Cross-modal hashing has gained widespread attention in cross-modal retrieval due to its low storage cost and significant computational efficiency. Existing cross-modal hashing methods primarily focus on learning modality invariance by mapping data from different modalities into a shared space and learn unified hash codes. Nevertheless, due to the inherent heterogeneity between different modalities, the common subspace may still exhibit modality discrepancies. This ultimately makes it challenging to achieve semantic alignment, thereby affecting the accuracy of cross-modal retrieval. To address this issue, we propose an Unsupervised Adversarial Contrastive Hashing (UACH) method for cross-modal retrieval. Specifically, we design a cycle generative adversarial network to learn the transformation relationships between different modality feature domains, effectively promoting semantic alignment across modalities. Additionally, we employ dual contrastive learning to simultaneously measure the representation learning and hashing learning components of each specific modality, and learn unified hash codes for each specific modality, thus mitigating the impact of modality discrepancies. Extensive experiments conducted on three cross-modal benchmark datasets demonstrate that our model outperforms the state-of-the-art baselines.

In recent years, research has illuminated the potency of implicit data processing in enhancing user preferences. Nevertheless, barriers remain in breaking through the constraints of implicit information. This study aims to bridge this gap by firstly constructing a triangular stable node network model, tailored to manage implicit information with precision. Recognizing the challenge of pinpointing novel structures within large-scale graphs, we propose SLTSNN— a triangular stable node network based on self-supervised learning for personalized prediction. SLTSNN innovates by maximizing mutual information between graph-level and patch-level representations, while augmenting graph representations through extended enhanced representations. Additionally, it incorporates triangular data associations and introduces a triangular allocation attention network, which emphasizes strongly correlated preference features among similar users. Furthermore, SLTSNN employs contrastive learning to maximize mutual information between graph vectors and hidden representations, distinguishing between high-order global and local representations. The model’s effectiveness in enhancing user preferences and capturing new graph structures is evidenced by its performance on hit rate and normalized discounted cumulative gain metrics across three datasets.

Multimodal Named Entity Recognition (MNER) aims to achieve more accurate entity recognition by incorporating image information to assist text, which is particularly significant on social media platforms. Current research disproportionately emphasizes enhancing text with images, overlooking that the core of the NER task remains textual. The modal differences between images and text inevitably introduces noise when incorporating image information. Therefore, when textual information is sufficient to independently complete the NER task, the introduction of image information is unnecessary. This paper proposes an Adaptive Logical Decision Framework (ALDF) capable of determining the sufficiency of textual information in NER tasks, deciding whether to introduce image information, avoiding unnecessary noise, and focusing more on information-scarce entities when introducing image information. Specifically, we designed a Logic Reasoning Neural Network (LRNN) that uses an evidence-theory-based method to simulate human decision-making logic and generate decision support degrees for deciding whether image information should participate in the recognition task. When incorporating image information, we utilize the generated decision support degrees to guide the multi-head self-attention mechanism, enhancing the model's focus on information-scarce entities. Additionally, we employ a modality-aware progressive training method that can use decision information in real-time during multimodal training and reduce information redundancy between modalities. Extensive experiments demonstrate that our model achieves state-of-the-art performance on popular public datasets.

The development of the Internet has made people more closely related and has put forward higher requirements for recommendation models. Most recommendation models are studied only for the long-term interests of users. In this paper, the interaction time between the user and the item is introduced as auxiliary information in the model construction. Interaction time is used to determine users’ long-term preferences and short-term preferences. In this paper, temporal features are extracted by building a convolutional gated recurrent unit with attention neural network (CNN-GRU-Attention). Firstly, for the problem of accurate feature extraction, CNN are constructed to extract higher-level and more abstract features of themselves and transform high-dimensional data into low-dimensional data; secondly, for the problem of social temporality, GRU are used to not only extract temporal information, but also effectively reduce gradient dispersion, making model convergence and training easier; finally, Graph Attention networks are used to aggregate the social relationship information of users and items respectively, which constitute the final feature representation of users and items respectively. In particular, a modified cosine similarity is used to reduce the error caused by data insensitivity when constructing the social information of the item. In this study, simulation experiments are conducted on two publicly available datasets (Epinions and Ciao), and the experimental results show that the proposed recommended model performs better than other social recommendation models, improving the evaluation metrics of MAE and RMSE by 1.06%-1.33% and 1.19%-1.37%, respectively. The effectiveness of the model innovation is proved.

With the advancement of the Internet, Graph Neural Networks based recommendation systems have become a topic of great concern in the research field. However, the current recommendation systems still have the following problems. First, it focuses on modeling users but ignores the problems of missing values of item rating vectors and non-corresponding positions of item rating vectors in the process of solving associated items; second, it focuses on the association relationship between users but pays less attention to the association relationship between items; third, there is insufficient research on the short-term attractiveness of items and the users' temporary preferences. To address the above problems, this study proposes the following solutions to better construct the item social graph and extract the short-term interest/attraction of users/items. Firstly, for problem one, this study reconstructs the item rating vector innovatively based on whether users have interaction with the items; secondly, for problem two, this study proposes to use Pearson similarity to calculate the association relationship between items so as to construct the item social graph. Again, for problem three, This paper investigates temporal information features and extracts short-term user preferences and item attractiveness. To achieve this, an attention network that focuses on temporal information features is constructed by combining channel attention and bidirectional long-term and short-term memory networks. Finally, it involves using a multilayer perceptron with a residual connection structure to combine user and item factors, leading to more accurate predictions. In this study, two publicly available datasets, Epinions and Ciao, were used in a comparative experiment. This model outperformed other baseline models in the experiment, resulting in a reduction of 1.42% and 1.24% in MAE values, and 1.47% and 1.38% in RMSE values, respectively. These findings suggest that incorporating short-term preferences and reconstructing item social graphs can enhance the precision of social recommendations.

Recent years, people have been gradually influenced by online socialization. The emergence and development of Graph Neural Networks (GNN) has shown great advantages in mining implicit data and expressing node relationships, etc. However, due to the arbitrary nature of establishing neighborly relationships, the judging reliability of trusting relationships between neighbors is a difficult issue. Therefore, this paper proposes an innovative approach to obtain the valid neighbor relationships based on real user-item interactions and friend trust relationships in the dataset. The proposed method can address the impact of invalid neighbors on the social model and achieve the effect of enhancing neighbor perception. For neighbor interaction in the graph neural network model, this paper establishes the direct connection between users and items through mapping of multilayer perceptron firstly. Then it integrates neighbor similarity and neighbor sampling to mitigate the interference information of invalid neighbors and achieves the effect of enhancing the perceived information of neighbor interaction. Finally, this paper establishes the item social space and user social space with enhanced neighbor perception according to the dyadic nature and organically integrated them to enrich the social data. Comparison experiments are carried out based on two publicly available datasets Epinions and Ciao, the recommended method performs better than other social recommendation models with the MAE and RMSE values being improved by 0.81%-1.09% and 1.15%-1.41% respectively.

It is hard for researchers to retrieve the desired information from the huge number of academic literatures efficiently. This paper proposes an automatic summarization method based on fuzzy ontology and machine learning to solve this problem. This method uses fuzzy ontology to model the literature context, and extracts the literature related context information, including the citation, publication time, the author's research interest, the recently published literature, etc. We integrate the domain element and the time element into the term frequency-inverse document frequency (TF-IDF) model, extract words from the document set that reflect the core content of the document. Then obtain topic-related words through the Latent Dirichlet Allocation (LDA) model. Finally, the important sentences are selected as abstracts according to these words. Experiments show that our method is superior to the standard TF-IDF method and LDA method.