Traditional Classification Methods Research Articles

Research on the application of graph neural networks in text classification

With the rapid development of the Internet and social media, a large amount of text data is constantly generated, and text classification has become an important task. Text classification is an important branch in the field of natural language processing and has attracted the attention of many researchers in recent years. Due to the ease of editing text data, most texts in the network are manually constructed and uploaded by users. Therefore, the standardization of online texts and the classification of texts with different granularity are of great significance in the field of information retrieval. This article introduces the basic concepts of text classification and three text classification methods: knowledge engineering (rule) based text classification methods, traditional machine learning-based text classification methods, and deep neural network-based text classification methods. This article also analyzes the advantages and disadvantages of the three methods. Subsequently, the basic concepts and common models of graph neural networks were explored. An application case of graph neural networks in text classification is provided. This article provides effective suggestions for text classification applications in the field of natural language processing.

SwitchTab: Switched Autoencoders Are Effective Tabular Learners

Self-supervised representation learning methods have achieved significant success in computer vision and natural language processing (NLP), where data samples exhibit explicit spatial or semantic dependencies. However, applying these methods to tabular data is challenging due to the less pronounced dependencies among data samples. In this paper, we address this limitation by introducing SwitchTab, a novel self-supervised method specifically designed to capture latent dependencies in tabular data. SwitchTab leverages an asymmetric encoder-decoder framework to decouple mutual and salient features among data pairs, resulting in more representative embeddings. These embeddings, in turn, contribute to better decision boundaries and lead to improved results in downstream tasks. To validate the effectiveness of SwitchTab, we conduct extensive experiments across various domains involving tabular data. The results showcase superior performance in end-to-end prediction tasks with fine-tuning. Moreover, we demonstrate that pre-trained salient embeddings can be utilized as plug-and-play features to enhance the performance of various traditional classification methods (e.g., Logistic Regression, XGBoost, etc.). Lastly, we highlight the capability of SwitchTab to create explainable representations through visualization of decoupled mutual and salient features in the latent space.

Deep Incomplete Multi-View Learning Network with Insufficient Label Information

Due to the efficiency of integrating semantic consensus and complementary information across different views, multi-view classification methods have attracted much attention in recent years. However, multi-view data often suffers from both the miss of view features and insufficient label information, which significantly decrease the performance of traditional multi-view classification methods in practice. Learning for such simultaneous lack of feature and label is crucial but rarely studied. To tackle these problems, we propose a novel Deep Incomplete Multi-view Learning Network (DIMvLN) by incorporating graph networks and semi-supervised learning in this paper. Specifically, DIMvLN firstly designs the deep graph networks to effectively recover missing data with assigning pseudo-labels of large amounts of unlabeled instances and refine the incomplete feature information. Meanwhile, to enhance the label information, a novel pseudo-label generation strategy with the similarity constraints of unlabeled instances is proposed to exploit additional supervisory information and guide the completion module to preserve more semantic information of absent multi-view data. Besides, we design view-specific representation extractors with the autoencoder structure and contrastive loss to learn high-level semantic representations for each view, promote cross-view consistencies and augment the separability between different categories. Finally, extensive experimental results demonstrate the effectiveness of our DIMvLN, attaining noteworthy performance improvements compared to state-of-the-art competitors on several public benchmark datasets. Code will be available at GitHub.

Multi-Label Supervised Contrastive Learning

Multi-label classification is an arduous problem given the complication in label correlation. Whilst sharing a common goal with contrastive learning in utilizing correlations for representation learning, how to better leverage label information remains challenging. Previous endeavors include extracting label-level presentations or mapping labels to an embedding space, overlooking the correlation between multiple labels. It exhibits a great ambiguity in determining positive samples with different extent of label overlap between samples and integrating such relations in loss functions. In our work, we propose Multi-Label Supervised Contrastive learning (MulSupCon) with a novel contrastive loss function to adjust weights based on how much overlap one sample shares with the anchor. By analyzing gradients, we explain why our method performs better under multi-label circumstances. To evaluate, we conduct direct classification and transfer learning on several multi-label datasets, including widely-used image datasets such as MS-COCO and NUS-WIDE. Validation indicates that our method outperforms the traditional multi-label classification method and shows a competitive performance when comparing to other existing approaches.

