[Objective] This study aims to systematically summarize sequence recommendation methods that incorporate knowledge features through a comprehensive literature review. [Coverage] Using “Sequential Recommendation * Knowledge” and “序列推荐*知识” as advanced search terms, we searched databases including Web of Science, DBLP, Google Scholar, and CNKI. A total of 97 articles were selected with special attention to the core content of specific chapters to ensure their alignment with research needs. [Methods] Employing the literature review approach, we categorized and analyzed sequence recommendation methods from three perspectives: research framework, real-world applications and evaluations, and future research directions. [Results] We constructed a research framework for the application of knowledge features in the sequential recommendation, which includes three components: knowledge feature representation, temporal knowledge enhancement, and sequence recommendation algorithms integrating knowledge features. We also analyzed the limitations of existing evaluation resources from datasets, evaluation metrics, and baseline models and explored future research directions. [Limitations] While the study provides a comprehensive overview of relevant works in the rapidly evolving field of knowledge-enhanced sequential recommendation, it may not cover all existing studies due to the breadth and volume of the literature. [Conclusions] Sequence recommendation algorithms that consider knowledge features enhance the accuracy of recommendations. Integrating multimodal knowledge features contributes to a deeper understanding of user needs.
[Objective] This paper comprehensively reviews the methods of text augmentation to reveal their current state of development and trends. [Coverage] Using “textual data augmentation” and “text augmentation” as search terms to retrieve literature from Web of Science, Google Scholar and CNKI, we screened out a total of 88 representative papers for review. [Methods] Text augmentation methods were categorized and summarized according to the objects of operation, the details of implementation and the diversity of generated results. On this basis, we conducted a thorough comparison of various methods with regards to their granularity, strengths, weaknesses and applications. [Results] Text augmentation approaches can be divided into text space-based methods and vector space-based methods. The former is intuitive and easily interpretable but may compromise the overall semantic structure of the text, while the latter can directly manipulate semantic features but incurs higher computational complexity. Current studies frequently necessitate external knowledge resources, such as heuristic guidelines and task-specific data. Moreover, the introduction of deep learning algorithms can enhance the novelty and diversity of generated data. [Limitations] We primarily offer a systematic examination of technical principles and performance characteristics of advanced methods, without assessing the developmental stage of platform tools quantitatively. Besides, the analysis is grounded in our chosen literatures and may not encompass all potential application scenarios of text augmentation methods. [Conclusions] Future work should pay more attention to enriching and refining the evaluation metrics for text augmentation techniques and increasing their robustness across different downstream tasks by prompt learning. Retrieval-augmented generation and graph neural networks should be taken seriously for addressing the challenges posed by lengthy texts and limited resources, which can further unlock the potential of text augmentation methods in the field of natural language processing.
[Objective] This paper aims to improve the accuracy of social media rumor detection and reduce their potential threat to social stability. [Methods] We proposed a rumor detection model integrating dynamic propagation and the Neural Hawkes Process. The model divided the propagation process into subgraphs according to the tweet timeline and constructed subgraph embeddings. Then, we input the embedding sequences into a global dynamic evolution encoding module. Third, we created the weighted sequence incorporating temporal encoding and fed it into the Neural Hawkes Process module to calculate the continuous conditional intensity function, modeling the self-exciting propagation phenomenon. The output after average pooling is passed through a feedforward neural network for rumor detection. In addition, we used a multi-task learning module to calculate the overall loss of two types of outputs to guide model training. [Results] The model achieves accuracy rates of 85.6% and 86.6% on the publicly available datasets Twitter15 and Twitter16, respectively, outperforming other mainstream baseline models for early rumor detection. [Limitations] The model currently only uses textual data and temporal attributes. Incorporating features such as tweet images and user attributes might enhance the model’s accuracy. [Conclusions] Encoding dynamic information and self-exciting characteristics of tweet propagation can improve rumor detection effectiveness.
[Objective] This paper addresses the issues of limited information in rumor data and the lack of associated commonsense information. It improves the accuracy of rumor identification. [Methods] We proposed a Multi-Branch Graph Convolutional Inference Network (MGCIN), which combines a bidirectional graph convolutional network with a commonsense inference module. These two components independently generated classification labels and achieved joint decision-making. [Results] We examined the model on three public datasets: Twitter15, Twitter16, and PHEME. The proposed method outperformed most baseline models, achieving accuracy rates of 87.8%, 89.8%, and 77.6%, respectively. It also demonstrated excellent early rumor detection performance. [Limitations] Further research is needed on the multimodality of background and commonsense information related to rumor data. [Conclusions] The proposed model effectively simulates human cognitive processes, successfully integrates textual features, propagation features, and commonsense knowledge, and provides new ideas and methods for rumor detection research.
