Application of GGNN inference propagation model for martial art intensity evaluation

Mingyuan Xie

doi:10.1515/nleng-2024-0071

Artikel Open Access

Application of GGNN inference propagation model for martial art intensity evaluation

Mingyuan Xie

Veröffentlicht/Copyright: 5. März 2025

Veröffentlicht von

Veröffentlichen auch Sie bei De Gruyter Brill

Manuskript einreichen Informationen für Autor*innen

Aus der Zeitschrift Nonlinear Engineering Band 14 Heft 1

Abstract

Moderate martial art exercise can enhance physical fitness and promote physical and mental health, but excessive exercise intensity may cause damage to the body. To address this issue, this study adopts the forward maximum matching method and a pre-trained conditional random field model to extract semi-structured and unstructured data to construct a knowledge graph for martial arts intensity assessment. A gated graph neural network is utilized to evaluate exercise intensity and to optimize it through the use of parallel computing. The sampling-optimized model exhibited a 0.13 ms reduction in processing time compared to the second-best model, while the model with an optimized training process demonstrated a 6.7 s reduction after 100 rounds of training, outperforming the second-best model. The evaluation accuracy of the strength evaluation model in the three groups was 81.6, 86.6, and 91.6%, respectively. The average evaluation accuracy of the optimized model was 7.2 and 13.4% higher than other models, respectively. The optimized martial art exercise intensity evaluation model has the potential to significantly reduce time consumption, thereby enhancing overall efficiency. By reducing computation time, the model can improve productivity, responsiveness, and user experience. Furthermore, it can improve evaluation efficiency and accuracy, making martial art exercise intensity evaluation more precise. Ultimately, the model can effectively assist individuals in selecting suitable exercises.

Keywords: assessment of exercise intensity; knowledge graph method; GGNN model; parallel computing; gated recurrent network

1 Introduction

The practice of martial arts has been demonstrated to have beneficial effects on both physical strength and moral character. It is a sport that is highly ornamental and competitive, and it constitutes an important part of Chinese culture. Learning martial art is to inherit excellent traditional Chinese culture and promote the Chinese spirit [1,2]. Martial art can be divided into three types based on the intensity of practice: relaxing, moderate, and vigorous, such as Tai Chi, Baduanjin, various boxing routines, as well as confrontational exercises such as Sanda and short combat [3,4]. Moderate martial art exercises can enhance various physical qualities such as flexibility, coordination, and strength, improve immunity, and reduce the risk of illness [5]. However, excessive martial art exercise may cause physical dysfunction, and in severe cases, limb injuries may occur [6]. Therefore, it is necessary to develop a suitable martial art exercise intensity assessment method to recommend appropriate exercise and reduce the risk of injury for people. Traditional martial arts exercise intensity detection methods are usually assessed through subjective, physical, and physiological aspects. Physical detection and subjective feeling intensity are relatively simple. However, they are highly dependent on an individual’s prior experience and require a more sophisticated positioning of the acceleration sensor. The advantage of physiological detection is that it is more scientifically sound. However, it requires more specialized equipment for measurement, which has poor universality. At the same time, existing machine learning-related assessment methods have problems such as poor assessment accuracy, insufficient assessment precision, and low assessment efficiency. Duan et al. proposed a new complementary attention gating network to improve the accuracy of pedestrian motion trajectory prediction. This network used both positive attention and negative attention to capture frequent and special modes in spatial and temporal patterns, and then used learnable weights for summation. This network could accurately estimate the distribution of multiple motion trajectories, and its performance in terms of average displacement error and final displacement error was superior to existing methods by 13.8 and 10.4%, respectively [7]. Zignoli proposed a new machine learning technique to improve the assessment accuracy of cardiopulmonary exercise tests. The technique combined convolutional neural networks, conditionally generated adversarial neural networks, and cooperative game theory to visualize and interpret the model. Experiments showed that the technique achieved expert-level accuracy in exercise intensity classification [8]. Biró et al. proposed an exercise monitoring method based on machine learning techniques to improve the safety of physical activity and reduce the risk of injury. The method employed machine learning to predict key body metrics in sports and used 3D acceleration sensors to predict the body’s acquired intensity. Experiments showed that the method achieved more than 97% prediction accuracy in sports such as running and soccer [9]. Zhang et al. proposed a new framework based on graph neural networks to address the issues of short duration of micro expressions and low exercise intensity. This framework used spatial graph aware neural networks to extract spatial and spectral domain features from images. It could effectively improve the performance of facial macro and micro expression recognition, and accurately recognized micro expressions in different image resolutions and formats [10].

Bloemheuvel et al. proposed a new gated convolutional network (GCN) architecture to reduce the dependence of deep learning prediction on time series length. This architecture used GCN to handle long time series in multivariate regression tasks. The root mean square error of this framework was reduced by 16.3% compared to the baseline method, and the required input data were 50% less than the baseline method [11]. Dang et al. proposed a new recognition method to improve the efficiency of spatiotemporal feature extraction. This method used hierarchical mining of node relationships and employed residual GCN and channel attention modules in the spatial module to extract temporal features of different scales. Its calculation speed was significantly faster than other methods, and it could effectively recognize human movements [12]. Sheng et al. proposed a new graph-based spatiotemporal convolutional network, which used GCN to process spatial interactions, captured temporal features using convolutional neural networks, encoded and decoded spatiotemporal features through gated recurrent unit networks. The prediction error and inference speed of this network were superior to other algorithms, and the model size was also smaller than other methods [13]. Hu et al. proposed a new spatiotemporal GCN based on graph learning to address the complex spatial and temporal dependencies in transportation networks. This network adopted a graph learning module to capture the dynamic spatial relationships of the traffic network and used an extended causal convolutional network with gating mechanism to capture long-term temporal correlations in traffic data, which had better performance [14].