Heart Sound Signals Classification with Image Conversion Employed

The number of patients with cardiovascular diseases worldwide is increasing rapidly, while medical resources are increasingly scarce. Heart sound classification, as the most direct means of discovering cardiovascular diseases, is attracting the attention of researchers around the world. Although great progress has been made in heart sound classification in recent years, most of them are based on traditional statistical feature methods and temporal dimension features. These traditional temporal dimension feature representation and classification methods cannot achieve good classification accuracy. This paper proposes a new partition attention module and Fusionghost module, and the entire network framework is named PANet. Without segmentation of the heart sound signal, the heart sound signal is converted into a bispectrum and input into the proposed framework for feature extraction and classification tasks. The network makes full use of multi-scale feature extraction and feature map fusion, improving the network feature extraction ability. This paper conducts a comprehensive study of the performance of different network parameters and different module numbers, and compares the performance with the most advanced algorithms currently available. Experiments have shown that for two classification problems (normal or abnormal), the classification accuracy rate on the 2016 PhysioNet/CinC Challenge database reached 97.89%, the sensitivity was 96.96%, and the specificity was 98.85%.

Multi-dimensional deep dense residual networks and multiple kernel learning for hyperspectral image classification

To address the issues of inadequate feature expression capacity and poor adaptability of feature fusion in traditional hyperspectral image classification methods, a new approach to classification utilizing a multi-dimensional deep dense residual network and multiple kernel learning is suggested. First, we construct a deep dense residual structure to obtain the spectral, spatial and spectral-spatial features of the image, respectively. Then, the multiple kernel learning is employed to map the three extracted features into the high-dimensional space to achieve the adaptive fusion between features. Finally, multiple kernel support vector machine is adopted to accurately classify the multiple kernel fused features. By randomly selecting 10%, 1% and 1% of samples in each ground object category as training samples, our proposed method demonstrated an enhanced overall classification accuracy of 99.58%, 99.52% and 99.66% according to the experimental results conducted on three typical hyperspectral images. Comparing with some multiple kernel learning based methods, our method can obtain higher classification accuracy. So our method has better recognition ability for hyperspectral image features, and has strong adaptability and robustness.

Intelligent non-cooperative optical networks: Leveraging scattering neural networks with small training data

Artificial intelligence (AI) is enabling intelligent communications where learning based signal classification simplifies optical network signal allocation and shifts signal processing pressure to each network edge. This work proposes a non-orthogonal signal waveform framework that leverages its unique spectral compression characteristic as a user address for efficiently forwarding messages to target users. The primary focus of this work lies in the physical layer intelligent receiver design, which can automatically identify different received signal formats without preamble notification in a non-cooperative communication approach. Traditional signal classification methods, such as convolutional neural network (CNN), rely on extensive training, resulting in a heavy dependency on large training datasets. To overcome this limitation, this work designs a specific two-layer scattering neural network that can accurately separate signals even when the training data is limited, leading to reduced training complexity. Its performance remains robust in diverse transmission conditions. Furthermore, the scattering neural network is interpretable because features are extracted based on deterministic wavelet filters rather than training based filters.

Object-Oriented Convolutional Neural Network for Forest Stand Classification Based on Multi-Source Data Collaboration

Accurate classification of forest stand is crucial for protection and management needs. However, forest stand classification remains a great challenge because of the high spectral and textural similarity of different tree species. Although existing studies have used multiple remote sensing data for forest identification, the effects of different spatial resolutions and combining multi-source remote sensing data for automatic complex forest stand identification using deep learning methods still require further exploration. Therefore, this study proposed an object-oriented convolutional neural network (OCNN) classification method, leveraging data from Sentinel-2, RapidEye, and LiDAR to explore classification accuracy of using OCNN to identify complex forest stands. The two red edge bands of Sentinel-2 were fused with RapidEye, and canopy height information provided by LiDAR point cloud was added. The results showed that increasing the red edge bands and canopy height information were effective in improving forest stand classification accuracy, and OCNN performed better in feature extraction than traditional object-oriented classification methods, including SVM, DTC, MLC, and KNN. The evaluation indicators show that ResNet_18 convolutional neural network model in the OCNN performed the best, with a forest stand classification accuracy of up to 85.68%.