[Objective] Tracking and observing the characteristics of public opinion circulation during emergencies can facilitate effective public opinion guidance, control, and shared governance. [Methods] Using the case study method, we construct a framework for understanding the macroscopic circulation of public opinion in emergencies. Using social network analysis, complemented by empirical research and natural language processing technology, we conduct an in-depth analysis of the circulation patterns of public opinion from a micro perspective, focusing on the dimensions of subjects, objects, and carriers. Validation analyses are conducted using data from public health emergencies. [Results] From a macro perspective, public opinion circulates across Cyber Space, Physical Space and Psychological Space, providing an interdisciplinary analytical framework for understanding and quantifying public behaviors and responses. At the micro level, public opinion circulates among multiple groups, media, events and platforms, exhibiting four effects respectively: homogeneous diffusion and heterogeneous traversal effect, field resonance and field escape effect, co-temporal and ephemeral effect, and amplified resonance and echo difference effect. [Limitations] The dynamics of social network sentiment are not considered. [Conclusions] By summarizing the laws of cross-domain circulation of public opinion from both macroscopic and microscopic perspectives and conducting empirical research linked to specific events, we provide new insights into the study of public opinion communication.
[Objective] To trace learners’ progress and assess their knowledge mastery in order to provide tailored academic support. [Methods] This research proposes a model named Fine-grained Learning Ability Boosted Interpretable Knowledge Tracing, that integrates knowledge mastery with fine-grained learning abilities. It extends Item Response Theory by incorporating error rate, allowing prediction of learners’ next-step attempting performance while providing interpretable insights. [Results] Experiments on three public datasets show that the proposed knowledge tracing model achieves a minimum improvement of approximately 2% on the AUC metric compared to the majority of baseline methods. [Limitations] The proposed model improves the interpretability of knowledge tracing through cognitive factor augmentation in learning analytics. Nevertheless, rigorous empirical verification remains imperative to advance the explanatory capacities in deep learning-based knowledge tracing. [Conclusions] The proposed model demonstrates enhancements in predictive accuracy while dynamically reconstructing learner cognitive profiles through rigorous multi-faceted representations, which improves the interpretability of the knowledge tracing.
[Objective] This study proposes a deep interaction and self-attention fusion recommendation model based on ID features, termed DFM-ID, to effectively exploit the deep semantic information embedded in ID features. [Methods] A deep learning framework specifically designed for ID features is proposed, incorporating three types of feature interaction layers and a self-attention-based fusion module. The model performs both low-order and high-order interactions among ID features, and utilizes the self-attention mechanism to generate deep semantic representations of ID features. [Results] Experiments conducted on three public datasets show that models incorporating DFM-ID achieve relative improvements of 16.03% in accuracy, 14.10% in precision, 20.97% in AUC, and 8.68% in F1 score compared to baseline methods. [Limitations] The experimental datasets exhibit a degree of homogeneity, which may limit the generalizability of the model to more diverse recommendation scenarios. [Conclusions] The proposed model effectively captures complex associations and in-depth information among ID features, significantly improving recommendation accuracy.
[Objective] To utilize influencing factors to mine disease-disease relationships in the biomedical literature and provide new perspectives for disease association analysis. [Methods] Based on the important role of influencing factor interventions in multimorbidity management, the extraction of disease-influencing factor entity relationships was completed by dependency analysis, combined with complex network analysis techniques for disease community discovery, and a disease association model based on influencing factors was constructed and validated using data from the Chinese Medical Association Journal Database. [Results] This model generated a network of 105 diseases, 453 influencing factors, and 2,067 edges, and discovered nine disease communities with strong internal associations mediated by influencing factors to realize the disease association analysis. [Limitations] The efficacy of acquiring disease-influence factors was found to be less effective for complex long sentences, which consequently reduced the number of multimorbidity associations established based on influence factors. [Conclusions] The disease association model based on influencing factors can obtain finer-grained disease-influencing factor relationships with better representativeness and interpretability. Furthermore, it can provide new research ideas for disease association analysis and multimorbidity co-management.
