Implicit Correlation Research Articles

Temporal Knowledge Graph Reasoning Based on Dynamic Fusion Representation Learning

ABSTRACTRecently, significant progress has been made in completing static knowledge graphs. However, knowledge tends to evolve with time, and static knowledge graph completion (KGC) methods struggle to capture the changes. Therefore, temporal knowledge graph (TKG) reasoning has become a focus of research. Most existing TKG methods incorporate temporal information into triplets and transform them into KGC tasks, ignoring the important influence of time information and implicit relationships between entities. In this paper, we propose a new method called TD‐RKG, which addresses the challenges of temporal variability and implicit entity correlations based on a dynamic fusion representation learning approach. The method consists of four modules: dynamic local recurrent encoding layer, dynamic implicit encoding layer, dynamic global information attention layer and decoding layer. Experimental results on three benchmark datasets demonstrate substantial improvements in TD‐RKG across multiple evaluation metrics.

Emotion Recognition Using EEG Signals and Audiovisual Features with Contrastive Learning.

Multimodal emotion recognition has emerged as a promising approach to capture the complex nature of human emotions by integrating information from various sources such as physiological signals, visual behavioral cues, and audio-visual content. However, current methods often struggle with effectively processing redundant or conflicting information across modalities and may overlook implicit inter-modal correlations. To address these challenges, this paper presents a novel multimodal emotion recognition framework which integrates audio-visual features with viewers' EEG data to enhance emotion classification accuracy. The proposed approach employs modality-specific encoders to extract spatiotemporal features, which are then aligned through contrastive learning to capture inter-modal relationships. Additionally, cross-modal attention mechanisms are incorporated for effective feature fusion across modalities. The framework, comprising pre-training, fine-tuning, and testing phases, is evaluated on multiple datasets of emotional responses. The experimental results demonstrate that the proposed multimodal approach, which combines audio-visual features with EEG data, is highly effective in recognizing emotions, highlighting its potential for advancing emotion recognition systems.

Collaborative trajectory representation for enhanced next POI recommendation

Point-of-Interest (POI) recommendation stands as the cornerstone within a variety of location-based applications and services that intend to anticipate upcoming movements that users may be interested in. The current state-of-the-art methods have effectively explored spatio-temporal contextual features and users’ long-term and short-term preference patterns. Nevertheless, most existing work lacks the ability to effectively capture group movement patterns from trajectory collaboration. Additionally, they pay close attention to the accuracy of personalized recommendations, neglecting recommendation diversity, which refers to offering broader location options that extend beyond a user’s typical preferences, thereby avoiding excessive homogeneity or repetition. To address these gaps, this study proposes a Collaborative Trajectory Representation model (CTRNext), which enhances the diversity of recommendations while maintaining the precision of personalized preferences. To be specific, we first design two trajectory embedding layers to extract joint semantic interactions and the explicit spatiotemporal context-aware representation. Then, a trajectory semantic similarity calculation module that captures collaborative signals from potentially similar-minded users and eliminates barriers caused by trajectory length is proposed. Next, the implicit correlation and further updated representation between different check-in records are achieved through a multi-head self-attention aggregation module. Finally, we put forward a dual-driven user preference matching module to generate the preference-based next POI recommendation while enhancing diversity. Our approach demonstrates its remarkable recommendation accuracy through extensive experimentation on four real-world datasets, surpassing the performance of state-of-the-art methodologies.

Linguistic-Driven Partial Semantic Relevance Learning for Skeleton-Based Action Recognition.

