9781665480185
The paper proposes a multi-factor interlinked POI recommendation model called MFIP. Extracting user similarity for user-sensitive implicit modeling to enrich user representation. Using contextual information such as sequential, geographical and social to construct a POI recommendation model with collaborative influence o f multi-factor, a lleviating data sparsity. A novel multi-factor interlinked strategy(FIS) is proposed that can dynamically adjust the user preferences of different factors to obtain the comprehensive impact of user personalization. In addition, we propose a active area selection algorithm based on segmentation to model the geographical information more effectively. Finally, we conduct a comprehensive performance evaluation for MFIP on two large-scale real world check-in datasets collected from Gowalla and Yelp. Experimental results show that MFIP achieves significantly superior recommendation performance compared to other state-of-the-art POI recommendation models.
Recently, the recommended method based on the Knowledge Graph (KG) has become a hot research topic in modern recommendation systems. Most researchers use assistive information such as entity attributes in KG to improv...
详细信息
ISBN:
(纸本)9783031159312;9783031159305
Recently, the recommended method based on the Knowledge Graph (KG) has become a hot research topic in modern recommendation systems. Most researchers use assistive information such as entity attributes in KG to improve recommendation performance and alleviate Collaborative Filtering (CF) sparsity and cold start problems. The most recent technical trend is to develop end-to-end models based on the Graph Convolutional Network (GCN). In this paper, we propose a Knowledge Graph Bidirectional Interaction Graph Convolution Network for recommendation (KBGCN). This method is used to refine the embedded representation of node by recursively delivering messages from the neighbors (attributes or items) of the node (entity) and applies the knowledge aware attention mechanism to distinguish the contributions of different neighbors based of the same node. It uses neighbors of each entity in KG as the view of this entity, which can be extended by expanding the view of Multi-hop neighbors to mine high-order connectivity information existing in KG automatically. We apply the proposed method to three real-world datasets. KBGCN is better than seven KG-based baselines in recommendation accuracy and the two state-of-the-art GCN-based recommendations frameworks.
Traditional approaches focus on an individual item of most interest to users. However, in most realistic scenarios, the platform needs to recommend a group of items at one time for users' convenience, called bundl...
详细信息
ISBN:
(纸本)9783031192135;9783031192142
Traditional approaches focus on an individual item of most interest to users. However, in most realistic scenarios, the platform needs to recommend a group of items at one time for users' convenience, called bundle recommendation. e.g., a music playlist containing multiple songs. The existing bundle recommendations usually use manual methods to artificially build bundles for different items, ignoring the obtained bundles and the potential relationships among the items in the bundle, especially the relationships between bundles. Therefore, how integrating multiple complex interactions into bundles and obtaining high-quality bundle recommendation is an important problem. To solve the problem, we propose a novel model named IMBR (short for Interactive Multi-Relation Bundle Recommendation with Graph Neural Network). Firstly, we construct a multi-relation interaction graph to capture the interaction relation from the user view. At the same time, we get bundle subordination relation from the item view. They can obtain richer representations of users, bundles, and items. Secondly, we design a bundle frequent term constraint algorithm (BFTC) to constrain the composition of items in a bundle and pay attention to the similarity between bundles. Finally, we leverage a multi-task learning framework to capture user personalized preferences to improve the performance of bundle recommendation. Extensive experiments on two real-world datasets with different scales show that our method can significantly outperform various baseline approaches.
For the node classification problem, the traditional graph convolutional network (GCN) and many of its variants achieve the best results at shallow layers, especially for sparse graphs, which do not take full advantag...
详细信息
ISBN:
(数字)9781728186719
ISBN:
(纸本)9781728186719
For the node classification problem, the traditional graph convolutional network (GCN) and many of its variants achieve the best results at shallow layers, especially for sparse graphs, which do not take full advantage of the higher-order neighbor information of the nodes in the graph. However, multiple stacked graph convolutional networks will suffer from the over-smoothing problem, resulting in ineffective differentiation between different classes of nodes in the graph. To address this problem, we propose an adaptive graph convolutional network based on decouple and residuals (ADR-GCN) to alleviate the over-smoothing problem. The model first obtains the initial representation of nodes using autoencoder and then decouples the representation transformation of nodes and feature propagation. In addition, we add initial residuals to the feature propagation of nodes and adaptively selects the appropriate local and global information of each node to obtain node representations containing rich information. Finally, we use a softmax classifier to generate the final node prediction. The experimental results on seven datasets show that the classification accuracy of ADR-GCN improves 1.3%-4.1% compared with GCN, which shows that the model can better alleviate over-smoothing.
Recently, graph neural networks (GNNs) have achieved significant success in many graph-based tasks. However, most GNNs are inherently restricted by over-smoothing, which limits performance improvement. In this paper, ...
详细信息
ISBN:
(纸本)9783031251573;9783031251580
Recently, graph neural networks (GNNs) have achieved significant success in many graph-based tasks. However, most GNNs are inherently restricted by over-smoothing, which limits performance improvement. In this paper, we propose an Enhanced Attribute-aware and Structure-constrained Graph Convolutional Network (EAS-GCN). Specifically, EAS-GCN first uses degree prediction to incorporate graph local structure information into autoencoder-specific representation. A delivery mechanism is then designed to pass the autoencoder-specific representation to the corresponding GCN layer. Autoencoder mainly assists GCN in learning enhanced attribute information, and node degree prediction assists GCN in learning local structure information. Furthermore, we theoretically analyze autoencoder could help alleviate the over-smoothing in GCN. Experimental results show that EAS-GCN enjoys high accuracy on the node classification task and can better alleviate over-smoothing.
暂无评论