Introduction

A person under the age of 18 who violates the law is considered a juvenile delinquent. Conduct disorder is a clinically recurrent pattern of antisocial behavior in which a person frequently violates social rules and participates in deviant behavior that causes others to be upset (Krohn and Lane, 2015). Adolescents during this phase of life develop more advanced patterns of reasoning. Hence, there is a higher risk of the development of such deviant behavior in juveniles. In India, the latest report published by NCRB (National Crime Records Bureau) states that a total of 31,591 crimes were committed by juveniles in the year 2018. The report states that boys make up more than 99 percent of the juveniles arrested for these offenses. This trend is concerning not just because the offense trends represent more serious offenses committed by these children, but also because these young criminals are more likely to continue their criminal activities. As deviant behavior usually develops during adolescence (Howell, 2003), it is very much essential for early identification (screening) of such individuals. Early intervention helps to prevent delinquent activity and promotes the growth of a young person’s assets and resilience. Screening such individuals by a psychologist is quite difficult in many cases since multiple behavioral and environmental dimensions need to be correlated before arriving at a final decision (Siegel and Welsh, 2014). Considering this problem, efforts must be made to identify at-risk adolescents with delinquent behavior early. Conventional diagnosis (Feld and Bishop, 2011) of deviant behavior in adolescents is based on psychological questionnaire measurements and the personal clinical experience of the psychologist.

Figure 1 depicts the steps followed for the conventional diagnosis of juvenile delinquency.

Fig. 1: The procedure followed for the conventional or traditional diagnosis of juvenile delinquency.
figure 1

This type of manual process is dependent upon the experience of the psychologist.

Full size image

The diagnostic evaluation of individuals with suspected delinquency begins with a careful review of the social and behavioral history, through psychiatric assessment or psychological screening. A formal assessment of delinquent behavior (Gearhart and Tucker, 2020) in adolescents is based on semi-structured diagnostic interviews and validated questionnaires (psychometric scales) under the supervision of a psychologist. Moreover, manual analysis becomes questionable not only because of the amount of work but also with regard to precision and the reproducibility of the results. The motivation to automate comes from the fact that besides being time-consuming, the results of manual diagnosis vary with the psychologist’s skill, experience, workload, and stress level. Automated systems (Rathinabalan and Naaraayan, 2017) can help overcome the dearth of trained psychologists. Owing to the existence of such ambiguity in the conventional diagnosis of deviant behavior, the traditional evaluation techniques need to be enhanced by using a quantitative approach in psychology. The principle of quantitative psychology (Ibrahim, 2016) focuses on the use of statistical and machine-learning models for the analysis of psychological patterns of human behavior. Additionally, computer-aided diagnostic systems can support the screening of adolescents in remote and rural parts of the country. In the last few years, various researchers have been attracted to digital psychology and have contributed to the area of modern quantitative psychology. Machine learning is one of the most efficient techniques for predictive data analysis, which is being implemented in multiple research dimensions (Alpaydin, 2021; Greener et al., 2022). Although extensive research has been carried out to understand different risk and protective factors of child delinquency, studies on the machine-learning methods on early risk assessment of delinquent behavior in juveniles are limited. Many schemes are reported to use individual risk factors along with statistical models to assess the significance of individual risk factors in detecting delinquency (Zhang et al., 2014; Svensson et al., 2013; Sciandra et al., 2013). However, most of them fail to provide a generic output. Few schemes have considered delinquency risk assessment as a multivariate pattern classification problem (Lansford et al., 2007; Bor et al., 2001; Britt, 1997). Empirical studies on delinquency showed that machine learning-based model outperforms statistical models in terms of accuracy. Some of the reported methods using single self-report scales claim to be reliable but they are not truly reliable since delinquency is multifaceted (Gray, 1987; Vaux and Ruggiero, 1983; Meldrum et al., 2015). These schemes generally use a single category of the risk factor for the prediction of delinquency in children (Castellana et al., 2014; Jacobsen and Zaatut, 2022).

