Construction Safety Risk Model with Construction Accident Network: A Graph Convolutional Network Approach

Abstract

Construction risk assessment (RA) based on expert knowledge and experience incorporates uncertainties that reduce its accuracy and effectiveness in implementing countermeasures. To support the construction of RA procedures and enhance associated decision-making processes, machine learning (ML) approaches have recently been investigated in the literature. Most ML approaches have difficulty processing dependency information from real-life construction datasets. This study developed a novel RA model that incorporates a graph convolutional network (GCN) to account for dependency information between construction accidents. For this purpose, the construction accident dataset was restructured into an accident network, wherein the accidents were connected based on the shared project type. The GCN decodes the construction accident network information to predict each construction activity’s severity outcome, resulting in a prediction accuracy of 94%. Compared with the benchmark feedforward network (FFN) model, the GCN demonstrated a higher prediction accuracy and better generalization ability. The developed GCN severity predictor allows construction professionals to identify high-risk construction accident scenarios while considering dependency based on the shared project type. Ultimately, understanding the relational information between construction accidents increases the representativeness of RA severity predictors, enriches ML models’ comprehension, and results in a more reliable safety model for construction professionals.

1. Introduction

The construction industry is considered a dangerous workplace, where accidents result in casualties and considerable economic losses. In this respect, construction risks are possible and probable events that have the potential to damage people and property. An important requirement for sustainable construction delivery is site safety, ensuring a sustainable environment that protects the safety of construction employees. Despite some improvements in the sector and years of effort put into risk minimization, the situation regarding occupational safety and health (OSH)-related injuries and fatalities remains alarming [1] and is a constant concern [2], displaying high incident rates [3,4,5]. Research has focused on workplace safety, due to the vulnerable position of construction workers [6]. Studies have shown that one third of occupational accidents take place at construction sites [7]. Identifying potential sources of harm within the complex environment of construction sites is the initial stage of OHS risk assessment (RA) to mitigate or prevent construction accidents. RA represents a combined effort to identify events with the potential to cause harm, particularly to people (hazards), and to assess their severity impact. Conventional RA comprises three key stages: risk identification, analysis, and evaluation. In addition, the output of an RA may be responded to by any of the three main risk-treatment strategies: avoidance, transfer, or control. These risk-response strategies may be implemented after a risk event (reactive) or planned to prevent a potential risk event (proactive). In this regard, the objective of reactive risk-response measures is to reduce damage, whereas proactive risk-response strategies prevent the occurrence of a risk event. This has resulted in proactive risk-response strategies being replaced with reactive methods [8]. An effective proactive risk-response strategy requires implementing control strategies through a safety model [9] that formulates accident scenarios ahead of their occurrence. Thus, the representativeness of safety models in regard to real-life scenarios and their ability to offer close-to-accurate insight into possible hazard scenarios and their impacts (severity outcomes) are important indicators of their reliability. These parameters are essential for determining the effectiveness of a model.

As a branch of artificial intelligence (AI), machine learning (ML) models have recently received attention as a way to support the RA procedure [10,11] and predict the outcome of construction accidents [11,12,13,14,15]. However, the adopted construction severity classifiers cannot understand the dependency information among construction accidents, which reduces the comprehensiveness of the adopted ML-based safety predictors. All the models presented in the literature were trained over a tabular dataset with Euclidean distance (Figure 1a), and they presented issues in processing the dependency information that exists between real-life datasets.

Unlike ordered datasets with rows and columns and uniform distances between structured datapoints (such as spreadsheets), graph-structured datasets have a non-Euclidean structure: adjacent datapoints may or may not be categorized under the same features (Figure 1b). This enhances the representativeness of the input dataset into the high-level language (human-friendly language), which cannot be processed using conventional ML approaches. As demonstrated in Figure 1b, graph-structured datasets better represent the complexity of real-life scenarios through networks, storing Euclidean datasets in their vertices (nodes) and connectivity information in their links (edges). Because graph-structured datasets comprise both information within the tabular dataset and dependency information between different datapoints, they can extract more value from the existing information to train ML-based safety predictors. Recently, there has been a growing interest in learning non-Euclidean geometric data. However, the conventional ML approaches presented in the literature cannot process dependency information within graph-structured datasets. This limits the ability of ML to capture the complexity of real-life accident scenarios and consider it within severity prediction procedure.