In summary, existing research methods have explored the accuracy and precision of exercise intensity assessment from multiple perspectives and have achieved certain results. However, current methods have problems such as strong subjectivity and low assessment accuracy. Therefore, this study proposes an optimized gated graph neural network (GGNN) based on knowledge graph. This method employs an innovative approach that utilizes the forward maximum matching method and a pre-trained conditional random field model to extract semi-structured and unstructured data, respectively. The aim of this study is to construct a knowledge graph for the evaluation of martial art exercise intensity. Parallel computing is used to optimize the sampling and training evaluation process of the GGNN. The research aims to improve the accuracy of assessing martial art exercise intensity, reduce computation time, and enhance assessment efficiency. This study is divided into three parts. The first part is to construct a knowledge graph and improve the GGNN. The second part is to evaluate the performance of optimization methods. The third part is a summary of this study and prospects for future research directions.

2 GGNN model for martial art exercise intensity evaluation

2.1 Construction of knowledge graph for martial art exercise intensity

The background of martial art is profound and rich, integrating various elements of traditional Chinese culture, philosophical thinking, and sports, which combines knowledge and practice of kicking, hitting, throwing, equipment attack, and defense, as well as related medical care. It not only focuses on physical exercise but also pays attention to spiritual cultivation. Martial art practice is not only a physical training but also a spiritual tempering. It integrates the ideas of Chinese philosophy, such as Taoism and Confucianism, emphasizing inner peace and self-control, and achieving spiritual sublimation through physical training. Martial art practice includes various boxing techniques, equipment use, etc. Each technique contains profound cultural significance and philosophical ideas. With the development of the times, the form and function of the martial art movement are constantly evolving, but its core values and spiritual connotations remain unchanged. Reasonable training intensity is crucial for improving athletes’ competitive level and protecting their health during training and competition. Therefore, it is necessary to conduct research on the evaluation of martial art training intensity and provide scientific training guidance for coaches and athletes.

Martial art sports can be divided into two categories based on their forms: routine sports and combat sports. Routine sports, in addition to being able to compete on stage, can also be improved to enhance physical fitness and strengthen the body, making them suitable for people of different ages and physical states [15]. When engaging in martial art exercise, it is necessary to choose the appropriate exercise according to each person’s physical condition. It is evident that excessive exercise intensity can precipitate physical damage, including joint and muscle strain. Furthermore, long-term high-intensity exercise has the potential to result in a reduction in body capacity, damage to the respiratory system, and, in extreme cases, the onset of chronic disease [16]. Therefore, it is necessary to conduct a martial art exercise intensity assessment to select suitable training methods for individuals, improve the scientificity, safety, and effectiveness of exercise training, promote personal health, and enhance exercise performance. This study uses a knowledge graph with simpler data expression, richer semantic information, and more intuitive relationships with entities to evaluate the intensity of martial art exercise. The evaluation process is shown in Figure 1.

Figure 1

Evaluation process of martial art exercise intensity based on knowledge graph.

In Figure 1, basic knowledge of the martial art movement is obtained through relevant data queries and cognitive consensus. A combination of top-down and bottom-up approaches is used to construct the knowledge graph, and real-time data from various martial art movements are integrated into the knowledge graph. After the construction is completed, GGNN is used to build an inference propagation model, optimize the inference training process of the model, and finally the evaluation results of the intensity of the martial art motion are output. However, existing knowledge graph constructions often use single source data, which can lead to issues such as limited coverage of knowledge domains. Therefore, the study adopts more complex and comprehensive multi-source data to construct a knowledge graph. During construction, information from different architectures is mutually verified, deepening the connections between each other and optimizing conflicting terms in the architecture [17]. The specific process of constructing a knowledge graph is shown in Figure 2.

Figure 2

Construction process of martial art sports knowledge map.

In Figure 2, relevant basic knowledge is obtained through methods such as literature review, guideline consensus, and consulting experts. Comma separated values are used for data storage to provide data for the knowledge graph. The acquired knowledge is modularized and structured, creating entities of the knowledge graph in a top-down manner, defining semantic relationships, and adding attributes to entities. The bottom-up approach is adopted to update acquired knowledge, including the deletion and modification of erroneous and outdated knowledge, as well as the addition of new knowledge. Cypher language is used to modify existing nodes, edges, or attributes in a knowledge graph. The evaluation indicators for martial art exercise intensity generally include rating of perceived exertion (RPE), heart rate (HR), blood lactate, oxygen consumption, and metabolic equivalent of task. However, an individual’s exercise intensity evaluation is related to multiple factors such as age and physical fitness. The node information in a knowledge graph is constructed based on these influencing factors, including name, gender, age, and past injury history [18]. Personal physical fitness node information includes height and weight, body fat percentage, exercise status, dietary habits, and unhealthy habits. The information of exercise time nodes includes exercise schedule, rest time after exercise, and repetition frequency. The information pertaining to the sports space nodes encompasses the surrounding environment during the act of exercise and whether or not equipment is utilized. Personal psychological factor node information includes neural excitement level, psychological fatigue, and RPE index. The node information for evaluating exercise intensity includes multiple physiological data such as HR and blood pressure. The exercise suggestion node information includes training items, training time, and training load. After the construction of node information is completed, the semantic relationships between each node of the entity are defined.