The accuracy of early detection of delinquency mostly depends on these risk factors and the use of a single category of risk factor may not deliver satisfactory performance in all situations. Thus, to improve the delinquency detection capability a multivariate approach involving different risk factors should be used.

The effect of sibling delinquency on adolescent delinquency on a broader scale with the least effects of other social domains on delinquency was analyzed (Huijsmans et al., 2019). They used hierarchical linear modeling (HLM) techniques using six waves of data for analysis. They suggested that the effect of peers, parents, and school are important factors to be analyzed in future research and interventions targeted to adolescent delinquency analysis. A study in the Philippines using a dataset collected from the City Social Welfare Development Office of Butuan City to develop predictive models to analyze the children at risk as well as children in conflict with the law was conducted (Castro and Hernandez, 2019). They applied algorithms, viz., decision tree, Naive Bayes, generalized linear model, and Logistic regression to the dataset. The Naive Bayes algorithm gave the best classification accuracy and the least classification error. The study showed that a large number of children around the age limit of 12−17 are affected by maltreatment and children in the age group 15−17 committed severe crimes.

Hence, exercising multiple risk factors and advanced machine-learning models in the automated detection of delinquency remains an active field of research to attempt this problem. Keeping the research directions in view, it has been realized that there exists enough scope to improve the delinquency detection performance for deviant individuals in different social and family environmental conditions. To improve delinquency detection capability, a multi-dimensional approach involving various independent risk factors has been carried out in this research.

Materials and methodology

In the present work, an effort has been made to perform delinquency behavioral risk assessment for adolescents using soft voting-based machine-learning ensemble modeling. In this section, we describe the details of the layout of the proposed system as well as the step-by-step approach followed. The block diagram of the proposed computational approach for the early diagnosis of delinquent behavior in adolescents is shown in Fig. 2, which is explained in detail in the following sections.

Fig. 2: The overall step-by-step computational approach for the early diagnosis of delinquent behavior in juveniles.
figure 2

The developed model can classify adolescent behavior into three classes, i.e., low, medium, and high risk.

Full size image

After carrying out the literature review, identifying the research, setting research objectives and research hypothesis following step-by-step methodology has been formulated.

Study area selection

The state of Jharkhand, India was chosen purposively for the study because in the recent past many juveniles have been arrested for heinous activities due to the high poverty rate.

Source of data and survey element

The primary data was collected by interviewing the adolescents studying in Class VII to Class XII of the selected schools. Secondary data of each individual includes his previous social behavior in the school and was obtained by interviewing the associated teachers. A standard delinquency questionnaire (ISRD-3 India Version) (Enzmann et al., 2018; Barranco et al., 2022; Marshall et al., 2019) with required rationalization according to regional conditions was used for the interview.

Participants selection and data collection strategy

A total of 182 students from the Ranchi district of Jharkhand, India with an age variation of 12 to 17 years participated in this study. The required consent to participate in the survey was sought from each school’s head (principal) and subsequently from the students in the classroom. The team of three members consisting of both the authors and a behavioral psychologist carried out data collection in 15 schools (both Government and private) of the Ranchi district of Jharkhand. Permission to survey each school was granted by the head of the school. Consent to participate in the survey was sought subsequently from the students in the classroom. Finally face to face interview with the students was carried out by the two members. A total of 182 students (96 boys and 86 girls) participated in the survey and their response was recorded. The first segment of planning involved an overview and discussion of the questionnaire so that students could understand the nature and objectives of the research and answer any questions that were likely to be asked since the interview was being carried out in the classroom. The second segment of planning focused on the data collection plan and discussed coordinating the team of the survey with teachers in each school, along with the procedures to be followed in the classroom. The response options obtained from the adolescents for each factor are converted into a numeric score based on various pre-defined scales as explained in the following sections.