Graph neural networks (GNNs) are ML approaches that can process the complexity of graph-structured datasets. Graph convolutional networks (GCNs) [16] are a recently developed GNN approach that have received attention owing to their success in various applications [17,18,19,20]. GCNs have proved successful in the construction safety management field for different tasks, such as tunnel-boring performance prediction and hazard classification [21,22]. Despite their ability to obtain richer information, the literature on construction safety still lacks a GNN-based construction severity predictor that accounts for dependency information among construction accidents. Therefore, this study implemented the GCN method and developed a more comprehensive construction safety model to improve the comprehensiveness of adopted safety models and associated RA procedures. Our model predicts the severity outcome of construction accidents by considering the dependency information between the accident datapoints. For this purpose, real-life construction accidents were first configured into a construction accident network, connecting construction accidents based on their shared project types.

In conducting this study to explore the prediction ability of a multilayer GCN algorithm that learns the connectivity between site accidents and project types, we had two primary motivations. First, we expected that obtaining richer information from existing construction safety datasets within the context of accidents connected according to their project type would offer an improved learning experience for the adopted predictive model. Second, we aimed to develop a construction accident severity prediction model that accounted for dependency information and could achieve a more accurate severity prediction. The novelty of this research is thus two-fold: (i) it provides a novel depiction of construction accidents through a graph-structured network (where the nodes are accidents connected according to project type), and (ii) it represents the first application of the multilayer GCN method for the severity prediction of construction accidents (accounting for dependency information between accident neighborhoods within a complex construction accident network). In order to demonstrate the effectiveness of the network representation of construction accidents in improving the learning ability of theML model, we developed a benchmark feedforward network (FFN) algorithm with roughly the same parameters as the GCN algorithm for the prediction of the severity outcomes of construction accidents.

2. Backgrounds

2.1. Machine Learning (ML) Methods as Construction Safety Models

Over the past decade, different ML approaches have been widely used, owing to their fast and accurate prediction ability [11,12,13,14,15,23]. Despite extensive research within other disciplines, the literature on ML in the field of construction is still scarce [11]. Artificial neural networks (ANNs) [12,24,25,26,27,28,29,30] and support vector machines (SVMs) [31,32,33] are the two most dominant ML approaches used in construction safety context. To a lesser extent, decision trees [32,34] and k-nearest neighbors (KNNs) [35] have also been explored as construction safety models. Recently, advanced ML approaches using deep neural network (DNN) layers have been explored. For example, a long short-term memory (LSTM) network was used to prevent gas-related accidents during tunnel excavation [36]. The procedures of previously adopted ML-based risk predictors mostly followed a similar pattern: the collected tabular dataset is pre-processed and split into training and test sets, and then the ML model is trained over the training set and evaluated over the test set. In addition, to increase the accuracy of individual ML-based safety predictor models, the bagging [37,38] and boosting [37,39] of ensemble ML models and the combination of several ML approaches have been investigated.

Moreover, hybrid models have been investigated in the literature to decrease the required computational effort and enhance the prediction accuracy of ML-based construction safety models by integrating ML with different optimization approaches, including engineering and dimensionality-reduction methods. An ANN-based model comprising latent class clustering analysis (LCCA) was adopted by Ayhan and Tokdemir [30]. Koc et al. [40] integrated discrete wavelet transform (DWT) with ANN, multivariate adaptive regression splines (MARS), and SVM to predict the number of daily occupational accidents in construction projects. Recently, convolutional neural network (CNN) [41] models have achieved success as predictive models for structured tabular data in applications such as traffic flow prediction [42]. The existing predictive ML models have focused on factoring in various aspects of the construction safety models that have been developed, such as the data collection method [11] and its balancing [43], the project type [12], and the correlation between accident attributes and their severity [44]. Toğan et al. [15] recently utilized a wide range of single (KNN, SVM, logistic regression, and Naïve Bayes) and ensemble (random forest, gradient boosting, and AdaBoosting) ML approaches within a customized automated ML model to predict the outcome of construction accidents. Their study suggested that the prediction ability of the adopted predictor depends on the input dataset and classification task. However, existing ML-based construction safety predictors cannot capture the connectivity among datasets embedded in the high-level language of a network. The reliability of ML depends on the selection of the ML approach and reliability of the input dataset. Otherwise, superficial knowledge can be learned from the input dataset, resulting in an unreliable outcome.