[Objective] To address the issues of accurately identifying scholars’ research interests, this study proposes a method integrating paper content and citation features. We constructed a paper recommendation model using an academic knowledge graph and random walk algorithms. [Methods] Firstly, we used pre-trained text embedding models and citation networks to learn vector representations of published papers and extracted research interests with similarity theory. Then, we utilized knowledge graph embedding, biased random walk, and attention mechanisms to calculate the probabilities of scholars’ interest in a paper. Finally, the model generated a paper recommendation list. [Results] We conducted experiments on the DBLPv14 dataset, and our model outperformed baseline models. The F1-score and MRR were improved by 0.041 and 0.031, respectively, with overall better performance across all evaluation metrics. [Limitations] Our model does not consider the impact of entity and relationship attributes on recommendation performance. [Conclusions] By incorporating paper content and citation features, the proposed model effectively identifies scholars’ research interests and improves the paper recommendation accuracy.
[Objective] This study addresses the issue of the traditional LDA model in processing short texts, particularly the abstracts of Traditional Chinese Medicine (TCM) papers where domain-specific terminology is abundant and topic term interpretability is limited. [Methods] We proposed an Improved LDA model (I-LDA) incorporating rough data reasoning. The method used a TextRank algorithm enhanced with rough data reasoning to extract the most representative keywords. Then, we constructed a domain-specific dictionary to increase the vocabulary’s weight. Finally, we expanded the range of topic term selection by integrating rough data reasoning. [Results] Compared to the traditional LDA model, the I-LDA model improved topic coherence and inter-topic distance by approximately 5.6 and 1.8 percentage points. [Limitations] Due to the large number of specialized terms in TCM abstracts, the pre-defined dictionary used in the experiment may not fully cover all relevant terminology, which could affect the model’s performance in topic modeling. [Conclusions] The I-LDA model demonstrates superior performance in topic modeling of TCM paper abstracts, with more representativeness and domain specificity identification.
[Objective] This study proposes a medical publication recommendation model that uses cross-modal information to improve recommendation accuracy. [Methods] First, the medical terminology system was employed to standardize label content and align image-text tags. Paired semantic labels were then utilized to align feature semantics between images and texts through contrastive learning. Based on the aligned semantic features, a cross-modal cross-attention mechanism was constructed, and user preferences for publications were predicted by analyzing their interest weights across different modalities. [Results] Comparative experiments with three state-of-the-art multimodal baseline models on two publication datasets showed that the proposed model achieved an average precision of 62.79%, F1-score of 53.62%, and NDCG of 61.17%, outperforming the baseline models in all metrics. [Limitations] Additional cold-start methods may be required for pre-training data containing only single-modality information. [Conclusions] The proposed model exhibits strong cross-modal feature fusion capabilities, effectively mitigating semantic gaps between modalities and improving the accuracy of medical publication recommendations.
[Objective] To enhance the efficiency of image retrieval for Dunhuang murals from a multi-label perspective, this paper designed a multi-label image retrieval model for Dunhuang murals. [Methods] First, Dunhuang murals images were collected, and multi-label labeling was performed based on their themes and contents. Then, the image features were extracted using DenseNet, and hashing was used to compress their encoding. Next, the label information was integrated to perform image matching using cosine similarity, with results ranked by similarity scores. Finally, the HyP2 loss function was adopted to evaluate and optimize hash code generation in the model. [Results] The DenseNet hashing-based multi-label image retrieval model (DNHMIR) shows strong performance on the constructed Dunhuang murals multi-label dataset, with mAP@7000 of 0.884, which reflects an improvement of at least 0.044 over the baseline models. [Limitations] Mapping image features to hash codes may result in partial loss of image information, and we ignore cognitive differences among user groups. [Conclusions] The DNHMIR model constructed for Dunhuang murals can accurately retrieve multi-labeled images, reduce storage space and search time, and improve the retrieval efficiency of Dunhuang murals.
[Objective] This paper addresses the issues of time-consuming and cumbersome manual sleep stage classification methods, as well as the long training time and poor recognition performance of existing automated sleep stage classification models. It also enhances the accuracy and robustness of sleep stage prediction. [Methods] We proposed a sleep stage classification model (WaveSleep) based on Discrete Wavelet Transform (DWT) and Residual Shrinkage Networks. First, we used the DWT to decompose the original electroencephalogram signals. Then, we extracted multi-resolution features using convolutional neural networks with varying kernel sizes. Third, we used a deep residual shrinkage network to model channel-layer feature correlation. Finally, we utilized a temporal context encoder with multi-head attention to capture the temporal dependencies of the extracted features. [Results] The proposed model achieved accuracy rates of 85.4%, 81.9%, and 84.4% on three public sleep datasets, respectively, representing improvements of 1.0, 0.6, and 0.2 percentage points compared to the baseline model. [Limitations] The improvement in the accuracy of the proposed model on different unbalanced datasets is limited. [Conclusions] The proposed model can effectively improve the sleep stage prediction’s efficiency and accuracy and exhibit significant robustness.