Graph encoding and the attention mechanism enable Graph Transformers (GTs) to extract features of graph structures. However, employing graph encoding and Transformer-based attention mechanism may lead to two defects: high computing complexity and sensitivity to graph structures. Therefore, this paper designs a graph feature extractor with a pure or lightweight attention mechanism that does not rely on graph encoding or feature projection matrices, called Collaborative Filtering Transformer (CFT). CFT uses the pure attention mechanism as the main structure of the encoder and introduces non-interaction information by constructing preference sentences. The core of CFT is that its encoder does not contain any collaborative information, but only plays a role in generating associations between different nodes, so that a node on the graph can notice its non-neighboring nodes. The utilization of collaborative information is only achieved through optimizing the loss function. In addition, through theoretical analysis and experimental verification, we prove that adding path-based graph encoding in CFT has a negative effect on the feature extraction process of the attention mechanism. Furthermore, experiments show that during the optimization process, the proposed pure attention mechanism can always assign higher attention scores to nodes with interactions, while making the attention scores between nodes without interactions approach zero. Finally, our model achieves the best performance when compared with the latest methods on five real-world datasets, and compared to the second-best baseline, the recommendation performance is improved by up to 20.60%. Moreover, CFT achieves considerably high training efficiency across all five datasets, with training time comparable to that of simple matrix factorization-based baselines.

Graph collaborative filtering has made great progress in the recommender systems, but these methods often struggle with the data sparsity issue in real-world recommendation scenarios. To mitigate the effect of data sparsity, graph collaborative filtering incorporates contrastive learning as an auxiliary task to improve model performance. However, existing contrastive learning-based methods generally use a single data augmentation graph to construct the auxiliary contrastive learning task, which has problems such as loss of key information and low robustness. To address these problems, this paper proposes a Perturbation-driven Dual Auxiliary Contrastive Learning for Collaborative Filtering Recommendation (PDACL). PDACL designs structure perturbation and weight perturbation to construct two data augmentation graphs. The Structure Perturbation Augmentation (SPA) graph perturbs the topology of the user-item interaction graph, while the Weight Perturbation Augmentation (WPA) graph reconstructs the implicit feedback unweighted graph into a weighted graph similar to the explicit feedback. These two data augmentation graphs are combined with the user-item interaction graph to construct the dual auxiliary contrastive learning task to extract the self-supervised signals without losing key information and jointly optimize it together with the supervised recommendation task, to alleviate the data sparsity problem and improve the performance. Experimental results on multiple public datasets show that PDACL outperforms numerous benchmark models, demonstrating that the dual-perturbation data augmentation graph in PDACL can overcome the shortcomings of a single data augmentation graph, leading to superior recommendation results. The implementation of our work will be found at https://github.com/zky77/PDACL.

Existing cross-domain few-shot hyperspectral image (HSI) classification methods based meta-learning face two main issues: one is that the optimization of class prototypes is insufficient with a small number of samples, resulting in poor balance between inter-class discriminability and intra-class compactness; the other is the lack of dynamic coordination of different fine-grained domain distributions, leading to limited domain adaptability. To address these issues, this paper proposes an adversarial domain adaptive network based on dynamic class prototype and graph convolution (DPGC) with two core strategies to improve the classification accuracy of cross-domain few-shot HSI classification. The first strategy, dynamic class prototype, optimizes the position distribution of class prototypes by adjusting the weights of real and pseudo-labels, thereby widening inter-class distances and compressing intra-class distances. The second strategy, graph-based domain adversarial strategy, extracts local and global domain features and forces the feature extractor to generate domain-independent features through adversarial training, mitigating performance degradation caused by domain distribution differences. Experiments show that DPGC achieves the best classification performance on three hyperspectral datasets and the code is publicly available at https://github.com/tiwchui/DPGC.