The literature on safety does not consider connectivity between construction accidents. Construction accident severity prediction using conventional ML approaches cannot benefit from the dependency information contained in real-life construction safety data. It is widely accepted that construction accidents have multiple causes [45,46] related to the specifics of the job, workers, and management. Linking construction accidents based on all their feature attributes and their project type, involves the creation of a complex network of accidents to form a comprehensive safety scenario. Achieving a more comprehensive safety scenario that contains richer information from a given construction accident dataset is expected to increase the model’s representativeness in relation to real-life complexity. Developing an ML-based safety model that better represents real-life construction accident scenarios would result in an intelligence closer to that of humans. The resulting network, compromising connectivity information, would better represent the nature of construction accidents and would thus be relatively reliable for predicting construction risks. However, dependency information in real-life datasets cannot be processed using existing ML-based construction safety models, such as ANN, LSTM, KNN, and SVM.

2.2. Graph Convolutional Network (GCN)

Feeding connectivity information to ML models facilitates learning by incorporating more representative datasets. One method to better represent accident records is to link them to different project types. Tunnel construction, for example, is far riskier than the construction of leisure parks or warehouses. In addition, the structural relationship between site accidents and project type must be illustrated and loaded into the machine using graph representation learning [47]. The utilization of relation-aware representation (RAR) for accident records should increase the reliability of construction accident prediction outcome. A graph convolutional network (GCN) is a RAR method that combines CNN and GNN models by utilizing linear activations among the convolutional layers. The convolutional mechanism is the filtering procedure for loading a matrix from the dataset into the adopted ML model. The GCN procedure is illustrated in Figure 2.

GCNs have proved successful in a variety of applications since 2017, including social networks [18], social recommendations [48], paper citations [17,19], and chemistry [19]. Gao et al. [49] utilized a GCN as an unsupervised feature-extraction technique to obtain a topological gas pipeline network that was distinct from its input features. GCNs have recently been utilized for various applications [17,18,19,20] within the engineering [50] and construction management domains. Pan et al. [22] predicted the performance of tunnel-boring machines by adopting a GCN model. Their study demonstrated the superior performance of GCNs over conventional ML approaches such as deep neural networks (DNNs), SVM, and RF. The rapid expansion of RAR and graph learning topics has encouraged the application of this approach in the field of construction management, particularly construction safety. To this end, recently, Tian et al. [21] adopted a GCN incorporating bidirectional long short-term memory (BiLSTM) to classify construction safety texts. Their study highlighted the ability of GCNs to encode syntactic and semantic feature attributes from a textural context. Motivated by the success of the GCN predictors and their ability to understand the dependency information within construction safety datasets, we develop a novel GCN accident severity classifier and a construction accident network that extracts richer information from existing datasets, aiming to increase the comprehensiveness of construction risk prediction.

3. Methodology

In this study, we adopted a GCN to classify the severity outcomes of construction accidents and increase the comprehensiveness and representativeness of theconstruction safety model, relying on its ability to process the dependency information within construction safety datasets. The developed construction accident severity predictor is shown in Figure 3.

As shown in Figure 3, the objective of this study was achieved through the following steps: first, a real-life construction safety dataset was processed; second, the conventional tabular dataset was reconfigured into a construction accident network; third, the resulting construction accident network was processed using the GCN model to predict the severity outcomes of construction accidents. A feedforward network (FFN) model was configured as a benchmark for the performance of the developed GCN approach. The developed predictor was coded using Python language in the Google Colab environment while adopting the Keras example library [51] on a paper citation classification project, an approach inspired by Kipf and Welling [16].