Due to the wide range of data sources and diverse forms of expression, most of them exist in three ways: unstructured, semi-structured, and structured. If the same knowledge extraction method is used, there may be problems such as ineffective definition of the final data, so targeted knowledge extraction methods need to be adopted. Structured data during martial art exercise can be directly converted into a two-dimensional table due to their high correlation with exercise intensity assessment. Semi structured data includes individuals’ physical fitness or past injury history, which are greatly influenced by their subjective factors. Therefore, this study uses the positive maximum matching method. First, a dictionary containing all node information is determined to ensure that the segmentation can recognize all data fields [19]. The length of the longest word in the dictionary is taken as the maximum matching length, which is used for matching. In the event of a successful match, the segment is divided and the matching length is gradually reduced to facilitate the continuation of the matching process. The final segmentation result is then output. For unstructured data, the bidirectional encoder representations from transformers (BERT) conditional random field model is adopted for extraction [20]. The BERT component within the model is responsible for generating a high-dimensional feature representation of each word within the sequence. This representation is then passed to the CRF layer, which considers both the transition probability between labels and the features of each word to predict the optimal label path for the entire sequence.

2.2 Construction of exercise intensity inference propagation model based on GGNN

In the actual construction of knowledge graphs, there are usually problems such as missing relationships or attributes and errors, and knowledge reasoning can identify and correct errors. In this study, GGNN is used for knowledge reasoning, and the running process of the GGNN inference propagation model is shown in Figure 3.

Figure 3

Inference propagation flow of martial art motion intensity evaluation by GGNN.

In Figure 3, state initialization is performed based on the characteristics of each node, and parameters are updated through information exchange with adjacent nodes. The update gate and reset gate in the gating loop are used to control information flow. The update gate determines the degree of node state update, while the reset gate determines the degree of ignoring previous information. Nodes update their state based on their own state and the information they obtain. GGNN undergo multiple iterations to enable nodes to obtain information from more distant nodes, thereby making the output results more comprehensive. The GGNN propagation model is recursively processed. First, the model is initialized and calculated as shown in Eq. (1) [21].

(1) H v = ( x v T , 0 ) T ,

where H v represents the initialized model. x v represents the input feature. T represents the information propagation depth of the model. If the dimension of the adjacency matrix is greater than x v ’s dimension, the remaining part of the adjacency matrix is filled with zeros. The information propagation calculation of the model is shown in Eq. (2) [22].

(2) a v = A v T ( h 1 ( t − 1 ) T , … , h ∣ v ∣ ( t − 1 ) T ) T + b ,

where a v represents the amount of model information transmission. A represents the adjacency matrix. h 1 ( t − 1 ) represents the state of node 1 at time step ( t − 1 ) . h ∣ v ∣ ( t − 1 ) indicates the state of node v at time step ( t − 1 ) . b represents the edge feature dimension of the information passing through the edges. The model performs information forgetting calculation through the update gate, as shown in Eq. (3).

(3) z v t = σ ( Q z a v t + P z h v ( t + 1 ) ) ,

where z v t represents the update gate. σ represents the activation function. Q represents the weight matrix of node state transformation. P represents the weight matrix of neighbor state aggregation. a v t represents the feature of node v at time step t . The information update calculation of the reset gate is shown in Eq. (4) [23].

(4) r v t = σ ( Q r a v t + P r h v ( t − 1 ) ) ,

where r v t represents the reset gate. The updated information is calculated as shown in Eq. (5).

(5) h ˆ v t = tanh ( Q a v t + P ( r v t ⊙ h v ( t − 1 ) ) ) h ˆ v t = tanh ( Q a v t + P ( r v t ⊙ h v ( t − 1 ) ) ) ,

where h ˆ v t represents the updated information. tanh represents the activation function. ⊙ represents the multiplication of each element step by step. The final state calculation of the node is shown in Eq. (6).

(6) h v t = ( 1 − z v t ) ⊙ h v ( t − 1 ) + z v t ⊙ h ˆ v t ,

where h v t represents the final state of the updated node. ( 1 − z v t ) ⊙ represents selective forgetting of some information. z v t ⊙ represents recording the updated new information. The graph level representation vector calculation of the GGNN is shown in Eq. (7) [24].

(7) h ζ = tanh ∑ v ∈ V σ [ i ( h v T , x v ) ] ⊙ tanh [ j ( h v T , x v ) ] ,

where h ζ represents a graph level representation vector. i represents a neural network that takes the concatenation result of h v T and x v as the input real valued vector. j represents a neural network that takes the concatenation result of h v T and x v as the output real valued vector. The model uses graph attention network (GAT) to extract neighborhood semantic information and achieve node level feature learning. In GATs, scalar attention coefficients act on the global features of adjacent nodes. When the model infers a multi-semantic graph, it needs to complete the global feature decomposition of all adjacent nodes and also pay attention to the fine-grained local semantics [25]. The root node in the model can be represented as a process of valuable information collection in adjacent domains from bottom to top, and the similarity of the collection process in the model is shown in Eq. (8).

(8) Y i j = a T ( Q v root i ∥ Q v root j ) ,

where Y i j represents the collection process of the model. v root i represents the root representation of node i . v root j represents the root representation of node j . This study sets the attention matrix of the GAT as global attention and the attention matrix that focuses on local semantics as local attention. The global matrix adjustment of the model is performed using local matrices, as shown in Eq. (9) [26].