Machine learning in the classification of asian rust severity in soybean using hyperspectral sensor

Traditional monitoring of asian soybean rust severity is a time- and labor-intensive task, as it requires visual assessments by skilled professionals in the field. Thus, the use of remote sensing and machine learning (ML) techniques in data processing has emerged as an approach that can increase efficiency in disease monitoring, enabling faster, more accurate and time- and labor-saving evaluations. The aims of the study were: (i) to identify the spectral signature of different levels of Asian soybean rust severity; (ii) to identify the most accurate machine learning algorithm for classifying disease severity levels; (iii) which spectral input provides the highest classification accuracy for the algorithms; (iv) to determine a sample size of leaves that guarantees the best accuracy for the algorithms. A field experiment was carried out in the 2022/2023 harvest in a randomized block design with a 6x3 factorial scheme (ML algorithms x severity levels) and four replications. Disease severity levels assessed were: healthy leaves, 25 % severity, and 50 % severity. Leaf hyperspectral analysis was carried out over a wide range from 350 to 2500 nm. From this analysis, 28 spectral bands were extracted, seeking to distinguish the spectral signature for each severity level with the least input dataset. Data was subjected to machine learning analysis using Artificial Neural Network (ANN), REPTree (DT) and J48 decision trees, Random Forest (RF), and Support Vector Machine (SVM) algorithms, as well as a traditional classification method (Logistic Regression - LR). Two different input datasets were tested for each algorithm: the full spectrum (ALL) provided by the sensor and the 28 spectral bands (SB). Tests with different sample sizes were also conducted to investigate the algorithms' ability to detect severity levels with a reduced sample size. Our findings indicate differences between the spectral curves for the severity levels assessed, which makes it possible to differentiate between healthy plants with low and high severity using hyperspectral sensing. SVM was the most accurate algorithm for classifying severity levels by using all the spectral information as input. This algorithm also provided high classification accuracy when using smaller leaf samples. This study reveals that hyperspectral sensing and the use of ML algorithms provide an accurate classification of different levels of Asian rust severity, and can be powerful tools for a more efficient disease monitoring process.

Multi-scale representation of surface-enhanced Raman, spectroscopy data for deep learning-based liver, cancer detection

The early detection of liver cancer greatly improves survival rates and allows for less invasive treatment options. As a non-invasive optical detection technique, Surface-Enhanced Raman Spectroscopy (SERS) has shown significant potential in early cancer detection, providing multiple advantages over conventional methods. The majority of existing cancer detection methods utilize multivariate statistical analysis to categorize SERS data. However, these methods are plagued by issues such as information loss during dimensionality reduction and inadequate ability to handle nonlinear relationships within the data. To overcome these problems, we first use wavelet transform with its multi-scale analysis capability to extract multi-scale features from SERS data while minimizing information loss compared to traditional methods. Moreover, deep learning is employed for classification, leveraging its strong nonlinear processing capability to enhance accuracy. In addition, the chosen neural network incorporates a data augmentation method, thereby enriching our training dataset and mitigating the risk of overfitting. Moreover, we acknowledge the significance of selecting the appropriate wavelet basis functions in SERS data processing, prompting us to choose six specific ones for comparison. We employ SERS data from serum samples obtained from both liver cancer patients and healthy volunteers to train and test our classification model, enabling us to assess its performance. Our experimental results demonstrate that our method achieved outstanding and healthy volunteers to train and test our classification model, enabling us to assess its performance. Our experimental results demonstrate that our method achieved outstanding performance, surpassing the majority of multivariate statistical analysis and traditional machine learning classification methods, with an accuracy of 99.38 %, a sensitivity of 99.8 %, and a specificity of 97.0 %. These results indicate that the combination of SERS, wavelet transform, and deep learning has the potential to function as a non-invasive tool for the rapid detection of liver cancer.

Research on Music Classification Technology Based on Integrated Deep Learning Methods

INTRODUCTION: Music classification techniques are of great importance in the current era of digitized music. With the dramatic increase in music data, effectively categorizing music has become a challenging task. Traditional music classification methods have some limitations, so this study aims to explore music classification techniques based on integrated deep-learning methods to improve classification accuracy and robustness.OBJECTIVES: The purpose of this study is to improve the performance of music classification by using an integrated deep learning approach that combines the advantages of different deep learning models. The author aims to explore the effectiveness of this approach in coping with the diversity and complexity of music and to compare its performance differences with traditional approaches.METHODS: The study employs several deep learning models including, but not limited to, Convolutional Neural Networks (CNN), Recurrent Neural Networks (RNN), and Long Short-Term Memory Networks (LSTM). These models were integrated into an overall framework to perform the final music classification by combining their predictions. The training dataset contains rich music samples covering different styles, genres and emotions.RESULTS: Experimental results show that music classification techniques based on integrated deep learning methods perform better in terms of classification accuracy and robustness compared to traditional methods. The advantages of integrating different deep learning models are fully utilized, enabling the system to better adapt to different types of music inputs.CONCLUSION: This study demonstrates the effectiveness of the integrated deep learning approach in music classification tasks and provides valuable insights for further improving music classification techniques. This approach not only improves the classification performance but also promises to be applied to other areas and promote the application of deep learning techniques in music analysis.