3.1. Data Description and Preprocessing

This study utilized 5224 construction site accident datasets [12] collected between 2014 and 2017 from 73 construction projects within the Euro–Asia region. The names of the construction projects were sscribbled to protect the privacy of the participating construction companies. Based on their types, each project was assigned to one of the following project categories: assembly, residential, institutional, commercial, industrial, education, healthcare, oil, gas, or park. This project categorization was carried out by considering the intrinsic safety environment of each construction project. For example, plants and refinery projects were placed under the category of industrial projects; offices and shopping malls under commercial projects; and the more hazardous oil and gas projects were grouped separately. The obtained accident dataset, containing 12 accident features and 94 attributes and their frequencies, is presented in

Table 1. Details of the accident dataset.

Variable	Feature	Attribute	Frequency
PC1	Accident ID	Numerical	5224
PC2	Project types	Industrial buildings	1709
		Commercial building projects	1265
		Residential building projects	756
		Educational building projects	620
		Oil and gas projects	284
		Institutional building projects	220
		Healthcare building projects	197
		Assembly building projects	163
		Park project	10
PC3	Accident severity	First aid	3070
		Medical intervention	1393
		Workday loss	397
		Near miss	244
		Material damage	116
		Fatality	4
PC4	Age	25–35	2193
		18–25	1583
		35–45	970
		45–65	478
PC5	Time of day	PM	2789
		AM	2435
PC6		0–1 month	1028
		1–3 months	1832
		3–6 months	1255
		6–12 months	882
		12–24 months	227
PC7	Human factors	Problems resulting from poor management system	1493
		Problems resulting from unbalanced workload	1187
		Insufficient skill and perception	1096
		Physical disability	441
		Faulty management system	357
		Problems related to education level	285
		Others	275
		Psychological disability	90
PC8	Risky behaviors	Inability to perceive external risks	3190
		Violation of safe work policy	505
		Incorrect physical movement	438
		Incorrect/absence of safe work policy	425
		Tending to use a shortcut	329
		Incorrect usage of equipment and handtools	214
		Others	119
		Working on energized equipment	4
PC9	Hazardous cases	Ncrs in the working environment	2652
		Ncrs in safety protection measures	818
		Others	754
		Mechanical hazards/Ncrs	382
		Natural hazards	176
		Radiation exposure	51
		Chemical hazards	26
		Fire or explosion	18
		NCRs in the usage of handtools/equipment/construction equipment	347
PC10	Occupation	Rough work crew	2140
		Mechanical assembly crew	1419
		Finishing work crew	649
		Repairman	412
		Others	221
		Engineer	140
		Administrative affairs	102
		Construction equipment operator	41
PC11	Workplace factors	Method of statement problems	1037
		Inadequate incident analysis systems	959
		Inadequate communication	873
		Inadequate management system	691
		Inadequate maintenance mechanism	485
		Insufficient control or tracking	445
		Others	365
		Lack of OHS training	202
		Inaccurate protection measures	107
		Inaccurate recruitment procedures	60
PC12	Activity	Daily activities	906
		Re-bar/formwork installation	805
		Assembly works	787
		Usage of handtools/equipment	707
		Lifting operations	470
		Welding/hot works	456
		Working with chemicals	348
		Finishing works	341
		Concreting	92
		Repair/maintenance works	59
		Excavation works	24
		Working at height	24
		Field measurement works	21
		Testing works	14
		Working with chemical materials	12
		Mobilization on/off site	9
		Landscaping works	7
		Cable-pipe assembly/working with containments	6
		Geotechnical works	2
		Material drop	1
		Transportation/construction equipment/usage of vehicle	133

.

The severity outcome of each accident (PC3) was first divided into six categories and then further categorized into two classes: high-severity (HS) and low-severity (LS). In this respect, near-miss and material damage outcomes were categorized as LS, whereas accidents that resulted in a fatality, medical/first-aid intervention, or workday loss were categorized as HS, similar to [30]. The outcome severity feature contained 4864 high-risk and 360 low-risk attributes.

3.2. Construction Accident Network

We first pre-processed the construction accident dataset into a construction accident network (Table 1) for later processing using the GCN severity outcome predictor. For this purpose, construction accident features, excluding accident IDs (PC1) and severity outcomes (PC3), were first dummy-encoded. This resulted in a dataset with 5224 rows and 75 binary columns, along with the accident ID and severity columns (Figure 4a).