Skeleton-based action recognition, renowned for its computational efficiency and indifference to lighting variations, has become a focal point in the realm of motion analysis. However, most current methods typically only extract global skeleton features, overlooking the potential semantic relationships among various partial limb motions. For instance, the subtle differences between actions such as "brush teeth" and "brush hair" are mainly distinguished by specific elements. Although combining limb movements provides a more holistic representation of an action, relying solely on skeleton points proves inadequate for capturing these nuances. Therefore, integrating detailed linguistic descriptions into the learning process of skeleton features is essential. This motivates us to explore integrating fine-grained language descriptions into the learning process of skeleton features to capture more discriminative skeleton behavior representations. To this end, we introduce a new Linguistic-Driven Partial Semantic Relevance Learning framework (LPSR) in this work. While using state-of-the-art large language models to generate linguistic descriptions of local limb motions and further constrain the learning of local motions, we also aggregate global skeleton point representations and textual representations (which generated from an LLM) to obtain a more generalized cross-modal behavioral representation. On this basis, we propose a cyclic attentional interaction module to model the implicit correlations between partial limb motions. Numerous ablation experiments demonstrate the effectiveness of the method proposed in this paper, and our method also obtains state-of-the-art results.

Cascaded learning with transformer for simultaneous eye landmark, eye state and gaze estimation

Eye tracking have garnered attention in human–machine interaction, disease monitoring, biometrics, etc. Existing investigations for eye tracking have predominantly concentrated on individual task for pupil detection or gaze estimation, overlooking the implicit relationships that exist among different tasks for eye tracking. In this work, we introduce a cascaded framework with transformer to collaboratively realize eye landmark detection, eye state detection and gaze estimation. Within our framework, we leverage Transformer to capture long dependencies with explicit eye-related structural information and implicit correlation among different tasks. Furthermore, the proposed cascade iteration framework alternatively optimize each task and boost the overall performance for pupil center, eye state and gaze estimation simultaneously. To address the problem of manual annotation, we further introduce the Control-Eye Diffusion Model (CEDM), a controllable eye image generation method conditioned on a simple contour with structure information. The proposed methods are evaluated on challenging datasets such as GI4E, BioID and MPIIGaze, and the results show that our methods outperform state-of-the-art methods in several tasks.

A deep learning-based surrogate model for probabilistic analysis of high-speed railway tunnel crown settlement in spatially variable soil considering construction process

In reality, many high-speed railway tunnels are built by sequential excavation method in spatially variable soil. This study introduces an efficient deep learning-based solution for the probabilistic analysis of tunnel crown settlements considering the spatial variability of surrounding soil. The two-dimensional convolutional neural network (2D-CNN) is employed to uncover the implicit correlation between the soil elastic modulus random field and tunnel crown settlements. The correlation coefficient (R) of the surrogate model is greater than 0.9664. Meanwhile, the mean relative error of the predictions remains within 2.38%, under various scales of fluctuation (SOFs). An additional 10,000 samples were produced to assess the practical performance of the trained model in probabilistic analysis. The relative errors of predictions generally fall within 5%, with a minimum confidence interval of 98.15%. These results demonstrate that the developed 2D-CNN model can effectively substitute the traditional method in predicting tunnel crown settlements considering soil spatial variability.

HyperArc performance in radiotherapy and its correlations with number and volume of multiple brain metastases