Prediction and Diagnosis of Breast Cancer Using Machine and Modern Deep Learning Models.

This research compares traditional machine learning classification methods to innovative techniques that use deep learning models. Established usual classification models such as k-Nearest Neighbors (kNN), Gradient Boosting, Support Vector Machine (SVM), Neural Network, CN2 rule inducer, Naive Bayes, Stochastic Gradient Descent (SGD), and Tree, and deep learning models such as Neural Decision Forest and Multilayer Perceptron used. The investigation, which was carried out using the Orange and Python tools, evaluates their diagnostic effectiveness in breast cancer detection. The evaluation uses UCI's publicly accessible Wisconsin Diagnostic Data Set, enabling transparency and accessibility in the study approach. The mean radius ranges from 6.981 to 28.110, while the mean texture runs from 9.71 to 39.28 in malignant and benign cases. Gradient boosting and CN2 rule inducer classifiers outperform SVM in accuracy and sensitivity, whereas SVM has the lowest accuracy and sensitivity at 88%. The CN2 rule inducer classifier achieves the greatest ROC curve score for benign and malignant breast cancer datasets, with an AUC score of 0.98%. MLP displays distinguish positive and negative classes, with a higher AUC-ROC of 0.9959. with accuracy of 96.49%, precision of 96.57%, recall of 96.49%, and an F1-Score of 96.50%. Among the most commonly used classifier models, CN2 rule and GB performed better than other models. However, MLP from deep learning produced the greatest overall performance.

Comparison of Crop Area Estimation for the Winter Season Using the Google Earth Engine (GEE) Platform and Traditional Method of Supervisor Classification

There is a significant development in the methods and techniques for classifying satellite images during these years. One of the most recently launched methods is using the Google Earth Engine platform for its contribution in providing a huge amount of satellite data with different dates and multiple satellite images sources, with a method of work that is easy to deal with, but it needed to learn the computer language such as Java and Python, which is the method of work depends on writing a script that gives application commands to classify the different satellite image. Therefore, this research aims to apply classification of satellite images using the Google Earth Engine platform for three models Classification and Regression Trees (CART), Support Vector Machine (SVM), and Random Forest (RF), then compare the results calculated from using images classification (Supervised classification: Maximum likelihood) in terms of the degree of accuracy of the classification and matching the results of the classified satellite images in addition. to boundary clarify of the classification methods. This study showed that, using the Google Earth Engine (GEE) platform is considered one of the important methods that facilitated researchers interested in using remote sensing techniques and geographic information systems, due to the availability of this platform of free satellite sources with multiple spatial resolutions, in addition to the presence of a powerful system that collects, processes, and classifies satellite images using different classification models and indexes. But we need to apply these modern techniques at a number of levels, such as at Markas, Governorates, International, and regional levels, to judge their accuracy and suitability for large areas.

Soil Classification and Crop Suggestion using Machine Learning Techniques

In agriculture, the optimization of crop selection based on soil characteristics plays a pivotal role in maximizing yields and ensuring sustainable agricultural practices. Traditional methods of soil classification and crop suggestion rely heavily on manual labour and expertise, often resulting in time-consuming processes with limited accuracy. However, the integration of machine learning techniques presents a promising solution to this challenge

Spam classification based on parallel optimized BERT

With the popularity of email and the increase in spam, effectively filtering spam has become an urgent need. This research presents a cutting-edge approach to spam filtering, leveraging the power of BERT and innovative parallel optimization techniques, improving email security. This study proposes a spam classification method based on the BERT (Bidirectional Encoder Representations from Transformers) model, aiming to improve the accuracy and efficiency of spam filtering. The study first investigated the shortcomings of traditional classification methods and then analysed the applicability of the BERT model to classifying spam. The study also performed two parallel optimizations on BERT, DDP (Distributed Data Parallel), and Gpipe. Among them, DDP was used to accelerate the large model BERT parallelly. After conducting experiments on spam classification, it was found that the average training time of an epoch was 54 seconds, with the accuracy on the test set reaching 98%, achieving significant performance improvements.