Additionally, the connectivity information was placed in the second dataset (the connectivity dataset), demonstrating the dependency between accident IDs (Figure 4a). In the connectivity dataset, accident IDs were linked based on shared project types (PC2). The configuration of the sample construction accident node is illustrated in Figure 5.

The nodal and accident connectivity datasets were used as the input for the GCN severity predictor, providing the structure of the construction accident networks. In this study, a construction accident network comprised 5224 accident nodes and 2,833,076 connectivity edges. Loading and processing the information within GNN-based models is a challenging task that is addressed through different approaches. In this study, the graph dimension was adjusted to 5224, allowing the model to accept the training input size as well as the test and evaluation sets. In other words, the construction of the graph model in the Keras library was utilized for structuring the properties of the prediction model; thus, it was not used as the input dataset for training the model. Figure 6 shows the configured accident network.

Note that due to the large number of nodes and edges, the depiction of the whole accident network (Figure 6c) could not illustrate the details of the nodal arrangement, and thus samples of single nodes (Figure 6a) and groups of connected accident nodes (Figure 6b) are also depicted. In addition, as shown in Figure 6c, the nine project type (PC2) attributes created nine independent subnetworks within the main construction accident network. We did not add weights to the connectivity information; thus, all the nodes in a neighborhood were separated by the same distance.

3.3. Configuring the GCN Model

The GCN predictor was configured using three main procedures: message preparation, aggregation, and updating. Figure 7 outlines the main steps of the GCN approach for processing the received construction accident network, along with severity outcome prediction.

Message preparation was performed using a feedforward (FFN) model by processing node information using a linear activation function (Figure 8).

The initial node representation was prepared using an FFN block comprising a batch normalization layer, dropout layer, and dense layer. Generally, in GCN models, nonlinear activation functions are used for processing the learned stacked layers [20]. In this study, we used the Gaussian error linear unit (GELU) and therefore collapsed the weight matrices between consecutive layers [20]. The message prepared by the first FFN block was processed with two GCN layers, comprising message aggregation and updating (Figure 9).

As detailed in Figure 9, upon receiving and stacking all the information from the connected accident nodes, the received node information was aggregated. Here, the summation method was used to aggregate the messages from the adjacent nodes in the neighborhood of each construction accident. The generated node representation was sent to the graph convolutional layers for convolutional operations. Unlike structured grids of image pixels, graph representation involves unstructured grids, so FFN blocks cannot be directly used. However, GCNs generalize the stacking operation of convolution layers from a structured dataset to a graph-structured dataset [52], which allowed us to extract high-level construction accident nodes and translate them into the language of the adopted model. Thus, the GCN algorithm received the information from all the accident nodes in the neighborhood of a single accident node and then stacked all the received information in that node (Figure 9). The developed GCN classifier comprised two graph convolutional layers, where each layer concatenated the mean information received from the neighborhood.

Furthermore, the FFN blocks were also utilized within the post-processing stages of the GCN algorithms; therefore, adjusting their parameters directly influenced the performance of the GCN algorithm (Figure 7). Because the node convolution operation involved partial information loss, four skip connections were used to send information outside the convolutional layers. The outcome of the convolutional layers and skip connections allowed the model to learn the embeddings of the construction accident records. The final embedding was post-processed using a Softmax layer to determine the severity level of the construction accidents.

3.4. Model Training

The input parameters used for configuring the FFN and GCN layers are listed in

Table 2. Hyperparameters used in FFN and GCN algorithms.

Hyperparameter	FFN and GCN Algorithms
Training set	2556
Test set	2268
Validation set	400
Hidden units	32 × 32
No. of epochs	300
Batch size	256
Dropout rate	0.5
Optimizer	Adam
Learning rate	0.01
Loss	Sparse categorical cross entropy
Activation (excluding the final layer)	GELU
Activation (final layer)	Softmax
Patience	50

.