(9) O ′ = L β + O ( 1 − β ) ,

where O ′ represents the adjusted global matrix. O represents the global matrix. L is the local attention matrix. β represents the importance adjustment parameter. Finally, the update of node expression is completed through information transmission, and the calculation is shown in Eq. (10).

(10) h i ′ t + 1 = σ ∑ j ∈ N i F i j Q h i ′ t ,

where h i ′ t + 1 represents the expression of node i at time step t + 1 . N i represents the set of all adjacent nodes of node i . F i j represents the information transmitted from node i to node j . h i ′ t represents the expression of node i at time step t . When calculating the model, the cross-entropy loss of each node in the knowledge graph is calculated as shown in Eq. (11) [27].

(11) U = 1 n ∑ i = 1 n − ∑ j = 1 C y i m log soft max m v root , i ,

where U represents the cross-entropy loss of each node. n represents the total number of nodes in the knowledge graph. y i j represents the true label of the i th node in the m -class node. C represents the total number of types of nodes. The calculation of the comprehensive evaluation index for exercise intensity is shown in Eq. (12) [28].

(12) E = ∑ v i ⋅ w i , i ∈ [ 1 , m ] ,

where E represents the comprehensive evaluation index of exercise intensity. v i represents the initial value of the evaluation factor. w i represents the weight size corresponding to the evaluation factor. The larger the value of E , the greater the corresponding user’s martial art exercise intensity.

2.3 Optimization of GGNN process based on parallel computing

GGNNs are better at dealing with long-term dependence problems. In the evaluation of exercise intensity, individual exercise habits and historical data have an important impact on the evaluation of current exercise intensity, and the increased gating mechanism can remember key motion features while ignoring irrelevant information. Therefore, it is widely used in the assessment of exercise intensity. However, existing GGNN consumes longer time and higher costs in training inference, sampling, and final evaluation. Therefore, this study proposes parallel computing to optimize it. The training and sampling process of GGNN is shown in Figure 4.

Figure 4

GGNN training and evaluation process.

In Figure 4, after initialization is completed, the neural network will be trained n times. After reaching the training limit, it will enter the evaluation step to calculate whether the current neural network’s prediction accuracy meets the requirements or whether it meets the maximum iteration limit. If so, the training evaluation process is complete, otherwise the training steps will be restarted. In the training process, different training methods will be selected based on the size of the graphics processor memory and dataset used in this study. When the graphics processor memory is sufficient, the full data training method will be adopted to fully learn the graph structure. When there is insufficient memory, a sampling-based batch training method will be used to reduce memory usage and improve computational efficiency. However, due to the low sampling and data propagation efficiency of existing graph neural networks, a parallel computing-based sampling process optimization method is proposed. The optimization method is shown in Figure 5.

Figure 5

Sampling process optimization based on parallel computing.

In Figure 5, before optimization, data sampling, data transmission, and GPU calculations are performed in batches, and the next batch of calculations will start after the previous batch is completed. After optimization, the asynchronous execution feature based on CUDA Stream allows the CPU and GPU to work simultaneously, enabling the sampling and transmission processes to be synchronized. The training process is divided into two parts: the main thread and the background thread. Sampling and data transmission are carried out in the background thread, and the data from the background thread is transmitted to the main thread through a queue, which then completes GPU calculations. After the main thread receives the relevant data, the background thread will proceed with the second round of sampling and data transmission process. The computational speed improvement after optimizing the sampling process is shown in Eq. (13) [29].

(13) Tar = M ( M − 1 ) ⋅ max ( X , Z , 1 − X − Z ) + 1 ,

where Tar represents the theoretical acceleration ratio before and after optimization. M represents the total number of batches calculated. X represents the total time consumed during the sampling process. Z represents the total time consumed during data transmission. The optimization of the training evaluation process based on parallel computing proposed in the study is shown in Figure 6.

Figure 6

Optimization of training evaluation process based on parallel computation.

In Figure 6, the entire training and evaluation process is divided into three parts: the main process, the training process, and the evaluation process. The primary process is tasked with initiating the training and evaluation process and overseeing its implementation to facilitate prompt adjustments and guarantee the effective execution of the training process. The training process is responsible for performing the requisite training steps and saving the trained model to a local file. The evaluation process, in turn, is tasked with monitoring the local folder and performing the evaluation step when it encounters a new, unevaluated model file. The primary process is responsible for allocating the requisite devices for the training and evaluation phases, respectively, and subsequently initiates the training phase. The status of the training and evaluation processes is queried, and if both have been completed, the training ends. If the training process is completed but the evaluation process is not completed and the device efficiency of the training process is higher, the unfinished evaluation process will be moved to the training device. The training process determines the end flag passed by the evaluation process. If it is true, the training ends and the model is saved. Otherwise, it determines whether the maximum number of iterations has been reached. If it is true, the training ends. Upon completion of the evaluation process and satisfaction of the termination criteria, the results should be conveyed to the training process. Otherwise, the next round of information will be received and the evaluation will continue. The computational speed improvement after optimizing the training evaluation process is shown in Eq. (14) [30].

(14) Tar = N ( N − 1 ) ⋅ max ( X , 1 − X ) + 1 ,

where Tar represents the increase in computing speed. N represents the total number of runs of the training and evaluation process. The maximum memory usage of the GGNN is calculated as shown in Eq. (15).

