《Mathematics》：Empirical Review of Traditional and Recent Balancing Techniques for Image-Based Physical Violence Classification Across Diverse Imbalance Scenarios and Multiple Datasets
研究人员针对图像中身体暴力的自动化分类面临一个关键的方法学障碍：类别不平衡（Class Imbalance），该问题会削弱深度学习（DL, Deep Learning）模型的判别能力，导致模型对多数类产生过拟合。尽管文献中已提出众多平衡策略，范围涵盖传统重采样到先进的生成模型，但针对身体暴力检测这一特定领域，在受控不平衡条件下对这些策略行为进行系统的经验评估仍然缺失。在本研究中，研究人员呈现了一项系统性实证综述，评估了在受控不平衡条件下的传统平衡技术、生成模型以及数据增强方法。该方法应用于五个数据集（UBI-FIGHTS、RLVSD、RWF-2000、RLVS和AIRTLAB），跨越由平衡百分比（Balance Percentage, BP）水平1%、5%、10%、25%、50%和75%定义的六种不平衡场景。为了将平衡策略的效应与架构混杂因素分离，研究人员在所有实验条件下采用单一稳定的骨干网络（Backbone）ResNet-18作为受控分类器。实验结果表明，不存在普遍最优的平衡技术：在高度不平衡场景中，传统过采样方法如ROS（Random Over Sampling）获得了最佳平均性能，而深度生成方法如DCGAN（Deep Convolutional Generative Adversarial Network）和DeepSMOTE在类别平衡度改善时竞争力逐渐增强。这些发现证实了平衡技术的有效性依赖于不对称程度和数据集内在特性两方面因素。因此，显然需要基于情境的策略选择而非一刀切（One-Size-Fits-All）的解决方案。除实证发现外，本工作为不平衡暴力检测的理论基础与最新方法提供了结构化综述。
1. Introduction
身体暴力（Physical Violence）是一个令人警醒的全球性问题，对个人和社会造成灾难性后果。在美洲地区，泛美卫生组织报告的凶杀率是全球平均水平的三倍，这凸显了开发有效检测和预防工具的紧迫性。在此背景下，深度学习（DL）的进展为应对这一挑战提供了重要潜力，使其能够自动化分析大规模视觉数据。然而，使用人工神经网络进行图像分类面临一个关键障碍：类别不平衡（Class Imbalance）。该现象指不同类别之间的基数差异，即每个类别可用样本数量的显著差异。
除已有文献所述之外，这种非对称分布会损害模型的预测能力，其性能关键依赖于输入的代表性质量。当输入包含噪声或未能充分捕捉每个类别的判别特征时，系统倾向于产生系统性偏向多数类的预测，显著降低其识别少数类实例的能力——而这些实例通常代表最具分析价值的案例。因此，传统分类算法倾向于以牺牲欠代表类别敏感性为代价来优化全局指标。
近期研究提出了多种技术来缓解这一问题，包括传统过采样和欠采样方法、合成数据生成以及度量学习。然而，在身体暴力检测这一特定领域，针对不同程度不平衡场景对这些策略进行全面评估仍有必要。本研究的主要目标是评估在神经网络分类图像身体暴力时，针对不同平衡场景使用平衡技术处理类别不平衡的效果。
研究所用的基本度量指标之一是平衡百分比（Balance Percentage, BP），定义为最少代表类别样本数相对于所有类别总观测数的比例。研究人员将标注数据集划分为暴力图像和非暴力图像两类，针对每个BP值p，调整后的少数类子集基数满足特定公式，从而产生从极端（1%）到近平衡条件（75-100%）的参数化不平衡场景族。随后，在训练分类器之前应用平衡技术以减少类别分布差异，进而评估模型在不同不平衡水平下的性能。
本研究被构思为一项系统性实证综述，具有三项主要贡献：一是提供关于文献如何处理基于图像的身体暴力检测中类别不平衡的结构化综述；二是形式化支撑所考虑平衡策略、评估指标和生成模型的理论基础；三是开展受控实证评估以分离平衡技术本身的效应，涵盖传统采样方法和用于合成数据生成的高级生成模型（DeepSMOTE和DCGAN），分析其缓解固有不平衡效应和丰富模型内部表征的能力。为此，研究人员刻意采用单一成熟的骨干网络（ResNet-18）作为受控变量，确保观察到的性能差异可归因于平衡策略而非混杂的架构因素。
2. Related Works
本节系统综述了近期关于深度学习图像分类中类别不平衡问题的研究，特别关注应用于固有不平衡数据集的方法及其在提升分类性能方面的有效性。研究对比了各文献的模型架构、平衡策略、数据集和性能结果，为将所提出的方法论置于当前技术背景下提供背景支撑。