ObjectiveTo assess the performance and advantages of HyperArc for radiotherapy of multiple brain metastases with dosimetric parameters with respect to coplanar volumetric modulated arc radiotherapy (c-VMAT), and characterize implicit correlations between tumor features and dosimetric parameters. MethodsThis retrospective study involved 40 patients with multiple brain metastases (4–16 tumors, 12.8–240.8 cm3) who received simultaneous radiotherapy at the Department of Radiation Oncology, Shenzhen People's Hospital from January 2019 to December 2022. Both HyperArc and c-VMAT plans were designed by the same and qualified physicist using the Eclipse system with Truebeam. A single isocenter and a 6 ​MV flattening filter-free (6FFF) photon beam were used for radiotherapy of multiple brain metastases in each plan. Dosimetric parameters of tumors and organs at risk (OARs) were compared between HyperArc and c-VMAT plans. Correlations between tumor features (number and volume) and dosimetric parameters were revealed and analyzed. Furthermore, monitor units (MUs) of both plan types were recorded to assess rays utilization and delivery efficiency. ResultsCompared to c-VMAT plans, HyperArc plans achieved higher maximum dose (Dmax) (5.23%, t ​= ​7.87, P ​< ​0.01), but lower minimum dose (Dmin) (3.94%, t ​= ​−2.12, P ​= ​0.04) and mean dose (Dmean) (1.05%, t ​= ​−3.29, P < ​0.01) for planning gross tumor volumes (PGTVs). The conformity index (CI), homogeneity index (HI), gradient index (GI) and R50% of PGTVs with HyperArc were better than those with c-VMAT, with the improvements of 20.78%, 1.68%, 19.83% and 36.20%, respectively (P ​< ​0.01). The dosimetric parameters of OARs were significantly improved in HyperArc plans compared to c-VMAT plans. Specifically, the Dmax and the Dmean to the brainstem decreased by 16.14% and 11.37%, respectively. The Dmax for eye_L decreased by 16.88%, for eye_R decreased by 11.67%, for optical nerve_L decreased by 12.56 ​%, and for chiasm decreased by 18.12%. Few correlations were observed between the differences of CI/HI/GI/R50% and the number of PGTVs between the two plan types. Whereas, regression analysis indicated that the differences of CI/ GI/ R50% were negatively correlated with the volume of PGTVs, respectively. Moreover, although HyperArc plans reduced the MUs by 4.74 ​% compared to c-VMAT, no significant difference was observed (P ​> ​0.05). ConclusionsThis study demonstrates the promising performance of HyperArc for radiotherapy of multiple brain metastases, including better dose conformity for the tumors, more rapid dose drop and less dose spillage outside the tumors area, effective reduction of dose bridges among close tumors, and sparing OARs. It also reveals the correlations between the number or volume of tumors and dose parameters, which can help optimize the use of HyperArc and predict clinical benefits.

Control of Interface Migration in Nonequilibrium Crystallization of Li2SiO3 from Li2O-SiO2 Melt by Spatiotemporal Temperature and Concentration Fields.

During liquid-solid transformation, bulk mass and thermal diffusion, along with the evolved interfacial latent heat, work in tandem to generate interfacial thermodynamic and kinetic forces, the interplay of which decides the solidification velocity and consequently the solidified phase attributes. Hence, access to interface dynamics information in dependence of bulk transfer processes is pivotal to tailor the desired quantity of solid phases of unique compositions. It finds particular application for engineering concentrated Lithium (Li) phases out of Li-ion battery slags, thus generating a high value-added product from a conventional waste process stream. However, considerable challenge exists to predict the impact of the diverse external cooling rates on the evolving internal transfer processes and thus tuning solidification routes for achieving phases of interest. Hence, in this work, a thermodynamically consistent nonequilibrium model, by considering spatiotemporal temperature and concentration fields, is developed and applied to study solidification of Li2SiO3 from a Li2O-SiO2 melt that constitutes an important subsystem of the Li containing battery-recycling slags. The approach treats the sharp solid/liquid interface as a moving heat source. In the presence of different heat extraction profiles, it evaluates the spatial temperature heterogeneity and its implicit correlation to internal material fluxes resulting from maximization of dissipation and consequently the interrelation to interface velocities. Model calculations revealed that irrespective of the external cooling rate, for an initial short time duration, the magnitude of which increased with decreasing cooling rates, the interface velocities show a reducing trajectory directly relatable to the reducing thermodynamic forces due to localized interfacial temperature rise from the generated latent heat of fusion from the initial solidification. This is followed by a thermodynamically controlled regime, whereby for each cooling rate, the interface velocities increase until a maxima, the magnitude of which decreases with decreasing cooling rates. Finally, the interface propagation speeds decrease as controlled by the kinetic regime.