(15) K = c + b 1 x 1 + b 2 x 2 ,

where K represents the maximum occupied memory. x 1 represents the number of vertices in the sampled graph. x 2 represents the number of edges in the sampled graph. c represents the constant term in the model. b 1 represents the weight coefficient corresponding to the number of vertices. b 2 represents the weight coefficient corresponding to the number of edges.

3 Experimental analysis of GGNN inference propagation model for evaluating exercise intensity

3.1 Performance testing of optimized GGNN modelt

This study used the UCF101 dataset and MADS dataset, and divided them into training set, validation set, and test set according to a 6:1:3 ratio. This study used GCN, GAT, and graph autoencoders (GAE) as comparative algorithms. The performance comparison of the GGNN model after optimizing the sampling process is shown in Table 1.

Table 1

Performance comparison of GGNN model after sampling process optimization

	Each phase of the algorithm in UCF101 takes time (ms)			Each stage of the algorithm in MADS takes time (ms)
Algorithm	GPU computing	Data transmission	Sampling	GPU computing	Data transmission	Sampling
GCN	5.4	0.9	4.8	5.9	1.4	5.3
GAT	11.6	0.9	7.2	8.2	1.3	6.4
GAE	14.3	0.9	5.7	13.6	1.3	5.9
GGNN	5.9	0.9	4.8	5.6	1.3	5.1
GGNN after optimization	5.9	0.0	0.8	5.6	0.0	1.8

In Table 1, due to the complexity relationship of the model being GAE > GAT > GGNN > GCN, the more complex the model, the longer the GPU computation time. After optimizing the GGNN model, the GPU computation time did not change. In the UCF101 dataset, the sampling time decreased from 4.8 to 0.8 ms, a decrease in 4.0 ms. In the MADS dataset, due to its larger size, the sampling time of the GGNN model before optimization was 5.1 ms, which was reduced to 1.8 ms after optimization, a decrease in 3.3 ms. The optimized data transmission was performed simultaneously with the sampling process, so time was not calculated separately. In the UCF101 dataset, the optimized algorithm saved a total of 4.9 ms in the sampling process, while in the MADS dataset, the optimized model saved a total of 4.6 ms in the sampling process. The optimization of the sampling process was applied to the other three algorithms, with a total batch count of 50. The theoretical acceleration ratio and actual situation comparison of different algorithms are shown in Figure 7.

Figure 7

Comparison between theoretical acceleration ratio of different algorithms and actual situation: (a) UFC101 and (b) MADS.

In Figure 7(a), the theoretical acceleration ratio was inversely proportional to the proportion of sampling time and data transmission time. The greater the complexity of the model, the higher the GPU computation time, and the larger the denominator, the lower the theoretical acceleration ratio. The maximum actual acceleration ratio of the GGNN model was 1.62, which was 0.13, 0.17, and 0.31 higher than GCN, GAT, and GAE, respectively. The GAE model had the closest theoretical acceleration ratio to the actual acceleration ratio and the best optimization effect. In Figure 7(b), due to the increased complexity of the dataset, the actual acceleration ratio of the GGNN model was slightly lower than that of the GAT model, which was 0.08 and 0.09 higher than that of the GAE and GCN models, respectively. The impact of different total batch times on model performance is shown in Table 2.

Table 2

Effects of different total batch times on model performance

	GGNN			GGNN after optimization
Total lot number	GPU computing (ms)	Data transmission (ms)	Sampling (ms)	GPU computing (ms)	Data transmission (ms)	Sampling (ms)
25	5.7	0.6	3.9	5.4	0.0	0.5
50	5.9	0.9	4.8	5.9	0.0	0.8
75	6.0	2.5	8.4	7.2	0.0	1.7
100	6.1	2.9	11.7	13.4	0.0	2.6
125	6.2	4.8	15.6	17.2	0.8	3.5
150	6.3	5.2	19.4	21.5	1.4	6.4
175	6.4	5.6	26.3	24.3	1.7	5.7

In Table 2, the time consumed by each step increased with the total number of batches. Before optimization, the GGNN model showed the fastest improvement in the time consumption of each step, which increased from 3.9 to 26.3 ms per batch. The fastest time improvement after optimization was the GPU computing step, which had increased from 5.4 to 24.3 ms per batch. When the total number of batches was the same, the average total time consumption of each batch in the optimized GGNN model was 4.2 ms lower than before optimization, indicating that the optimized model can effectively improve computational efficiency. Due to the short duration of each round of training and evaluation, the total time spent on 100 rounds of training evaluation was recorded. The performance comparison of the optimized model after training and evaluation process is shown in Figure 8.

Figure 8

Comparison of model performance after optimization of training evaluation process: (a) UFC101 and (b) MADS.

In Figure 8(a), there was no linear relationship between the training evaluation steps before and after optimization and the complexity of the model. The GGNN model had the shortest time consumption, with pre- and post-optimization times being 3.4 and 6.7 s less than the second-best model, respectively. After optimization, the time consumption was reduced by 3.1 s compared to before optimization. In Figure 8(b), the optimization time of the GGNN model was relatively small compared to GAT, and the total optimization time was 7.6 s less than GAT. The impact of the number of training rounds on the performance of the model after optimizing the training evaluation process is shown in Figure 9.

Figure 9

Influence of the number of training rounds on the performance of the model after the optimization of the training evaluation process: (a) UFC101 and (b) MADS.

In Figure 9(a), the theoretical and actual acceleration ratios of the GGNN model approached convergence when the number of training rounds reached 2,000, with convergence values of 1.57 and 1.14, respectively. The maximum ratio of optimization effects was 0.72. In Figure 9(b), due to the increasing complexity of the dataset, the number of convergence rounds for theoretical and practical acceleration ratios was around 3,000, with convergence values of 1.45 and 1.41, respectively. The maximum ratio of optimization effects was 0.68. The impact of relative batch size on model performance after the superposition of two optimization processes is shown in Figure 10.