In this study, half of the accident records associated with each attribute of the severity features were selected for training the models. The other half were allocated to model testing and validation, i.e., 85% were separated to test the model on the unseen dataset, and 15% were used to validate the model prediction during training and enhance its generalization ability. Therefore, when the model received the 2556 training examples for the two-level classification task, it created an output with a shape of 2556 × 2. In addition, a random generator from the NumPy library was used to select the index values of the accident records. This resulted in statistically different training, test, and validation datasets. Regarding the activation function, linear activation functions in both the FFN and GCN algorithms were employed to reduce the model complexity. This procedure could be performed using different numbers of GCN layers. The addition of several convolution layers increases the prediction accuracy but also smoothens the model by obtaining similar embeddings for all accident nodes [51]. Therefore, a trade-off must be achieved between the accuracy obtained by the number of GCN layers and the generalization ability of the model, considering the size and dimensions of the utilized dataset. In this study, a GCN algorithm with a dense input and output layer containing two graph convolution layers was built over the FFN blocks. The final GCN model was created using the parameters listed in Table 2 and the hyperparameters described in

Table 3. Details of benchmark FFN and GCN layers.

Model Description	FFN Model	Multilayer GCN Model
Total trainable parameters	12,728	17,144
Total non-trainable parameters	726	982
Total parameters	13,454	18,126

.

To compare the developed GCN severity predictor with conventional ML approaches, an ANN-based severity predictor was created using FFN, adopting the procedure outlined in Figure 8 and with roughly the same parameters as the GCN algorithm (Table 3).

3.5. Model Evaluation

The learning performance of both the FFN and GCN algorithms was measured in terms of accuracy, generalization ability, and computational cost. The FFN and GCN accuracies are shown in Figure 10.

Both benchmark FFN and GCN algorithms displayed high prediction accuracies of 94.2% and 94.47%, respectively, after 116 and 186 epochs, respectively. The difference between the training and test accuracies and loss plots determined the generalization ability of both models. In this regard, as Figure 10 shows, there was a 2% difference between the training and test results in the FFN algorithm, whereas the GCN algorithm showed good generalization ability in earlier epochs and had improved by the last epochs. While the accuracy reflected the number of correct predictions, the loss metric indicated the number of incorrect severity predictions between the FFN and GCN models (Figure 11).

The GCN algorithm outperformed the benchmark FFN algorithm in terms of prediction accuracy and generalization ability. The FFN and GCN algorithms were each evaluated 10 times to assess their robustness, with different training, test, and validation sets were selected at random for each evaluation (Table 3).

As shown in

Table 4. FFN and GCN algorithm accuracy across 10 evaluations.

Evaluation	FFN Algorithm		GCN Algorithm
	Epochs	Accuracy (%)	Epochs	Accuracy (%)
1	123	93.73	160	93.80
2	51	93.43	51	93.43
3	136	93.68	104	94.13
4	90	94.35	83	93.97
5	125	93.26	129	94.15
6	110	93.94	81	92.79
7	113	94.51	51	94.33
8	103	94.17	159	94.33
9	129	93.44	129	93.89
10	111	94.45	150	94.73

, the FFN and GCN models demonstrated 93.90% and 93.96% average prediction accuracies, respectively. While the main objective of this study was to increase the quality of construction safety models in terms of enhancing their representation ability in terms of real-life construction accident datasets, the results demonstrated a slight improvement in prediction accuracy and better generalization ability compared with FFN. However, the FFN algorithm cannot learn dependency information, which restricts its ability to understand the context of construction safety datasets to the same extent as the GCN model. The context of the input dataset is of vital importance to the reliability of ML approaches, because they are prone to superficial learning by understanding the pattern even within irrelevant dataset, which reduces the reliability of predictor. The GCN predictor adopted in this study achieved a good prediction accuracy as a deep learning approach by processing a greater amount of complex and challenging contextual information from the construction accident dataset.

Additionally, as shown in

Table 5. CPU and wall time values of benchmark FFN and GCN algorithms.

Hyperparameter	Benchmark FFN Algorithm (s)	Multilayer GCN Algorithm (s)
User CPU time	30.2	422
System CPU time	2.3	6.57
Total CPU time	32.5	448
Total wall time	30.3	802

, the computational costs of the FFN and GCN algorithms were evaluated in terms of their CPU usage and actual wall time.