Label enhancement via manifold approximation and projection with graph convolutional network

Label enhancement (LE) aims to enrich logical labels into their corresponding label distributions. But existing LE algorithms fail to fully leverage the structural information in the feature space to improve LE learning. To address this key issue, we first apply manifold learning to map the relatedness between low-dimensional feature samples to the label space. Based on the smoothness assumption of manifolds, the implicit correlation between low-dimensional feature and label spaces effectively promotes the LE process, enabling the learning model to accurately capture the mapping relationship between feature and label manifolds. This leads to an LE based on feature representation (LEFR) algorithm. We also propose an LE algorithm based on graph convolutional network (GCN), called LE-GCN. Inspired by the relationship between threshold connections and label connections, we extend GCN to the LE field for the first time to fully exploit the hidden relationships between nodes and labels. By enhancing node information with threshold connections and label connections, the label learning accuracy reaches a new level. Experiments on real-world datasets show that our LEFR and LE-GCN outperform several state-of-the-art LE algorithms.

Exploration of the methods and rules of syndrome/pattern differentiation and treatment of headache from the acupuncture-moxibustion prescriptions of ancient literature based on the data mining technology.

The study aims to identifying and exploring the methods and rules of the syndrome/pattern differentiation and treatment of headache through collating acupuncture-moxibustion prescriptions recorded earliest in ancient literature. Using Excel2016 software, the structural data table was prepared with "name of disease", "location of disease", "etiology and pathogenesis", "complicated symptoms", "sites for acupuncture and moxibustion" and "techniques of acupuncture and moxibustion" included. The normative approach was conduced on "name of disease", "etiology and pathogenesis", "complicated symptoms" and "nomenclature of acupoint". Using conventional literature statistical method, combined with Apriori algorithm of association rule, the implicit multi-dimensional correlation rules were explored among various elements of syndrome/pattern differentiation of headache and corresponding therapeutic methods. Based on the findings of the study, the regularity was distinct regarding the treatment at "distal acupoints along the affected meridian and the local acupoints at the affected area" after identifying the location of headache; the strong association was presented between "etiology and pathogenesis" and "acupoint selection", and between "etiology and pathogenesis" and "therapeutic methods", including 9 and 12 rules, respectively. Guanyuan (CV 4) selected in treatment of headache was associated with kidney deficiency, the combination of Zhongwan (CV 12) and Zusanli (ST 36) was with phlegm, Fengfu (GV 16), Fengchi (GB 20), Xinghui (GV 22) and Baihui (GV 20) was with wind, and Hegu (LI 4) was with cold. Moxibustion was dominant in treatment if headache was caused by pathogenic cold or related to deficiency syndrome; acupuncture was used specially for the case caused by phlegm, or interaction of wind and phlegm or wind and heat. For heat syndrome, either acupuncture or moxibustion was applicable, in general, acupuncture was more commonly used in comparison with moxibustion for headache. There were 6 association rules regarding the acupoint selection and the techniques of acupuncture and moxibustion. Moxibustion was generally applied to Xinghui (GV 22), Shangxing (GV 23) and Baihui (GV 20) ; and acupuncture was to Fengfu (GV 16), Hegu (LI 4) and Zusanli (ST 36). There were few association rules between the complicated symptoms and acupoint selection. Among nearly 100 complications, there were only 3 feature associations. Zhongwan (CV 12) was selected for the case with poor appetite, Chengjiang (CV 24) was with neck stiffness, and Fengchic (GB 20) combined with Fenglong (ST 40) or Jiexi (ST 41) was used if vertigo was present. In the ancient time, regarding the treatment of headache, acupuncture and moxibustion are delivered based on the three aspects, i.e. the location of illness, the etiology and pathogenesis, and the complicated symptoms. For acupoint selection, in line with the courses of affected meridians, the adjacent and distal acupoints are combined according to the location of headache. The acupoint prescription is composed in terms of the etiology and pathogenesis. The techniques of acupuncture and moxibustion are optimized in consideration of the sites where acupuncture and moxibustion are operated.