Research on multi-role classification task of online mall based on heterogeneous graph neural network

With the rapid development of e-commerce, online shopping malls have become an indispensable part of daily life. In order to better meet the needs of consumers, marketplace platforms need to accurately identify and categorize different user personas to provide personalized services and recommendations. In traditional role classification methods, basic information and behavioral data of users are typically used for classification. However, this approach often ignores the complex relationships between users and multiple heterogeneous data such as goods, reviews, social networks, and more. Therefore, we propose a new approach based on heterogeneous graphs to model different types of data in the form of graphs to better capture the connections between users and various elements in the marketplace. In this study, graph embedding technology is used to map nodes in heterogeneous graphs into low-dimensional vector spaces to capture similarities and relationships between nodes. Then, using the vector representation of these nodes, we can apply algorithms such as attention mechanisms for multi-role classification. Specifically, we use algorithms such as support vector machines to train classification models and use heterogeneous graph attention mechanisms to obtain the final feature representation of nodes. Experimental results show that our method shows significant advantages in multi-role classification tasks. Finally, the results of this study are discussed and summarized. We found that the classification model based on heterogeneous graph can effectively classify multiple roles in the online mall to provide personalized services and recommendations for the mall. At the same time, we also find that the construction of heterogeneous maps and the choice of graph embedding technology have important impacts on the classification results, which need further research and optimization. Therefore, multi-role task classification of online shopping malls based on heterogeneous graph neural networks is of great significance for improving the user experience and recommendation effect of online shopping malls, and also provides new ideas and methods for research in related fields.

Multi-View Scene Classification Based on Feature Integration and Evidence Decision Fusion

Leveraging multi-view remote sensing images in scene classification tasks significantly enhances the precision of such classifications. This approach, however, poses challenges due to the simultaneous use of multi-view images, which often leads to a misalignment between the visual content and semantic labels, thus complicating the classification process. In addition, as the number of image viewpoints increases, the quality problem for remote sensing images further limits the effectiveness of multi-view image classification. Traditional scene classification methods predominantly employ SoftMax deep learning techniques, which lack the capability to assess the quality of remote sensing images or to provide explicit explanations for the network’s predictive outcomes. To address these issues, this paper introduces a novel end-to-end multi-view decision fusion network specifically designed for remote sensing scene classification. The network integrates information from multi-view remote sensing images under the guidance of image credibility and uncertainty, and when the multi-view image fusion process encounters conflicts, it greatly alleviates the conflicts and provides more reasonable and credible predictions for the multi-view scene classification results. Initially, multi-scale features are extracted from the multi-view images using convolutional neural networks (CNNs). Following this, an asymptotic adaptive feature fusion module (AAFFM) is constructed to gradually integrate these multi-scale features. An adaptive spatial fusion method is then applied to assign different spatial weights to the multi-scale feature maps, thereby significantly enhancing the model’s feature discrimination capability. Finally, an evidence decision fusion module (EDFM), utilizing evidence theory and the Dirichlet distribution, is developed. This module quantitatively assesses the uncertainty in the multi-perspective image classification process. Through the fusing of multi-perspective remote sensing image information in this module, a rational explanation for the prediction results is provided. The efficacy of the proposed method was validated through experiments conducted on the AiRound and CV-BrCT datasets. The results show that our method not only improves single-view scene classification results but also advances multi-view remote sensing scene classification results by accurately characterizing the scene and mitigating the conflicting nature of the fusion process.

Predicting lodging severity in dry peas using UAS-mounted RGB, LIDAR, and multispectral sensors