Figure 10

Impact of relative batch size after superimposing two optimization processes on model performance: (a) UFC101 and (b) MADS.

In Figure 10(a), optimization A referred to the optimization of the sampling process, optimization B referred to the optimization of the training evaluation process, and the total training time gradually decreased with the increase in the relative batch size. The performance of optimization A + B was always the best, with an average time that was 8.4 and 3.6 s less than optimization A and optimization B, respectively. In Figure 10(b), the superposition optimization effect fluctuated at a relative batch size of 10%, which was caused by training errors. However, the optimization time for A + B was always the least, with an average time that was 7.2 and 2.5 s less than optimization A and B, respectively.

3.2 Experimental analysis of martial art exercise intensity evaluation of optimized GGNN model

The martial art exercise for testing included some intense activities, so the experimental population was selected from 20 middle-aged people, 20 teenagers, and 20 young people. The physical condition of the experimental population was good and there were no major physical illnesses. Before the experiment, basic information such as the name, age, height, and weight of the testers were recorded. The testers were required to perform sufficient warm-up activities before each test to prevent unexpected situations from occurring. The experiment selected relaxing martial art Tai Chi, moderate martial art Baguazhang, and full-strength martial art boxing based on the intensity of the exercise. To prevent personal injury, the testing was conducted using a single player routine. The comparative methods used in the study included physiological indicator method (PIM), RPE, and knowledge graph method (KGM). The study selected the accuracy rate as the evaluation index, as this metric can reflect the performance of the model. A higher accuracy rate indicates that the predicted results are closer to the actual structure. Accuracy rate provided a clear performance measure for testers, which was easy to understand and apply. The optimized GGNN model for assessing the intensity of martial art exercise is shown in Table 3.

Table 3

Evaluation results of martial art exercise intensity of GGNN model after normalization

Age group	Test dataset	Martial art exercise intensity	Evaluation result			Average accuracy (%)
Age group	Test dataset	Martial art exercise intensity	Low intensity	Medium strength	High strength	Average accuracy (%)
Middle age	20	Low intensity	16	4	0	81.6
	20	Medium strength	2	14	4
	20	High strength	0	1	19
Teenager	20	Low intensity	19	1	0	86.6
	20	Medium strength	2	17	1
	20	High strength	0	3	16
Young person	20	Low intensity	20	0	0	91.6
	20	Medium strength	2	17	1
	20	High strength	0	2	18

In Table 3, the overall accuracy of the GGNN model based on knowledge graph for assessing the intensity of martial art exercise was 86.6%. The accuracy of intensity assessment for middle-aged, young, and adolescent individuals was 81.6, 86.6, and 91.6%, respectively. The middle-aged group exhibited the lowest accuracy rate, reflecting the considerable inter-individual variability in physical fitness and subjective perception of exercise intensity, even at low levels. Four test subjects classified the intensity as moderate, despite the exercise being below the recommended threshold. The accuracy rate for teenagers is the highest, with less variation in subjective perception of exercise intensity among individuals. The comparison of the intensity assessment results of different methods for martial art exercise is shown in Figure 11.

Figure 11

Comparison of martial art exercise intensity evaluation results of different methods: (a) The juvenile population, (b) youth population, and (c) middle-aged people.

In Figure 11, when combined with the three subplots (a), (b), and (c), the fluctuation of RPE was the largest, with even greater fluctuations in the adolescent population, as RPE was more influenced by the tester’s experience. The test results of PIM were greatly influenced by individual qualities, and the fluctuations increased with age. The KGM had the smallest fluctuation and was relatively close to the test results of PIM. Therefore, the KGM could effectively evaluate the intensity of martial art exercise. The accuracy of assessing the intensity of martial art exercise using different evaluation models is shown in Figure 12.

Figure 12

Accuracy of martial art exercise intensity evaluation with different evaluation models: (a) middle-aged people and (b) youth group.

In Figure 12(a), the test results of three types of graph neural networks combined with KGM showed that the evaluation accuracy gradually decreased with the increase in exercise intensity. GGNN had the highest evaluation accuracy, with average values of 7.2 and 13.4% higher than GCN and GAE, respectively. In Figure 12(b), the average evaluation accuracy of GGNN was 8.1 and 10.9% higher than that of GCN and GAE, respectively. The accuracy was the lowest at exercise intensity level 5, and increased as exercise intensity continued to increase.