A novel complex-high-order graph convolutional network paradigm: ChyGCN

Abstract In recent years, there has been a growing interest in Graph Convolutional Networks (GCN). However, existing GCN and variants are predominantly based on simple graph or hypergraph structures, which restricts their ability to handle complex data correlations in practical applications. These limitations stem from the difficulty in establishing multiple hierarchies and acquiring adaptive weights for each of them. To address this issue, this paper introduces the latest concept of complex hypergraphs and constructs a versatile high-order multi-level data correlation model. This model is realized by establishing a three-tier structure of complexes-hypergraphs-vertices. Specifically, we start by establishing hyperedge clusters on the foundational network, utilizing a second-order hypergraph structure to depict potential correlations. For this second-order structure, truncation methods are applied to assess and generate a three-layer composite structure. During the construction of the composite structure, an adaptive learning strategy is implemented to merge correlations across different levels. We evaluate this model on several popular datasets and compare it with recent state-of-the-art methods. The comprehensive assessment results demonstrate that the proposed model surpasses existing methods, particularly in modeling implicit data correlations (the classification accuracy of nodes on five public datasets Cora, Citeseer, Pubmed, Github Web ML, and Facebook is 86.1±0.33, 79.2±0.35, 83.1±0.46, 83.8±0.23, and 80.1±0.37, respectively). This indicates that our approach holds advantages when handling datasets with implicit multi-level structures.

DSAMR: Dual-Stream Attention Multi-hop Reasoning for knowledge-based visual question answering

Knowledge-based visual question answering aims to associate external knowledge facts for answering questions about images. Most existing methods emphasize high-order associations between knowledge facts and questions, and fail to consider the negative effects of unnecessary knowledge facts in multi-hop reasoning. In this paper, we propose a Dual-Stream Attention Multi-hop Reasoning (DSAMR) architecture that constructs two different attention streams to mitigate unnecessary knowledge facts. This dual-stream mechanism enables the model to reduce the attention weights on unnecessary knowledge while gathering essential knowledge by learning the implicit correlations between knowledge facts and questions. In addition, we designed a hypergraph knowledge extraction module in the architecture to extract optimal knowledge facts by evaluating the relevance of each knowledge fact to the question. The experimental results demonstrate the effectiveness of our method not only on the knowledge-based visual question answering dataset KVQA, but also on the multi-hop question answering dataset PathQuestion.

Can a coarse-grained water model capture the key physical features of the hydrophobic effect?

Coarse-grained (CG) molecular dynamics can be a powerful method for probing complex processes. However, most CG force fields use pairwise nonbonded interaction potentials sets, which can limit their ability to capture complex multi-body phenomena such as the hydrophobic effect. As the hydrophobic effect primarily manifests itself due to the nonpolar solute affecting the nearby hydrogen bonding network in water, capturing such effects using a simple one CG site or “bead” water model is a challenge. In this work, we systematically test the ability of CG one site water models for capturing critical features of the solvent environment around a hydrophobe as well as the potential of mean force (PMF) of neopentane association. We study two bottom-up models: a simple pairwise (SP) force-matched water model constructed using the multiscale coarse-graining method and the Bottom-Up Many-Body Projected Water (BUMPer) model, which has implicit three-body correlations. We also test the top-down monatomic (mW) and the Machine Learned mW (ML-mW) water models. The mW models perform well in capturing structural correlations but not the energetics of the PMF. BUMPer outperforms SP in capturing structural correlations and also gives an accurate PMF in contrast to the two mW models. Our study highlights the importance of including three-body interactions in CG water models, either explicitly or implicitly, while in general highlighting the applicability of bottom-up CG water models for studying hydrophobic effects in a quantitative fashion. This assertion comes with a caveat, however, regarding the accuracy of the enthalpy–entropy decomposition of the PMF of hydrophobe association.

MusicFace: Music-driven expressive singing face synthesis

It remains an interesting and challenging problem to synthesize a vivid and realistic singing face driven by music. In this paper, we present a method for this task with natural motions for the lips, facial expression, head pose, and eyes. Due to the coupling of mixed information for the human voice and backing music in common music audio signals, we design a decouple-and-fuse strategy to tackle the challenge. We first decompose the input music audio into a human voice stream and a backing music stream. Due to the implicit and complicated correlation between the two-stream input signals and the dynamics of the facial expressions, head motions, and eye states, we model their relationship with an attention scheme, where the effects of the two streams are fused seamlessly. Furthermore, to improve the expressivenes of the generated results, we decompose head movement generation in terms of speed and direction, and decompose eye state generation into short-term blinking and long-term eye closing, modeling them separately. We have also built a novel dataset, SingingFace, to support training and evaluation of models for this task, including future work on this topic. Extensive experiments and a user study show that our proposed method is capable of synthesizing vivid singing faces, qualitatively and quantitatively better than the prior state-of-the-art.