Afzal等人的研究采用AlexNet架构与基于几何变换的数据增强技术，应用于OASIS（MRI）数据集，模型准确率从85.11%提升至98.41%。Kumar等人提出SMOTE与ENN混合策略，应用于UBI-FIGHTS和RLVSD数据集，F1分数从0.82提升至0.91。Dablain等人引入DeepSMOTE技术，整合自编码器与合成数据生成，应用于CIFAR-10-LT数据集，较SMOTE等传统技术提升了F1分数。Vosta和Yow提出基于ResNet-50和ConvLSTM的混合CNN-RNN架构，采用随机过采样处理UCF-Crime数据集，AUC达81.70%。Abundez等人基于主动学习和知识迁移，使用DenseNet121、EfficientNetB0和MobileNetV2在AIRTLAB、RVLS和SCVD数据集上报告AUC高达0.989。Ye等人提出IFW-SMOTE新型过采样技术，应用于CIFAR-10数据集，准确率93.10%。Gao等人开发CL-SR技术，整合对比学习与SMOTE-RST，在ImageNet-LT数据集上获得Top-1准确率68.70%。Wang等人提出基于EfficientNet-B0的轻量模型，针对MovieFight和RWF-2000数据集进行暴力检测，准确率90.25%。Azzakhnini等人提出LAVID系统，基于DSCNN-BiLSTM与迁移学习，应用于RWF-2000、RLVSD和Hockey Fights数据集，分别获得91.06%和92.73%的准确率。Chen等人采用HRUMamba模型配合代价敏感方法处理高分辨率土地覆盖分类中的不平衡，Macro-F1分数达92.25%。Yuan和Li提出GAN辅助的YOLOv5n，用于检测苹果花爆发，mAP达67%。
3. Theoretical Foundations
3.1. Deep Learning
深度学习是机器学习的一个子领域，依赖于具有多个隐藏层的人工神经网络。这些网络能够学习数据的分层表征，从而建模输入输出变量之间的复杂非线性关系。与传统机器学习方法需要手动干预特征提取不同，深度学习自动化这一过程，使模型能够直接从原始数据中发现相关模式。
形式上，深度学习模型可表示为参数化函数，其中x为输入向量，$\hat{y}$为估计输出，L为层数，$\theta^{(i)}$表示深度模型中第i层的可调参数集合。训练过程中，模型参数通过梯度下降和反向传播进行优化以最小化损失函数$\mathcal{L}$，该函数量化预测与真实标签之间的误差。
为最小化损失函数，采用梯度下降算法。链式法则使反向传播成为可能；它描述了梯度如何从一层传递到另一层以及损失相对于模型参数的变化，梯度通过逐层乘偏导数并在训练样本上取平均获得。
3.2. Convolutional Neural Networks
卷积神经网络（CNN, Convolutional Neural Network）是深度学习中的一种基础架构，专门设计用于处理具有空间结构的数据，如图像。其有效性在于应用卷积操作提取局部特征，通过在输入上滑动滤波器来识别边缘、纹理和形状等模式，同时保持像素之间的空间关系。
CNN已成为计算机视觉的基石，在图像分类和目标检测等任务中表现卓越。其成功催生了VGGNet、Inception、DenseNet和ResNet等高级架构，这些架构增强了特征提取和表征学习能力。鉴于其自动学习判别性视觉特征的能力，CNN特别适用于城市环境中身体暴力检测等应用。
从数学角度，输入图像I与滤波器K之间的二维离散卷积定义为：$S(i,j) = \sum_m \sum_n I(i+m, j+n)K(m,n)$，其中$I(m,n)$表示位置$(m,n)$处的输入图像值，$S(i,j)$表示激活图在位置$(i,j)$的值。
在实际CNN实现中，卷积操作由两个关键参数控制：步幅（Stride）参数s指定滤波器在输入图像上的位移；填充（Padding）参数p在图像边界周围添加零值像素。这些参数共同控制输出特征图的空间维度。
由于CNN本质上是 introduced by [48], are an extension of the original GAN architecture that incorporates specific convolutional constraints for image generation. In the context of this study, the DCGAN is employed as a generative model to create artificial images that exhibit the visual characteristics of the minority class, equalizing the cardinality of both classes without duplicating existing samples.