Current hyperspectral image (HSI) classification methods face challenges such as insufficient feature representation, inadequate exploration of sample relations, and loss of boundary details. Convolutional neural networks (CNNs) are also limited by small receptive fields and insufficient attention to global information. Moreover, existing frequency-domain analyses for HSI classification are relatively simple and prone to noise interference. To address these issues, this paper proposes a network based on combined frequency-domain filtering and multi-scale spatial feature communication (CFFMSFC) for HSI classification. The method integrates CNNs with frequency-domain processing to achieve collaborative modeling in frequency and spatial domains, improving classification performance. Specifically, the combined frequency-domain filtering module (CFF) transforms the image into the frequency domain and adaptively fuses details in high-frequency with global information in low-frequency. The frequency feature extraction (FFE) network is built with stacked inverted residual blocks to reduce complexity. The multi-scale spatial feature communication (MSFC) module communicates multi-scale spatial features in both horizontal and vertical directions, and employs a multi-head attention mechanism to better fuse features with different scales. To fully utilize high-level semantic features, gated fusion is employed to integrate different types of features. Experimental results demonstrate that the CFFMSFC outperforms the existing approaches, achieving superior classification performance even with limited labeled samples.

Click-Through Rate (CTR) prediction is a critical task in commercial recommender systems. Feature embedding is the cornerstone of CTR models. Most previous work on CTR prediction to improve the model performance focused on modelling feature interactions. However, the quality of embeddings and the effective information of embeddings carried are overlooked. As a result, performing feature interactions on inferior embeddings will be a bottleneck to capture informative knowledge. To address this problem, we optimize embeddings from the perspective of expanding the weight gap and masking noisy information of the embedding vector. We first propose a novel module named Simple Stacked Gate (SSG), which selects the salient information of feature embeddings and enhances the weight gap between it and other information. Based on this, we further propose an Automatic Stacked Gate (AutoSG) framework, which utilizes Neural Architecture Search (NAS) method to automatically identify and mask noisy information of the embedding vector obtained after the SSG module. Both SSG and AutoSG can be seen as plug-and-play components and added to any existing deep CTR models flexibly. Extensive experiments on four benchmark datasets demonstrate that these two methods can further improve the performance of the backbone CTR models. Even only half of the bit information retained in the optimized embeddings, better performance can still be achieved compared to the original model. The implementation of our work can be found at https://github.com/SG-CTR/Stacked-Gate.

Session-based recommendation (SBR) aims to recommend items based on anonymous behavior sequences. Methods based on graph neural networks always ignore two issues. First, they ignore the multiple user interests hidden in sessions. Second, they do not pay much attention to the impact of the over-smoothing phenomenon in graph neural networks. To address these two problems, this paper proposes a novel multi-interest awareness (MIA) model by modeling multiple interests explicitly for SBR. First, we design a multi-interest extraction module to extract user interests hidden in sessions. Specifically, a bilinear mapping matrix models multiple user interests by mapping items in the session to different interest capsules. A gating network allows the session to select different user interests through the gating mechanism. Then, the interest aggregation module uses an interest anchor to aggregate effective user interest information at each layer, so that the over-smoothing problem is alleviated. Experimental results show that our approach is competitive compared to the state-of-the-art models.

Convolutional Neural Networks (CNNs) have achieved remarkable performance in remote sensing image classification tasks. To address the issue of high model complexity, we propose a block-level pruning strategy based on the semantic similarity analysis that no fine-tuning is required during the pruning process. By employing this strategy, we effectively reduce the complexity of the model. Furthermore, to restore the overall performance of the pruned model, we propose a teacher-student collaborative distillation strategy that enables knowledge transfer through the collaboration of the original model and the dropped-blocks model to promote the performance of the compact pruned model. Experimental results demonstrate that our pruning and distillation strategies outperform other approaches, thereby achieving favorable performance while reducing model complexity.

Hundreds of bands that make up a hyperspectral image (HSI) contain a large amount of redundant information. It is essential to use band selection method to select the effective and representative bands. Besides, the hyperspectral cube typically contains some pixels that do not fall within the same category as the central pixel, which can affect the classification effect. To address above issues, this paper proposes a multi-attention network with multi-level spatial-spectral feature fusion based on band selection (MAFFBS) for HSI classification. The band selection module (BSM) is utilized to choose some bands with substantial amounts of information. Furthermore, we create a new pixel attention block (PAB) to reduce the impact of interfering pixels on the classification. Then, the multi-level spatial-spectral feature fusion block is used to extract and fuse spatial-spectral features at different levels. Finally, the feature reutilization module consisting of the dense block and the spatial attention block is adopted to alleviate the vanishing gradient problem and to produce more discriminative spatial features. Experimental results on two HSI datasets show that the MAFFBS significantly outperforms some state-of-the-art deep learning-based methods.