In terms of the computational cost, the benchmark FFN algorithm was computationally more efficient than the GCN algorithm. This was due to the loading of the two graph convolutional layers with the entire graph dataset. The proposed GCN algorithm must be fitted entirely within the system, which results in CPU overloading when using big data.

4. Discussion and Recommendations

An important principle for maintaining sustainability within construction organizations is to protect the safety of the employees, especially site workers. Maintaining site safety during construction is a complex managerial challenge that demands the holistic supervision of all site jobs, processes, teams, and workers. Therefore, identifying the cause of construction accidents is a complicated task involving many factors. In this study, the first challenge was to connect the accident nodes within the network and load the connectivity information from high-level to low-level language that the machine could understand. Training an ML model over a non-Euclidean data structure is challenging [51] because it requires encoding both nodal and connectivity datasets into the machine’s language. High-level information, such as networks, is user-oriented and closer to the understanding of human intelligence, and is thus not machine-oriented. Considering this, the developed GCN could process complex non-Euclidean structure datasets [53] through an encoding and decoding procedure that translated the high-level relational information within the construction accident network into the low-level language of machine. Here, the GCN functioned as a special case of a CNN [54] that generalized its convolution mechanism [52] to receive the decoded network information, generalizing the convoluted layers. The encoding and decoding mechanisms allowed both nodal and connectivity information to be received. In contrast, the convolutional mechanism benefits from the high prediction accuracy of strong ML predictors such as CNNs.

It should also be noted that the results confirmed the model’s suitability for a dataset of less than 10k; for a larger dataset, the number of convolutional layers would have to be increased. However, it is important to maintain a balance between the accuracy obtained by increasing the number of graph convolutional layers and the smoothness of the model (the overfitting issue). In addition, the GCN algorithm using the accident network developed in this study provided high prediction accuracy without overfitting. The high accuracy achieved by this model encourages the development of more complicated accident networks in which the relationships among other accident features are also considered. The effective utilization of the method developed herein would require the creation of meaningful links between construction accidents rather than an exclusive focus on model accuracy. Compared with the benchmark FFN algorithm, the GCN algorithm achieved improved accuracy and generalization ability. This may have been because of the richer information obtained from the existing accident dataset. Thus, applying the method developed in this study on the other datasets, used in existing ML-based safety models, is expected to achieve a better performance.

The main objective of this study was to develop a construction severity predictor that factored in the dependency information among accidents at different construction sites during different projects. This enabled the multilayer GCN algorithm to better learn the embedded features within the dataset by receiving accident-to-project connectivity information. Despite the configured construction accident network, implementing the developed GCN severity predictor was straightforward due to the automation of node selection, encoding, embedding representation, and graph formation up to severity classification, which made the model practical and effective. Depicting connectivity between worker age and experience, or between hazardous activity and time of day, for example, should allow a richer learning process.

In this study, we developed a novel GNN-based construction safety model, which remains to be explored further from practical and technical perspectives. The results of this study showed that the GCN demonstrated a better generation ability and slightly improved prediction accuracy compared with FFN while processing unstructured (graph-structured) construction accident datasets. From a practical perspective, the model could be further explored using different connectivity information that better represents the reality of construction accident records. In addition, the selection of safety features as nodal attributes could be explored using different feature engineering approaches. Below, we highlight the SWOT analysis of the developed GCN construction severity predictor from a technical perspective.

Strengths: ML allows the solution of practical problems through learning from past information. In this regard, GCNs’ ability to understand the complex dependency information that exists in real-life datasets (such as construction safety datasets) comes at the price of difficult regularizing the complex non-Euclidean datasets. Our results suggested that ML-based construction safety models can move beyond the limits of structured datasets and understand a greater extent from the contextual dependency information within complex real-life safety datasets. Although different ML-based predictors have been developed in the construction safety literature, the quality of the safety model depends on the quality of the input safety dataset. The ability to incorporate the dependency information between construction accidents along with the accident dataset enhances the quality of the input safety dataset, which, in turn, enhances the quality of the resulting safety model.