This study investigated the feasibility of using data collected by sensors mounted onto unmanned aerial systems (UASs) and traditional classification methods (Maximum Likelihood, Mahalanobis Distance, Minimum Distance, and Adaptive Coherence Estimator) and four machine learning models (Random Forest [RF], Support Vector Machine [SVM], K-Nearest Neighbor [KNN], and Extreme Gradient Boosting [XGBOOST]) to predict the severity of dry pea lodging. Data were collected in two environments in North Dakota during the 2022 crop season, where lodging was visually estimated at the plot level using a lodging scale range of 0 to 9. Aerial images were collected using two UASs (DJI Matrice 200 and 300) during early, mid, and late stages of growth. Synthetic Minority Oversampling Technique combined with Edited Nearest Neighbors (SMOTE-ENN) was used to balance the imbalanced dataset. Different feature selection models (recursive feature elimination, lasso regression, principal component analysis, and correlation-based methods) were used and compared to identify the most important features in the dataset. Our study found that similar channels can be used interchangeably for analyzing lodging without compromising accuracy, simplifying data acquisition, and saving time and resources. The maximum plant height values obtained from LiDAR and RGB sensors showed a strong correlation with actual values. Recursive feature elimination was the most effective feature selection method for handling high-dimensional datasets. The XGBOOST model outperformed all traditional and machine learning classification models, with the lowest variation in precision values. RF showed the most pronounced overfitting among the machine learning models, while Mahalanobis distance showed the least overfitting among all traditional classification models. Overall, the study demonstrated the potential of UAS and machine learning models for predicting lodging severity in dry peas, with XGBOOST being the best-performing model.

Phylogenetic relationships and genetic differentiation of two Salamandrella species as revealed via COI gene from Northeastern China.

Due to traditional classification methods' limitations, some cryptic species remain undiscovered. To better explore the existence of the Schrenck salamander (Salamandrella tridactyla, a cryptic species of Siberian salamander S. keyserlingii) in China, we conducted a molecular phylogenetic analysis to confirm the taxonomic relationship among Salamandrella species and investigate genetic variation. We used complete sequences of the mitochondrial COI gene from 65 specimens collected across a wide range in Northeastern China. Thirty-five haplotypes were obtained from six populations. They showed medium-high haplotype diversity (Hd) and low nucleotide polymorphism (π). The phylogenetic tree and haplotype network analysis revealed that populations from Greater Khingan Ridge (Huma: HM) and Lesser Khingan Ridge (Tieli: TL) belong to S. keyserlingii, while populations from Changbai Mountain (Shangzhi-zhuziying: SZ, Shangzhi-cuijia: SC, Hailin: HL, and Baishan: BS) belong to S. tridactyla. This indicates the monophyly of Salamandrella and each of the two species. There was a substantial level of genetic differentiation between different species and within populations of the same species. This differentiation was significantly related to geographical distance. At last, the mismatch distribution and neutrality analyses indicated that the TL populations have undergone expansion of history. The study supplements the distributional range of Schrenck salamander. And it provides a theoretical basis for species conservation of Salamandrella species.

Rock Mass Structure Classification of Caves Based on the 3D Rock Block Index

In large-scale water conservancy and hydropower projects, complex rock structures are considered to be the main factor controlling the stability of hydraulic structures. The classification of rock mass structure plays an important role in the safety of all kinds of large buildings, especially underground engineering buildings. As a quantitative classification index of rock mass, the rock block index is very common in the classification of borehole and dam foundation rock mass structures. However, there are few studies on the classification of underground engineering rock masses. Moreover, their classification criteria have disadvantages in spatial dimension. Therefore, this paper takes the long exploratory cave CPD1 in the water transmission and power generation system of the Qingtian pumped storage power station in Zhejiang Province as the research object and launches a study on the structural classification of the rock mass of a flat cave based on the 3D rock block index. According to the group distribution of joints, the sections are statistically homogeneous. Additionally, the Monte Carlo method is used to carry out random simulations to generate a three-dimensional joint network model. The virtual survey lines are arranged along the center of the shape of the three different orthogonal planes of the 3D joint network model to represent the boreholes, and the RBI values of the virtual survey lines on each orthogonal plane are counted to classify the rock mass structure of the flat cave in a refined manner using the rock block index of the rock mass in 3D. The above method realizes the application of the 3D rock block index in underground engineering and overcomes the limitations of traditional rock mass classification methods in terms of classification index and dimension. The results show that: (1) Three-dimensional joint network simulations built on statistical and probabilistic foundations can visualize the structure of the rock mass and more accurately reflect the structural characteristics of the actual rock mass. (2) Based on the 3D rock block index, the rock mass structure of the long-tunnel CPD1 is classified, from that of a continuous structure to a blocky structure, corresponding to the integrity of the rock mass from complete to relatively complete. The classification results are consistent with the evaluation results of horizontal tunnel seismic wave geophysical exploration. (3) Based on the 3D joint network model, it is reasonable and feasible to use the 3D rock block index as a quantitative evaluation index to determine the structure type of flat cave rock masses. The above method is helpful and significant in the classification of underground engineering rock mass structures.