Existing hyperspectral image (HSI) classification methods rarely consider open-set classification (OSC). Although some reconstruction-based methods can deal with OSC, they lack adaptive threshold strategies and heavily rely on the labeled samples. Therefore, this article proposes a self-supervised multitask learning (SSMTL) framework for few-shot open-set HSI classification, including three stages: pretraining stage (PTS), fine-tuning stage, and testing stage. The model consists of three modules: data diversification module (DDM), 3-D multiscale attention module (3D-MAM), and adaptive threshold module (ATM), as well as a backbone network: dense feature pyramid network (DFPN). In the PTS, we construct a self-supervised reconstruction task with unlabeled samples for model initialization, where DDM aims to improve the robustness of the model and 3D-MAM applies 3-D multiscale convolution to focus on key information spatially and spectrally. In the fine-tuning stage, we further optimize the model with a few labeled samples based on both reconstruction task and classification task, where ATM implements adaptive threshold strategies based on uncertainties of predicted probability and reconstruction loss, and DFPN is helpful to retain the detailed information. The experimental results on three common HSI datasets show SSMTL performs significantly well and even surpasses many advanced closed-set and open-set HSI classification methods.

Graph Convolutional Neural Networks (GCNs) have performed well in many recommendation scenarios. In spite of this, recommendation models based on GCNs still face problems such as insufficient information mining and high complexity for some existing models. To address the above problems, we propose a Graph Convolutional Neural Network for Recommendation Based on Community Detection and the Combination of Multiple Heterogeneous Graphs (GCN-CMHG). This model uses the community detection algorithm to detect the communities in the user-item interaction heterogeneous graph (UIIHG), Finds the regional central nodes of communities, and then creates edges between the regional central node of each community and all other nodes in the UIIHG to construct the heterogeneous partial adjacent graph. Then, a Heterogeneous Partial Adjacent Auxiliary (HPAA) layer is designed to aggregate information on the heterogeneous partial adjacent graph. HPAA layer expands the influence of distant nodes on target nodes, enables target nodes to receive global information, and enhances the ability of GCN-CMHG to mine information. Specially, due to the low complexity of HPAA layer and the abandonment of redundant information, GCN-CMHG is easier to implement and train. Under the exact same experimental setting, GCN-CMHG’s time consumption is only about 1/10 of another model based on GCN called Graph Convolutional Neural Network for Recommendation Based on the Combination of Multiple Heterogeneous Graphs (GCN-MHG). Experiments on multiple real-world datasets show that GCN-CMHG achieves better results compared with several advanced models. The implementation of our work can be found at https://github.com/GCNRSs/GCN-CMHG.

High-Efficient landslide recognition on remote sensing images is of great importance to hazard monitoring. In this paper, we introduce MobileL-K, a light Convolutional neural networks(CNNs) to achieve highly efficient landslide recognition. In this network, the depthwise separable convolutions with a large kernel is borrowed to exploit global features on an image. Moreover, an improved EC-based network pruning method was proposed based on continual masking. We prune the MobileL-K to apply to landslide recognition. The experiment results on a benchmark dataset show that proposed method outperforms other compared methods with smaller model size, less FLOPs and higher running speed on GPU.

Deep convolutional neural network (DCNN) has been widely used in hyperspectral image classification. However, due to the large number of parameters of DCNN, it is difficult for the network to converge during the training. Besides, under the condition of limited samples, DCNN will suffer from the over-fitting problem. In this paper, we propose a multi-branch network based on weight sharing and attention mechanism (MNWA), in which multiple branches share the same parameters and the number of parameters in the proposed neural network is greatly reduced. On the other hand, the proposed attention mechanism can give different weight values to different bands in hyperspectral images, which reduces the transmission of noise bands in the network and increases the transmission of bands useful for classification. Experiments on two datasets show that MNWA is superior to the state-of-the-art methods in the case of limited training samples.