4 Conclusion

Aiming at the problems of low accuracy and long time consumption of existing methods for assessing the intensity of martial art exercise, a model for assessing the intensity of martial art exercise based on knowledge graph and improved GGNN was proposed. The optimized GGNN model sampling process saved 4.9 and 4.6 ms in the UCF101 dataset and MADS dataset, respectively, with theoretical acceleration ratios 0.13, 0.17, and 0.31 higher than GCN, GAT, and GAE, respectively. In two datasets, the total time required for training and evaluating the optimized GGNN model was 6.7 and 7.6 s less than the second-best model, respectively. In the UFC101 dataset, the acceleration ratio of the GGNN model was close to convergence after 2,000 rounds of training, with an actual acceleration ratio of 1.14 and a maximum optimization effect of 0.74. The average time consumption after combining the two optimization processes was 8.4 and 3.6 s less than optimizing A and B separately. The GGNN model, based on a knowledge graph, demonstrated accuracy rates of 81.6, 86.6, and 91.6% in intensity assessment among three groups, respectively. This was due to significant differences in individual physical fitness and subjective perceptions of exercise intensity. The KGM, which comprehensively considered the characteristics of PIM and RPE methods, had the smallest fluctuation in evaluation results, thereby further indicating that the KGM could effectively evaluate the intensity of martial art movements. In the test results of the three graph neural networks combined with the KGM, the average evaluation accuracy of GGNN was 7.2 and 13.4% higher than that of GCN and GAE, respectively. There are still some issues with this study. For example, in the experiment, the model collects very comprehensive information and the data are relatively rich. However, in actual application, it is difficult to obtain some information and the performance of the model will correspondingly decrease when there is a lack of information. Therefore, in the subsequent optimization of the KGM, it is necessary to increase the knowledge of the exercise intensity evaluation path and optimize the prediction speed and accuracy. At the same time, more convenient information collection devices should be added to further improve the accuracy of exercise intensity assessment.

Funding information: The research is supported by: Zhengzhou University of Industrial Technology’s Key Research Project at the School Level, Research on the actual difficulties and corresponding paths of online exams for public physical education courses in universities based on core competencies (No. 2023ZD011); Humanities and Social Sciences Research Project of Henan Provincial Department of Education, Research on the repositioning and ecological system construction of ideological and political teaching in physical education courses in universities with the concept of “fence consciousness” (No. 2024-ZDJH-540).
Author contributions: The author has accepted responsibility for the entire content of this manuscript and approved its submission.
Conflict of interest: The author states no conflict of interest.
Data availability statement: All data generated or analyzed during this study are included in this published article.

References

[1] Brissman E, Johnander J, Danelljan M, Felsberg M. Recurrent graph neural networks for video instance segmentation. Int J Comput Vis. 2023;131(2):471–95.10.1007/s11263-022-01703-8Suche in Google Scholar

[2] Bessadok A, Mahjoub MA, Rekik I. Graph neural networks in network neuroscience. IEEE Trans Pattern Anal Mach Intell. 2022;45(5):5833–48.10.1109/TPAMI.2022.3209686Suche in Google Scholar PubMed

[3] Firdaus M, Singh GV, Ekbal A, Bhattacharyya P. Affect-GCN: a multimodal graph convolutional network for multi-emotion with intensity recognition and sentiment analysis in dialogues. Multimed Tools Appl. 2023;82(28):43251–72.10.1007/s11042-023-14885-1Suche in Google Scholar

[4] Ahmed A, Sadanandan SK, Pandey S, Basumallik S, Srivastava AK, Wu Y. Event analysis in transmission systems using spatial temporal graph encoder decoder (STGED). IEEE Trans Power Syst. 2022;38(6):5329–40.10.1109/TPWRS.2022.3226209Suche in Google Scholar

[5] Raabe D, Nabben R, Memmert D. Graph representations for the analysis of multi-agent spatiotemporal sports data. Appl Intell. 2023;53(4):3783–803.10.1007/s10489-022-03631-zSuche in Google Scholar

[6] Kiningham K, Levis P, Ré C. GRIP: A graph neural network accelerator architecture. IEEE Trans Comput. 2022;72(4):914–25.10.1109/TC.2022.3197083Suche in Google Scholar

[7] Duan J, Wang L, Long C, Zhou S, Zheng F, Shi L, et al. Complementary attention gated network for pedestrian trajectory prediction. Proc AAAI Conf Artif Intell. 2022;36(1):542–50.10.1609/aaai.v36i1.19933Suche in Google Scholar

[8] Zignoli A. Machine learning models for the automatic detection of exercise thresholds in cardiopulmonary exercising tests: from regression to generation to explanation. Sensors. 2023;23(2):826–45.10.3390/s23020826Suche in Google Scholar PubMed PubMed Central

[9] Biró A, Szilágyi SM, Szilágyi L, Martín-Martín L, Cuesta-Vargas AI. Machine learning on prediction of relative physical activity intensity using medical radar sensor and 3D accelerometer. Sensors. 2023;23(7):3595–618.10.3390/s23073595Suche in Google Scholar PubMed PubMed Central

[10] Zhang J, Sun G, Zheng K, Mazhar S, Fu X, Li Y, et al. SSGNN: A macro and microfacial expression recognition graph neural network combining spatial and spectral domain features. IEEE Trans Hum-Mach Syst. 2022;52(4):747–60.10.1109/THMS.2022.3163211Suche in Google Scholar

[11] Bloemheuvel S, Van den, Hoogen J, Jozinović D, Michelini A, Atzmueller M. Graph neural networks for multivariate time series regression with application to seismic data. Int J Data Sci Anal. 2023;16(3):317–32.10.1007/s41060-022-00349-6Suche in Google Scholar

[12] Dang R, Liu C, Liu M, Chen Q. Channel attention and multi-scale graph neural networks for skeleton-based action recognition. AI Commun. 2022;35(3):187–205.10.3233/AIC-210250Suche in Google Scholar

[13] Sheng Z, Xu Y, Xue S, Li D. Graph-based spatial-temporal convolutional network for vehicle trajectory prediction in autonomous driving. IEEE Trans Intell Transp Syst. 2022;23(10):17654–9.10.1109/TITS.2022.3155749Suche in Google Scholar