The mathematical foundation of the DCGAN is based on a min-max optimization, which is inherent to generative adversarial networks. The architecture is composed of two neural networks trained jointly: a generator $G(z; \theta_g)$, parameterized by $\theta_g$, whose objective is to generate synthetic samples indistinguishable from real data, and a discriminator $D(x; \theta_d)$, parameterized by $\theta_d$, which seeks to differentiate between real samples and those artificially generated by the generator [48]. Formally, the generative model can be expressed as shown in Equation (32).
In turn, the discriminator $D(x; \theta_d)$ receives as input both real samples and synthetic samples produced by $G$, and outputs a probability that the input is real, i.e., $D(x) \in [0,1]$. Both networks are optimized adversarially through the min-max objective function defined in Equation (33).
During the optimization process, the discriminator adjusts its parameters $\theta_d$ to maximize this function, progressively improving its discriminative capacity through the gradient expressed in Equation (34).
Complementarily, the generator optimizes its parameters $\theta_g$ to minimize the probability that its samples are detected as synthetic, which translates into the gradient presented in Equation (35).
This adversarial dynamic establishes a competition between both networks in which, ideally, the generator produces data progressively closer to the original distribution $p_{data}$, converging toward a faithful estimation thereof.
DCGAN was configured with specific parameters to capture the minority class's characteristic visual patterns, increasing diversity without duplicating instances (Table 5). The hyperparameters used in its implementation were selected based on previous studies and empirically tuned to ensure training stability.
4.4.2. DeepSMOTE
It combines synthetic data generation with an architecture that ensures diversity in the new samples, enabling the model to generalize better under extreme imbalance conditions [12]. The main component of DeepSMOTE is an autoencoder architecture, designed to learn a compact representation of the data in a lower-dimensional feature space.
In Equations (36) and (37), the architecture used for the implementation of this technique is formally expressed.
The main objective of the autoencoder is to minimize the reconstruction loss, which is calculated as the mean squared error (MSE) between the original image $x$ and its reconstruction $\hat{x}$. Equation (38) formally presents this loss.
The hyperparameters of DeepSMOTE were configured to control both the latent-space dimensionality and the variability of the generated samples. Table 6 presents the values used in this implementation, which were selected based on previous studies and empirically tuned to ensure stability during training.
Generative models are incorporated as a complementary balancing strategy and evaluated under controlled imbalance scenarios (Figure 5). Specifically, the DCGAN and DeepSMOTE architectures are used to generate synthetic minority-class samples, with the number of generated instances corresponding to the difference between the majority and minority classes. Unlike traditional oversampling approaches based on replication or interpolation, these methods learn the underlying data distribution and create new samples from latent representations, thus increasing the diversity and representativeness of the minority class. By enriching the training data with semantically meaningful instances of physical violence, generative models can reduce the bias toward the majority class and improve classification performance under varying levels of imbalance.
4.5. Data Augmentation
Data augmentation applies transformations to existing images, such as rotations, scaling, and adjustments to brightness and contrast, to simulate common variations in surveillance scenarios [49]. In image recognition tasks, these transformations increase the diversity of the training set without requiring additional data collection, improving generalization and reducing overfitting. Furthermore, when applied preferentially to minority classes, data augmentation can partially mitigate class imbalance by generating additional training samples and promoting the learning of more robust and discriminative feature representations [14].
Data augmentation techniques are applied, consisting of geometric and photometric transformations on images from the minority class. These include rotations, scaling, adjustments to brightness, contrast, and saturation, as well as horizontal flipping, simulating common variations in urban surveillance environments. As shown in Figure 6, these transformations are applied across the different controlled imbalance scenarios. This approach helps improve the model's robustness to variations in lighting, perspective, and composition, enhancing generalization in real-world contexts.