Behavioral attributes creation and quantification

Suitable representation of behavioral attributes of adolescents is essential for computer-aided diagnosis of delinquency. The questionnaire set based on ISRD-3 is quantified into 53 attributes based on specifications provided by a two-member panel of clinical Psychologists. Each data field of a particular factor/feature measurement has been converted to a proportionate numeric value. Standard scaling methods in psychology (modified Likert Scale) have been used to quantify individual, parental, and environmental factors as per the recommendation of the two-member panel. Adding to the above factors, other features, viz., victimization and gang relationships are also analyzed, measured, and represented numerically for computer-based analysis.

Feature selection

To determine meaningful qualities for the current dataset, a standard statistical feature selection technique called analysis of variance (ANOVA) was applied (Saefi et al., 2020). The ANOVA method analyzes group variances and means to see if they are overlapping. The features are considered statistically significant if they have a smaller p-value. The p-value was set at 0.05 in this study to indicate that the features are statistically significant. Out of the 53 features tested, 47 were found to be statistically significant (p < 0.05).

One-way ANOVA test

The one-way ANOVA is applied to verify whether any significant differences exist between the means of three or more independent (unrelated) groups. The null hypothesis is rejected if any of the group means deviate considerably from the overall mean (Pauly and Smaga, 2020). Table 1 shows the major statistically significant risk factors and predictors analyzed for juvenile delinquency detection.

Table 1 Significant socio-demographic risk factors and predictors for juvenile delinquency detection and classification.
Full size table

Figure 3, Fig. 4, and Fig. 5 illustrate the box plots for some significant features, which were tested and analyzed for juvenile delinquency detection.

Fig. 3: The box plots of two major attributes viz. neighbourhood association and aggression.
figure 3

X-axis indicates the 3 classes and Y-axis indicates the given feature.

Full size image
Fig. 4: The box plots of two major attributes viz. trip without permission and victimization.
figure 4

X-axis indicates the 3 classes and Y-axis indicates the given feature.

Full size image
Fig. 5: The box plots of two major attributes viz. attitude towards delinquency and peer attachment.
figure 5

X-axis indicates the 3 classes and Y-axis indicates the given feature.

Full size image

Reliability test (using Cronbach’s alpha)

The degree to which a test or a measuring technique produces consistent results over time is referred to as reliability (Sobri et al., 2019). Reliability is the trend toward consistency exhibited in repeated measurements of the same phenomena. Exploratory factor analysis is a major technique for validating dimensionality. Cronbach’s Alpha (Taber, 2018) is commonly used to evaluate internal consistency or reliability. Nunnally and Bernstein, on the other hand, generalize an alpha coefficient of 0.70 or above to be generally acceptable (Nunnally, 1975). Cronbach’s alpha is mathematically explained as a function of the test items count and the inter-correlation mean among the features. It is defined as:

$$\alpha = \left[ {n\,C/\left( {V + \left( {n - 1} \right)\,C} \right.} \right]$$
(1)

where n is the number of features, C is the average covariance among the features, and V is the average variance. Table 2 shows the result of applying the above reliability test on all features of our dataset. Table 3 shows the values of Cronbach’s alpha of individual feature items.

Table 2 Reliability statistics for frequentist scale.
Full size table
Table 3 Reliability statistics of frequentist individual item.
Full size table

Feature scaling using Min–Max normalization

Implementation of machine-learning algorithms over measured behavioral attributes/features is unfeasible as the raw data vary widely. Hence scaling of features is essential to normalize the range of independent attributes. Different scaling methods were applied and a comparative study was done to identify the outstanding approach. It is important to normalize the dataset with a wide range of values before classification. In this component, normalization (Jain et al., 2005) of each of the 47 attributes present in the dataset has been carried out using Min–Max normalization within a range of 0 to 1.