Weaknesses: It should also be noted that the developed safety model was tested over a randomly selected subset of the collected accident dataset, and that the remained data was used for training the model. Although the test and training sets were separated, they were still obtained from the same dataset and had a consistent data collection system. Thus, testing the developed GCN severity predictor over an accident dataset collected according to different features and attributes might demonstrate the detachment of the predictor model from the reality of the input dataset, using a shortcut learning approach. However, the abovementioned knowledge issue related to the ML-based predictor is not specific to the developed GCN model and is still an ongoing issue for other DNN approaches, such as FFN.

Opportunities: This study contributes to existing ML-based approaches in the construction safety literature through its data-centric approach. The literature has proven the abilities of different ML approaches for various prediction tasks using different datasets. The developed GCN construction severity predictor adopted a DNN predictor that extracted richer contextual information from construction safety records. As this study considered only one type of dependency information, the topic remains to be investigated in regard to other types of meaningful connectivity information. Likewise, the arrangement of information on accident nodes should be explored further through the creation of other construction safety networks, e.g., by connecting the root causes of construction accidents.

Threats: As discussed earlier, the entire construction safety network was fed into the GCN model, using entire dataset for constructing the predictor and training it over the 2556-item training set. This increased the computational cost of the GCN model; thus, the model could be further improved through approaches such as random walk.

Finally, the GCN model trained using the past accident dataset was able to classify the severity outcomes of new construction accident records. This would allow construction professionals to flag severe accidents by evaluating different scenarios for a specific construction project using the GCN model. For example, the user defines the prediction input by detailing accident scenarios (e.g., worker age and experience) associated with a housing project. The GCN model predicts the severity outcome of the entered construction accident scenario while considering the dependency of the input dataset on previous accidents within housing projects. Perceiving accident information within project types allows decision-makers to consider project-specific risks. The type of project routinely determines the kind of construction activities and the associated hazards, managerial structure, and health and safety culture of the project team. Therefore, linking construction accidents through their shared project attributes would allow the mentioned items to be factored in as existing accident attributes. The improved representativeness and use of input accident data by the proposed GCN model provides an enhanced risk prediction model for construction sites.

5. Conclusions

Through a graph-structured accident network, in this study we proposed an accident severity prediction model that benefits from connectivity and embedded safety information to extract richer information from existing safety datasets, resulting in more accurate severity prediction of construction accidents. The conclusions were as follows:

(1)

This study created a construction accident network by establishing connectivity among construction accidents based on shared project types. The accident network consisted of nine project-type edges, which formed nine separate subgraphs within the main accident graph. Ten safety features with 98 attributes were added to each accident node to define the characteristics of each accident.

(2)

High-level graph information was represented by mapping the construction accident in terms of its project type, which resulted in a safety model that better represented the reality of construction accidents in the context of different project types. The enhanced representation of the accident dataset as the model input provided a real-world dataset for the risk-prediction model.

(3)

The proposed severity prediction method has other desirable properties, including good generalization ability and high prediction accuracy for different dataset sizes, achieved by varying the number of graph convolutions.

(4)

The GCN algorithm displayed a better prediction accuracy and generalization ability than the benchmark FFN algorithm. Compared to the structured safety features loaded into the FFN algorithm, the graph-structured dataset containing embedded connectivity information improved the prediction ability of the GCN algorithm. By comparing an enhanced accident input and a newly developed GCN algorithm, the proposed risk-prediction model offers risk prediction with better visibility for construction professionals.

(5)

The computational cost of the GCN algorithm was higher than that of the benchmark FFN algorithm. This was due to the processing of the complete graph of the GCN model, which increased the loading on the available RAM. However, adopting optimization methods or sampling techniques, such as taking an unbiased random walk inside the graph space, would reduce the computational cost of the GCN algorithm.

(6)

The developed construction safety model could be enhanced to predict construction risks based on project type, along with other input attributes, such as the age and experience of the worker, the activity performed, and workplace and human factors. This would allow construction professionals to evaluate different accident scenarios and flag high-risk accident scenarios while considering dependency information based on the construction project type. This is expected to enhance the information gain of ML-based severity predictors and, in turn, improve associated RA procedures.