4.6. Execution Environment Setup
To ensure reproducibility and optimal performance in data processing, the experiments were conducted in a carefully configured hardware and software environment. The main model was developed in PyTorch, although for specific cases of synthetic data generation (DCGAN and DeepSMOTE), TensorFlow was used. Table 7 details the hardware configuration and the main libraries employed.
All subsets generated through the different sampling strategies (traditional, generative, and data augmentation) are used to train a ResNet-18 model adapted for binary classification. The performance of each technique is primarily evaluated using the g-mean metric, due to its sensitivity to class-wise performance in imbalanced contexts. Additionally, complementary metrics such as Accuracy, Precision, Recall, and F1-Score are reported to provide a more comprehensive assessment of the classifier's behavior. This experimental setup enables identification of the most effective balancing technique based on the degree of imbalance, confirming that the choice of method should be tailored to the specific distributional characteristics of each dataset.
To ensure a fair and deterministic comparison across the extensive configuration space, all model training runs were executed under a fixed random seed protocol. This approach guarantees that performance differences are attributable to the balancing techniques rather than to stochastic initialization or variability in data shuffling. While this design prioritizes comparative consistency and reproducibility over stochastic variance estimation, it is consistent with standard practices in large-scale benchmarking studies [50].
4.7. Significant Statistical Test
To statistically validate the differences observed among the balancing techniques, the nonparametric Friedman test was applied. This test allows comparison of multiple algorithms across different datasets by analyzing average ranks [51]. The Friedman test statistic is calculated according to Equation (39).
In cases where the null hypothesis was rejected, multiple comparison post-hoc tests were conducted using the Holm and Shaffer methods to identify pairs of techniques with statistically significant differences in performance. The Holm procedure involves ordering the p-values from smallest to largest and sequentially comparing them with adjusted thresholds of the form $\alpha/(k-i+1)$, where k is the total number of comparisons and i is the hypothesis index. If a hypothesis cannot be rejected, all subsequent hypotheses are retained. In contrast, the Shaffer method applies a stricter correction, using $\alpha/(k-i)$ as the threshold, where $t_i$ represents the maximum number of hypotheses that can be true given that $i-1$ have already been rejected [53].
5. Results and Discussion
This section presents the empirical analysis that constitutes the third core contribution of this systematic review: the controlled evaluation of balancing techniques across multiple datasets and imbalance scenarios. The experimental results are interpreted in light of the theoretical framework, providing mechanistic insights into why specific techniques outperform others under varying degrees of class asymmetry. The findings summarized in Table A1 reveal a clear correlation between the proportion of minority class representation and classification performance, while also demonstrating that no single balancing strategy achieves universal superiority. Instead, the effectiveness of each method is shown to be context-dependent, contingent on both the level of imbalance and the intrinsic characteristics of each dataset. These results validate the experimental design and contribute to the structured synthesis of state-of-the-art methods applicable to imbalanced violence detection, fulfilling the integrative objective of this empirical review.
As shown in Figure 7, in an extreme imbalance scenario of 1%, ROS achieves the best performance with an average g-mean of 0.750, outperforming more sophisticated techniques such as SMOTE (0.706) and DCGAN (0.687). In contrast, Near Miss exhibits the lowest performance (0.131), highlighting the limitations of undersampling methods when the minority class is extremely small.
The superior performance of ROS in the 1% scenario can be interpreted through the lens of sample scarcity and distributional learning. When the minority class is extremely underrepresented (only 1% of the total samples), generative models such as DCGAN and DeepSMOTE face a fundamental limitation: they lack sufficient real instances to learn a stable and representative latent distribution. In such conditions, the autoencoder in DeepSMOTE or the generator in DCGAN is forced to extrapolate from a highly sparse sample space, producing synthetic instances that may introduce noise or mode collapse rather than meaningful diversity. This phenomenon aligns with theoretical analyses of generative oversampling in low-sample regimes, where the quality of synthetic data degrades as the cardinality of the minority class decreases [12,48].
In contrast, ROS, despite its well-known risk of overfitting, preserves the original distribution of the minority class without introducing synthetic artifacts. By duplicating existing samples, ROS provides the classifier with a stronger and more reliable learning signal under extreme scarcity, even if it does not increase feature diversity. This interpretation is further supported by the poor performance of Near Miss (g-mean = 0.131), which confirms that aggressive undersampling in extreme imbalance scenarios discards critical information from the majority class, collapsing the decision boundary and rendering the classifier unable to discriminate between classes. The results suggest that under extreme imbalance, the priority is to maximize the representation of the minority class, even at the cost of diversity, rather than to generate synthetic samples from an unreliable distribution.