Classification of delinquent behavior of juveniles using machine learning

The problem of child behavior characterization is considered here a supervised machine-learning (or classification) problem. Moreover, adolescents are categorized based on the level of risk of showing delinquent behavior, i.e., low, medium, and high risk. Classification is a category of machine learning, which comes under supervised learning (Aggarwal, 2015). It specifies the data element’s class and is appropriate when the output has both finite and discrete values. In multi-class classification, instances are allocated to one among a range of pre-defined classes. This can be achieved by fitting several binary classification models for each class versus all other classes (One-vs.-Rest) or by using a single model for each pair of classes (called One-vs.-One) (Wang et al., 2021). In this work, a multi-class classification model using ensemble learning has been created to characterize adolescent behavior into three classes, i.e., low, medium, and high based on multiple factors. Specific risk and protective indicators of juvenile delinquency, viz., psychological factors and socio-demographic factors have been studied based on the literature review and the ISRD3 questionnaire (India version) has been used along with required modifications to record the student’s responses through interviews.

Decision tree (C4.5)

In this research, the C4.5 technique has been used which is a widely used algorithm in machine learning. There are several parameters in C4.5, i.e., the parameter to test the effectiveness of the Decision tree with pruning (Polat and Guneş 2009; Al-qazzaz et al., 2021). The number of folds (n) determines the amount of data used for pruning and the minimum count of instances per leaf is denoted by l. The confidence Factor is set at 0.25. All the parameters are correctly set and pruning has been applied.

Support vector machines (SVM)

SVM does classification by generating n-dimensions and then by maximizing the margin to obtain the best classification results. SVMs are based on the concept of hyperplane classifiers or linear separability. In general, the SVM classifier (Tharwat and Gabel, 2020) is a binary classifier that can be used to classify data instances. It may be readjusted to handle a greater number of classes (>2) by implementing one vs. rest or one vs. one methodology. SVMs may also be readjusted to generate a non-linear decision boundary. This is accomplished by converting the input from its current dimension to a higher one. The goal is to obtain a non-linear decision boundary since the link between the input space and converted space is non-linear (Xian and Yang, 2021). Kernel functions are functions that are used to transform data into a non-linear format. SVM’s main goal is to find the best (or optimal) separating hyperplane. The optimal hyperplane has the highest distance from the margins and they always give a better generalization rate. SVM’s objective function (Ricci and Perfetti, 2007) can be expressed as:

$$L_P = \left( {1/2} \right)\left\| w \right\|^2 - \mathop{\sum}\limits_{i = 1}^n\alpha _i\left\{ {y_i\left( {w_i^Tx_i + w_0} \right) - 1} \right\}$$
(2)

where w indicates the weight vector and xi refers to the input vector. The transformation function that is used determines the degree of non-linearity of the decision boundary. On data that is non-linearly separable as well as overlapping data, SVMs proves to be extremely robust and efficient. To handle this, a slack variable (Vapnik and Izmailov, 2021), viz., ei is utilized. The SVM optimization is also termed by a factor (‖w ‖2/2) + C (∑i ei)k.

A kernel K(xi, xj) is considered “valid” if there exists feature space “Φ” so

$$K\left( {x_i,x_j} \right) = \Phi \left( {x_i} \right) \ast \Phi \left( {x_j} \right)$$
(3)

where Φ(xi) is the depiction of xi in higher dimensional space.

The polynomial kernel which is a non-linear kernel (Morales et al., 2021) is used for classification in this work and is given by (Xi Xj + 1)h. The complexity parameter is an indication of the extent of avoidance of misclassification for each training example in SVM optimization. The complexity parameter is set at value one.

Multi-layer perceptron (MLP)

It is one of the most efficient systems of neural networks. The efficiency of the multi-layer perceptron arrives (Taud and Mas, 2018) from non-linear activation functions. In maximum cases, sigmoid activation functions are used, which are given by:

$$f\left( s \right) = 1/\left( {1 + e^{ - s}} \right)$$
(4)

The learning process is based on the minimization of errors between network outputs and desired outputs. Hence a backpropagation of error value through a network similar to that which is learned is implemented. The main task is to find the minimum error function e(w) about the connecting weights. In this research, backpropagation MLP has been used. The learning rate for weight updates and the momentum applied to weight updates are selected properly. There are 2 hidden layers with two nodes each in the network.