Improve label embedding quality through global sensitive GAT for hierarchical text classification

Hierarchical text classification aims to assign text to multiple labels in a label set stored in a tree structure. The current algorithms mainly introduce the priori information of the label hierarchy, but the implicit correlation between labels in the hierarchy is rarely applied. At the same time, we also found that the inherent class imbalance of chain labels will also lead to poor classification effects of lower-level labels through a large number of studies. Therefore, a label structure enhanced hierarchy aware global model (LSE-HiAGM) is proposed. Firstly, the common density coefficient of labels is defined to measure the importance of a pair of labels in the hierarchical structure. Secondly, the common density coefficient is used as the weight of the label to update the topological structure features, so that the label can be linked with all labels globally. Finally, the topological structure feature, text features, and label hierarchical features are fused to make full use of all features to improve the embedding quality of low-level labels. In addition, to alleviate the class imbalance problem, a new loss function is used to constrain the model training. The probability of the label being sampled relative to all the labels of the sample is taken as the weight of the loss function. Therefore, a small penalty is imposed on the upper label and a large penalty on the lower label. A large number of experiments on datasets such as RCV1, WOS and NYT show that LSE-HiAGM performs better than the baseline models in hierarchical text classification.

Irradiated mechanical properties predicted by a machine learning method with the Fourier-transform-based feature extension

ABSTRACT High-dimensional nonlinear relationships between the irradiated yield strength and its influencing factors, including doses, temperatures, and crystal structures, are difficult to explicitly characterize in the absence of a comprehensive database. In this study, we developed a machine learning method with the Fourier-transform-based feature extension, successfully constructing the prediction model of irradiated yield strength by a relatively small and sparse database of irradiated material properties. The analysis suggests that the proposed feature extension method improves the training performances of machine learning with small dataset. And the present model is accurate and feasible for predicting the irradiated yielding behaviors. Furthermore, we attempt the inverse machine learning model to determine material properties and irradiation conditions according to the desired yield strength. Since the parameter combinations commensurate with a fixed strength are diverse, the optimal model is helpful in reversely calculating and optimizing material performances. The data-driven machine learning method, which can detect the implicit correlations among numerous data, exhibits great prospects in investigating irradiated mechanical properties and exploring multiscale links in the nuclear material field. This work holds the promise for optimizing the design of in-pile structural components and can be further extended to other machine learning problems with the small dataset.

Boosting Scene Graph Generation with Contextual Information

Scene graph generation (SGG) has been developed to detect objects and their relationships from the visual data and has attracted increasing attention in recent years. Existing works have focused on extracting object context for SGG. However, very few works have attempted to exploit implicit contextual correlations among relationships of the objects. Furthermore, most existing SGG schemes rely on high-level features to predict the predicates while overlooking the potential inherent association of low-level features with the object relationships. We present in this article a novel scheme to capture enhanced contextual information for both objects and relationships. We design a Dual-branch Context Analysis Transformer (DCAT) architecture to extract both object context and relationship context from the visual data with dual transformer branches and then effectively fuse both high-level and low-level features by an adaptive approach to facilitate relationship prediction. Specifically, we first conduct feature representation learning to enrich relation representations by the visual, spatial, and linguistic feature extractors. Next, two transformer branches are designed to leverage the modeling of global associative interaction and mine the hidden association among objects and relationships. Then, we devise a novel feature disentangling method to decouple contextualized high-level features with guidance from the visual semantics. Finally, we develop a refined attention module to perform low-level feature recalibration for the refinement of the final predicate prediction. Experiments on Visual Genome and Action Genome datasets demonstrate the effectiveness of DCAT for both image and video SGG settings. Moreover, we also test the quality of the generated image scene graphs to verify the generalizability on downstream tasks like sentence-to-graph retrieval and image retrieval.