[14] Hu N, Zhang D, Liang W, Hsieh MY. Graph learning-based spatial-temporal graph convolutional neural networks for traffic forecasting. Connect Sci. 2022;34(1):429–48.10.1080/09540091.2021.2006607Suche in Google Scholar

[15] Peng L, Wang N, Xu J, Zhu X, Li X. GATE: Graph CCA for temporal self-supervised learning for label-efficient fMRI analysis. IEEE Trans Med Imaging. 2022;42(2):391–402.10.1109/TMI.2022.3201974Suche in Google Scholar PubMed

[16] Rahmani S, Baghbani A, Bouguila N, Patterson Z. Graph neural networks for intelligent transportation systems: A survey. IEEE Trans Intell Transp Syst. 2023;24(8):8846–85.10.1109/TITS.2023.3257759Suche in Google Scholar

[17] Wein S, Schüller A, Tomé AM, Malloni WM, Greenlee EW, Lang MW. Forecasting brain activity based on models of spatiotemporal brain dynamics: A comparison of graph neural network architectures. Netw Neurosci. 2022;6(3):665–701.10.1162/netn_a_00252Suche in Google Scholar PubMed PubMed Central

[18] Pineda J, Midtvedt B, Bachimanchi H, Noé S, Midtvedt D, Volpe G, et al. Geometric deep learning reveals the spatiotemporal features of microscopic motion. Nat Mach Intell. 2023;5(1):71–82.10.1038/s42256-022-00595-0Suche in Google Scholar

[19] Dong G, Tang M, Wang Z, Gao J, Guo S, Cai L, et al. Graph neural networks in IoT: A survey. ACM Trans Sens Netw. 2023;19(2):1–50.10.1145/3565973Suche in Google Scholar

[20] Zhao Y, Du H, Liu Y, Wei S, Chen X, Zhuang F, et al. Stock movement prediction based on bi-typed hybrid-relational market knowledge graph via dual attention networks. IEEE Trans Knowl Data Eng. 2022;35(8):8559–71.10.1109/TKDE.2022.3220520Suche in Google Scholar

[21] Tao M, Gao S, Mao D, Huang H. Knowledge graph and deep learning combined with a stock price prediction network focusing on related stocks and mutation points. J King Saud Univ Comput Inf Sci. 2022;34(7):4322–34.10.1016/j.jksuci.2022.05.014Suche in Google Scholar

[22] Choudhuri S, Adeniye A, Sen A. Distribution alignment using complement entropy objective and adaptive consensus-based label refinement for partial domain adaptation. Artif Intell Appl. 2023;1(1):43–51.10.47852/bonviewAIA2202524Suche in Google Scholar

[23] Liu Y, Zhang H, Li Y, He K, Xu D. Skeleton-based human action recognition via large-kernel attention graph convolutional network. IEEE Trans Vis Comput Graph. 2023;29(5):2575–85.10.1109/TVCG.2023.3247075Suche in Google Scholar PubMed

[24] Bok D, Rakovac M, Foster C. An examination and critique of subjective methods to determine exercise intensity: The talk test, feeling scale, and rating of perceived exertion. Sports Med. 2022;52(9):2085–109.10.1007/s40279-022-01690-3Suche in Google Scholar PubMed

[25] Hansen D, Abreu A, Ambrosetti M, Cornelissen V, Gevaert A, Kemps H, et al. Exercise intensity assessment and prescription in cardiovascular rehabilitation and beyond: why and how: a position statement from the secondary prevention and rehabilitation section of the European association of preventive cardiology. Eur J Prev Cardiol. 2022;29(1):230–45.10.1093/eurjpc/zwab007Suche in Google Scholar PubMed

[26] Hutchinson MJ, Goosey-Tolfrey VL. Rethinking aerobic exercise intensity prescription in adults with spinal cord injury: time to end the use of ‘moderate to vigorous’ intensity? Spinal Cord. 2022;60(6):484–90.10.1038/s41393-021-00733-2Suche in Google Scholar PubMed PubMed Central

[27] Dror N, Carbone J, Haddad F, Falk B, Klentrou P, Radom-Aizik S. Sclerostin and bone turnover markers response to cycling and running at the same moderate-to-vigorous exercise intensity in healthy men. J Endocrinol Invest. 2022;45(2):391–7.10.1007/s40618-021-01659-5Suche in Google Scholar PubMed

[28] Stickland MK, Neder JA, Guenette JA, Donnell DE, Jensen D. Using cardiopulmonary exercise testing to understand dyspnea and exercise intolerance in respiratory disease. Chest. 2022;161(6):1505–16.10.1016/j.chest.2022.01.021Suche in Google Scholar PubMed

[29] DeZoort G, Battaglia PW, Biscarat C, Vlimant JR. Graph neural networks at the Large Hadron Collider. Nat Rev Phys. 2023;5(5):281–303.10.1038/s42254-023-00569-0Suche in Google Scholar

[30] Aspandi D, Sukno F, Schuller BW, Binefa X. Audio-visual gated-sequenced neural networks for affect recognition. IEEE Trans Affect Comput. 2022;14(3):2193–208.10.1109/TAFFC.2022.3156026Suche in Google Scholar

Received: 2024-09-19

Revised: 2024-11-25

Accepted: 2024-11-28

Published Online: 2025-03-05

This work is licensed under the Creative Commons Attribution 4.0 International License.

Artikel in diesem Heft

https://doi.org/10.1515/nleng-2024-0071

Schlagwörter für diesen Artikel

assessment of exercise intensity; knowledge graph method; GGNN model; parallel computing; gated recurrent network

Creative Commons

BY 4.0