Ensemble of classifiers

Ensemble-based methods (Kasim, 2021) are techniques that create multiple models and then combine them to produce better model performance in terms of, viz., accuracy, F-score and Kappa value, and other evaluation parameters. Ensemble methods generally produce more accurate solutions than a single model. One of the main issues in combining different classifiers is the framing of appropriate combination rules. Voting is one of the efficient ways of combining the predictions from multiple machine-learning algorithms.

Soft voting

Soft voting (Kumar and Batra, 2021) produces the best outcome by averaging the probabilities calculated by individual techniques. In soft voting (Kieu et al., 2020) the final prediction is the class with the highest class probability averaged over all individual classifiers.

It is mathematically given by:

$$Y^{\prime} = {\mathrm{argmax}}_i\mathop {\sum}\nolimits_j^m {w_j,p_{ij}}$$
(5)

where wj is the weight that can be assigned to the jth classifier.

In this research, soft voting technique has been used to classify juvenile delinquents into three classes, viz., low, medium, and high. For base classifiers, Decision Tree (C4.5), Support Vector Machines (Polynomial Kernel of degree 3 using One Versus One technique) and Multi-Layer Perceptron (2,2), i.e., two hidden layers with two nodes each have been utilized.

Proposed soft voting ensemble model for multi-class classification of juvenile delinquency

The proposed approach is to build a soft voting ensemble with 3 base classifiers, viz., C4.5 (Decision Tree), MLP (2,2), and SVM (Polynomial kernel with degree 3 using One vs. One technique). The voting model (soft voting) model has been developed to classify each data instance of the dataset which is shown in Fig. 6.

Fig. 6: The proposed soft voting ensemble architecture developed for automated assessment of adolescent delinquency.
figure 6

The base classifiers utilized are, C4.5 (Decision Tree), MLP and SVM.

Full size image

All three base learners can effectively handle non-linear data as well as do multi-class (>2 class) classifications. After identifying the base learners, the next step was to create a voting ensemble. The voting ensemble that was created combined the base learners by calculating the average predicted probability of each class label for each base classifier’s prediction. In the average predicted probability technique (soft voting) the final (target) class label is derived from the class label having the highest average probability estimate. This process takes into account the confidence of each voter. Each of the three class probabilities (low, medium, high) is generated by each classifier and the ensemble calculates the average of each class label (overall 3 classifiers) and chooses the class label with maximum average probability as the target class for each instance. The training and testing of the ensemble model will be done using 10-fold cross-validation to find a suitable technique for classifying juvenile delinquency (Marcot and Hanea, 2021).

Hyperparameter tuning

Hyperparameters are learning parameters that are set before the system is trained and have a direct impact on the model’s performance and efficiency (Schratz et al., 2019). For our analysis, we have implemented four machine learning-based algorithms, viz., C4.5 (decision tree), MLP (multi-layer perceptron), SVM (support vector machines), and the proposed soft voting ensemble-based classifier are explained in detail in section “Validation”. Before utilizing these machine learning-based models, the hyperparameters were effectively tuned so that they could predict and analyze more precisely and efficiently than other models (Duan et al., 2003). Hyperparameters and their tuning values for each model were illustrated in Table 4.

Table 4 Tuning hyperparameters for machine-learning framework.
Full size table

Validation

In this work, 10-fold cross-validation has been utilized for evaluating the machine-learning algorithms. K-fold Cross-validation is a resampling technique used to implement machine-learning models on a small data sample (Grimm et al., 2017). The term K denotes the number of groups into which a given data sample should be divided. Compared to alternative methods such as a straightforward train/test split, cross-validation is simple to grasp and typically yields a less-biased assessment of the model’s performance. In K-fold cross-validation, the initial sample is randomly divided into k subsamples of equal size. The remaining k – 1 subsamples are utilized as training data, and one subsample from the total of k subsamples is kept as the validation data for testing the model. Thereafter, the cross-validation procedure is carried out K times and each of the K subsamples is utilized once as the validation set. To create a single estimation, the K findings are finally averaged (Saud et al., 2020). This technique of validation helps to avoid overfitting and bias.