Noise in hyperspectral images (HSIs) may degrade the HSI classification result. Robust principal component analysis (RPCA) is an excellent method to obtain low-rank (LR) representation of data and is widely used in image denoising and also in HSI classification. However, data are drawn as a union from multiple subspaces in HSIs, so LR subspace estimation (LRSE) is necessary when using RPCA, which is complicated and time-consuming. To solve this problem, this letter proposes a novel LR-based method for HSI classification called two-branch network combined with RPCA, which combines RPCA with a neural network. Specifically, both the LR component and the sparse component are preserved and used for feature extraction in two independent convolutional branches. This way, we can avoid information loss without using accurate LRSE. A concatenate operation and a pointwise convolution are then adopted to realize the feature fusion. Finally, fused features are constructed to indicate the ground truth of each pixel in the classification process. The proposed method constructs a convenient model for HSI classification by discarding the LRSE and combining denoising, feature extraction, feature fusion, and classification into an end-to-end network. The experimental results on three data sets demonstrate that the proposed method outperforms many state-of-the-art methods including ones based on LR representation and ones based on deep learning. In addition, it maintains good classification performance for the cases of small samples and class imbalance.

Recommender system (RS) plays an important role in helping users find the information they are interested in and providing accurate personality recommendation. It has been found that among all the users, there are some user groups called “core users” or “information core” whose historical behavior data are more reliable, objective and positive for making recommendations. Finding the information core is of great interests to greatly increase the speed of online recommendation. There is no general method to identify core users in the existing literatures. In this paper, a general method of finding information core is proposed by modelling this problem as a combinatorial optimization problem. A novel Evolutionary Algorithm with Elite Population (EA-EP) is presented to search for the information core, where an elite population with a new crossover mechanism named as ordered crossover is used to accelerate the evolution. Experiments are conducted on Movielens (100k) to validate the effectiveness of our proposed algorithm. Results show that EA-EP is able to effectively identify core users and leads to better recommendation accuracy compared to several existing greedy methods and the conventional collaborative filter (CF). In addition, EA-EP is shown to significantly reduce the time of online recommendation.

This paper proposes a change detection algorithm in synthetic aperture radar (SAR) images based on the salient image guidance and an accelerated genetic algorithm (S-aGA). The difference image is first generated by logarithm ratio operator based on the bi-temporal SAR images acquired in the same region. Then a saliency detection model is applied in the difference image to extract the salient regions containing the changed class pixels. The salient regions are further divided by fuzzy c-means (FCM) clustering algorithm into three categories: changed class (set of pixels with high gray values), unchanged class (set of pixels with low gray values) and undetermined class (set of pixels with middle gray value, which are difficult to classify). Finally, the proposed accelerated GA is applied to explore the reduced search space formed by the undetermined-class pixels according to an objective function considering neighborhood information. In S-aGA, an efficient mutation operator is designed by using the neighborhood information of undetermined-class pixels as the heuristic information to determine the mutation probability of each undetermined-class pixel adaptively, which accelerates the convergence of the GA significantly. The experimental results on two data sets demonstrate the efficiency of the proposed S-aGA. On the whole, S-aGA outperforms five other existing methods including the simple GA in terms of detection accuracy. In addition, S-aGA could obtain satisfying solution within limited generations, converging much faster than the simple GA.

Community detection has received a great deal of attention in recent years. Modularity is the most used and best known quality function for measuring the quality of a partition of a network. Based on the optimization of modularity, we proposed a memetic algorithm with a local search operator to detect community structure. The local search operator uses a quality function of local community tightness based on structural similarity. In addition, the tactics of vertex mover is used for reassigning vertices to neighboring communities to improve the partition result. Experiments on real-world networks and computer-generated networks show the effectiveness of our algorithm.

Many systems in social world can be represented by complex networks. It is of great significance to detect the community structure and analyze the functions for networks. In recent years, plenty of research and works have been focused on this problem. In this paper, we propose an enhanced algorithm based on ant colony optimization (ACO) for the community detection problems. In order to avoid redundant computing in ACO, we divide the ant colony into two groups, original group and intelligent group, which search the solution space simultaneously. In the intelligent group, due to the locus-based adjacency representation of the solution, we let some of them have an ability of self-learning and others can learn from the optimal solutions proactively. Experiments on synthetic and real-life networks show the proposed algorithm can explore in an efficient and stable way.