Capturing spatial–temporal correlations with Attention based Graph Convolutional Network for network traffic prediction

Network traffic prediction is essential and significant to network management and network security. Existing prediction methods cannot well capture the temporal–spatial correlations hidden in the network traffic and suffer from a prediction accuracy degradation for two reasons. First, the common recurrent neural network models which are leveraged to learn the temporal relations in network traffic exhibit a poor performance in long-term prediction for its limited receptive field. Second, the existing methods for modeling the spatial relationship of network traffic focus on capturing the distance relationship between nodes in the network topology, which cannot reflect the implicit correlation between network nodes. To tackle these two problems, we propose an Attention-based Graph Convolutional Network model (AGCN) for capturing both the spatial and temporal correlations in network traffic. To catch the hidden spatial dependencies in network traffic, we combine graph attention network with graph convolutional network to mine the spatial relationships of network traffic. To efficiently learn the temporal long-term relations embedded in network traffic, we design a dilated convolution module to enable an exponentially growing receptive field for handling long sequences. Experimental results on three network traffic datasets show that AGCN has excellent performance in terms of prediction accuracy and inference time compared to current mainstream methods.

A hybrid framework based on knowledge distillation for explainable disease diagnosis

Deep learning (DL) techniques provide highly accurate disease predictions by extracting implicit higher-order correlations between features. However, DL opinions are not advisable for direct clinical practice because they lack interpretability in their reasoning logic. Fuzzy rule-based systems (FRBSs) can provide transparent and effective diagnostic results, but the rule explosion problem and the long-tail distribution of rules weaken their performance. Therefore, DL and FRBS complement each other. This paper proposes a hybrid framework based on DL models and fuzzy rules for explainable disease diagnosis. This framework has three components: interpretable component, uninterpretable component, and knowledge distillation component. In the interpretable component, we propose a strong rule extraction method to avoid the rule explosion problem and the long-tail distribution of rules by discarding weak rules. In the uninterpretable component, we use DL models to capture the implicit higher-order interactions between features to improve the model's performance. The knowledge distillation component embeds the hidden knowledge extracted from DL models into SFRBSs to guide the parameter tuning. This framework can predict new instances and provide interpretable prediction results by extracting explicit knowledge (fuzzy rules) and feature weights. Experimental results based on three real-world medical datasets show that the framework achieves the highest AUC and accuracy, which is more accurate and interpretable than other diagnostic models.

CTransCNN: Combining transformer and CNN in multilabel medical image classification

Multilabel image classification aims to assign images to multiple possible labels. In this task, each image may be associated with multiple labels, making it more challenging than the single-label classification problems. For instance, convolutional neural networks (CNNs) have not met the performance requirement in utilizing statistical dependencies between labels in this study. Additionally, data imbalance is a common problem in machine learning that needs to be considered for multilabel medical image classification. Furthermore, the concatenation of a CNN and a transformer suffers from the disadvantage of lacking direct interaction and information exchange between the two models. To address these issues, we propose a novel hybrid deep learning model called CTransCNN. This model comprises three main components in both the CNN and transformer branches: a multilabel multihead attention enhanced feature module (MMAEF), a multibranch residual module (MBR), and an information interaction module (IIM). The MMAEF enables the exploration of implicit correlations between labels, the MBR facilitates model optimization, and the IIM enhances feature transmission and increases nonlinearity between the two branches to help accomplish the multilabel medical image classification task. We evaluated our approach using publicly available datasets, namely the ChestX-ray11 and NIH ChestX-ray14, along with our self-constructed traditional Chinese medicine tongue dataset (TCMTD). Extensive multilabel image classification experiments were conducted comparing our approach with excellent methods. The experimental results demonstrate that the framework we have developed exhibits strong competitiveness compared to previous research. Its robust generalization ability makes it applicable to other medical multilabel image classification tasks.