Evaluation metrics

Proposed ensemble models are evaluated based on different standard performance metrics, i.e., Accuracy (Benussi et al., 2021), MCC (Matthews Correlation Coefficient) (Chicco and Jurman, 2020), and Kappa Value (Alizad et al., 2020). Moreover, ROC-area (Bowers and Zhou, 2019), Precision, Recall, and F1 score (DeVries et al., 2021)-based results have also been provided for comparative performance validation.

$${\mathrm{Classification}}\,{\mathrm{Accuracy}} = \left( {{\mathrm{TP}} + {\mathrm{TN}}} \right)/\left( {{\mathrm{TP}} + {\mathrm{TN}} + {\mathrm{FP}} + {\mathrm{FN}}} \right)$$
(6)
$${\mathrm{TPR}} = {\mathrm{TP}}/\left( {{\mathrm{TP}} + {\mathrm{FN}}} \right)$$
(7)
$${\mathrm{FPR}} = {\mathrm{FP}}/\left( {{\mathrm{TN}} + {\mathrm{FP}}} \right)$$
(8)
$${\mathrm{Precision}} = {\mathrm{TP}}/\left( {{\mathrm{TP}} + {\mathrm{FP}}} \right)$$
(9)

Results

Experimental simulation and analysis were conducted using an Intel Core i7-8700 CPU @ 3.20 GHz PC, along with 16 GB RAM on Windows 10 operating system. Python version 3.9.5 with Pandas and Scikit-learn libraries and Weka have been used for simulation and analysis. For statistical analysis, JASP software version 0.16 has been used. An automated ensemble classification system for the early diagnosis of juvenile delinquency has been developed, experiments were performed using the above configuration, and the corresponding results have been provided in this section. The overview of the performance of the proposed model and its comparative analysis with individual classifiers over 10-fold cross-validation is provided in Table 5.

Table 5 Comparative classification performance of proposed soft voting ensemble model in terms of accuracy, Kappa, and MCC over 10-fold CV.
Full size table

The results show that the multi-class classification using the proposed soft voting ensemble model outperforms the performance of individual classifiers with higher accuracy, Kappa value, and MCC.

The classification performance of the proposed ensemble model in terms of TPR, FPR, Precision, and F-score and its comparison with different classifiers are shown in Table 6.

Table 6 Comparative classification performance of proposed soft voting ensemble model in terms of TPR, FPR, Precision, and F-score.
Full size table

The achieved F-score value indicates that the proposed model has successfully been able to classify juvenile delinquency into three classes with substantial accuracy.

ROC analysis of proposed soft voting ensemble model and comparative performance evaluation with different classifiers

A depiction of a true-positive rate (TPR) versus a false-positive rate (FPR) is known as a ROC curve. ROC is a probability curve and AUC (area under ROC-curve) is used to measure the degree of separability of the classes (Espasandin et al., 2021). In our case of multi-class classification, the ROC curve is plotted and analyzed separately for each class label. AUC is also calculated separately for each class label and then finally averaged as shown in Table 7.

Table 7 Class-wise AUC comparison of proposed soft voting ensemble model with standard classifiers for detection of juvenile delinquency.
Full size table

These values of AUC indicate that the proposed soft voting ensemble model is highly effective in differentiating between the class labels and outperforms individual classifiers. Overall multi-class classification using the soft voting ensemble technique proves to be an efficient mode of juvenile delinquency detection at an early stage.

The individual class-wise plots of ROC for the proposed soft voting ensemble model showing the classification performance for classes low, medium, and high are illustrated in Fig. 7, Fig. 8, and Fig. 9, respectively.

Fig. 7
figure 7

The individual class-wise ROC plot for the proposed soft voting ensemble model showing the classification performance for class ‘Low’.

Full size image
Fig. 8
figure 8

The individual class-wise ROC plot for the proposed soft voting ensemble model showing the classification performance for class ‘Med’.

Full size image
Fig. 9
figure 9

The individual class-wise ROC plot for the proposed soft voting ensemble model showing the classification performance for class ‘High’.

Full size image

These values of AUC and analysis of ROC-plots indicate that the proposed soft voting ensemble model is highly effective in differentiating between the class labels and outperforms individual classifiers. Overall multi-class classification using the soft voting ensemble technique proves to be an efficient mode of juvenile delinquency detection at an early stage.

Performance validation of proposed architecture using confusion matrix

A confusion matrix or error matrix is a special tabular structure that permits the visualization of the classifier performance in the case of supervised learning. The instances in the actual class are represented in each row of the matrix, whereas the instances in a predicted class are represented in each column or vice versa. The confusion matrix is an n*n matrix, with n denoting the number of target class labels, which is shown in Fig. 10. For 2 class labels, a 2 * 2 matrix is generated. In our case for three class labels 3 * 3 confusion matrix is generated for each classifier after 10-fold cross-validation, which is explained in Fig. 10.

Fig. 10
figure 10

Generated 3*3 confusion matrix for each classifier and proposed architecture after 10-Fold CV.

Full size image

All the above results and analysis using confusion matrix, ROC-plots and indicate that the proposed soft voting model using 10-fold CV outperformed all individual machine-learning techniques with 87.50% accuracy, 0.82 MCC, AUC of 0.94, Kappa value of 0.81, and F-score of 0.88. C4.5 (Decision Tree) showed less efficiency when used as an individual classification technique but it was observed that the overall performance of the soft voting ensemble model increased when C4.5 was used as the base learner in the ensemble. MLP emerged as the best individual classifier. Overall multi-class classification using soft voting ensemble modeling emerged as the best technique in terms of model performance.

The comparison of classification performance of proposed soft voting ensemble model with classical ensemble frameworks, viz., AdaBoost (Hastie et al., 2009), Random Forest (Speiser et al., 2019), and CatBoost (Jabeur et al., 2021) in terms of Classification performance accuracy is presented in Table 8.

Table 8 Comparison of classification performance of proposed soft voting ensemble model with classical ensemble frameworks.
Full size table

From the comparison of the proposed model’s performance with the above-mentioned classical ensemble techniques, it can be inferred that the proposed soft voting ensemble model outperforms other ensemble techniques in terms of performance accuracy.

Conclusion

Individual factors and family effects are the most important risk factors leading to deviant attributes in adolescents. A single identified risk variable or risk factor cannot be blamed for adolescent delinquency. Specific derived indicators, such as stress levels and aggressiveness are the product of multiple influences across time, ranging from heredity to the child’s surroundings. Automated quantitative assessment of delinquent behavior is essential for the early diagnosis of delinquency among adolescents. Such type of quantitative delinquency risk assessment will facilitate clinicians in the early identification of deviant behavior among adolescents for individual timely therapeutic interventions. In this study, we have proposed a novel soft voting-based ensemble model comprising three base classifiers, viz., Decision Tree (C4.5), Multi-Layer Perceptron, and Support Vector Machine for early detection of delinquency among juveniles using socio-behavioral features of adolescents as input. Encouraging results in terms of multiple evaluation metrics for quantitative grading of delinquent behavior in juveniles were obtained using the proposed ensemble classification model. Furthermore, the established classification model has provision to include additional behavioral attributes to perform advanced identification of complex behavioral disorders among juveniles. In the future, such quantitative behavioral studies will further expedite the process of automated juvenile delinquency detection involving a huge number of clinical parameters using deep-learning-based models.