In the second scenario, Figure 8 shows a notable increase in the average g-mean values as the representation of the minority class rises to 5%. ROS maintains its superiority (0.922), followed by TL (0.904) and Data Augmentation (0.883). These results confirm the effectiveness of oversampling strategies under highly imbalanced conditions, while NM, although improved (0.398), continues to exhibit lower performance.
As the minority class representation increases to 5%, ROS maintains its superiority (g-mean = 0.922), but the performance landscape begins to shift. Tomek Links (TL) emerges as a competitive alternative (g-mean = 0.904), and Data Augmentation achieves a notable performance (g-mean = 0.883). This behavior can be explained by the fact that, with a slightly larger minority sample, the decision boundary between classes becomes more defined, allowing boundary-refinement techniques like TL to operate effectively. TL removes majority class instances that form ambiguous pairs with minority samples, thereby sharpening the classification frontier without discarding the entire majority class, as occurs with Near Miss or Random Under Sampling.
Data Augmentation also benefits from the increased minority representation, as geometric and photometric transformations (rotation, flipping, brightness adjustment) can generate meaningful variations without altering the semantic content of violent actions. However, generative models such as DCGAN (g-mean = 0.687) and DeepSMOTE (g-mean = 0.671) still underperform relative to ROS, confirming that 5% imbalance remains insufficient for these architectures to learn a stable latent distribution. The results indicate that in highly imbalanced scenarios (1–5%), traditional oversampling and boundary-refinement methods are more robust than generative approaches, which require a minimally representative minority sample to function effectively.
If the minority class increases to 10%, as shown in Figure 9, a generalized improvement in classifier performance is observed. ROS continues to lead (0.961), while generative techniques such as DCGAN (0.951) and DeepSMOTE (0.941) show comparable results. This suggests that, in moderately imbalanced scenarios, synthetic data generation approaches can achieve performance levels close to those of traditional oversampling methods.
When the minority class proportion reaches 10%, classifier performance improves substantially across all balancing techniques, and the differences among methods become less pronounced. ROS remains the best-performing approach (g-mean = 0.961), while DCGAN (g-mean = 0.951) and DeepSMOTE (g-mean = 0.941) achieve comparable results. This trend indicates that generative oversampling becomes increasingly effective as more minority-class samples are available for learning the underlying data distribution.
The improved performance of generative models can be attributed to their ability to construct more accurate latent representations from a larger set of minority-class examples. In particular, DeepSMOTE benefits from a more faithful characterization of the minority-class manifold, generating synthetic instances that increase feature diversity, while DCGAN achieves greater training stability through a more balanced adversarial learning process. These results suggest that a 10% minority-class representation constitutes a practical threshold at which generative approaches become reliable alternatives to traditional oversampling methods, providing competitive performance in moderately imbalanced scenarios.
In the experiments where the training set is balanced at 25%, as shown in Figure 10, most methods exceed an average g-mean of 0.90. ROS (0.979) and DCGAN (0.976) achieve the best results, while RUS (0.959) demonstrates that undersampling methods can remain competitive under moderately low imbalance levels.
In the 25% balance scenario, most methods achieve an average g-mean above 0.90, with only minor performance differences among techniques. ROS (g-mean = 0.979) and DCGAN (g-mean = 0.976) deliver the best results, while RUS also performs competitively (g-mean = 0.959). As the class distribution becomes more balanced, the classifier receives a less biased learning signal, reducing the impact of the specific balancing strategy and leading to a convergence in performance.
The competitive results obtained by RUS indicate that, under moderate imbalance, removing majority-class samples does not significantly affect the model's ability to learn effective decision boundaries. This contrasts with highly imbalanced scenarios (1–5%), where the loss of information caused by undersampling severely degrades performance. Consequently, the effectiveness of undersampling methods appears to be strongly dependent on the degree of imbalance. Moreover, the small difference between ROS and DCGAN suggests that, at this level of imbalance, both traditional and generative oversampling strategies provide sufficient minority-class representation to support effective learning.
Furthermore, in Figure 11, with a 50% balance, a convergence in the performance of the different methods is observed. ROS and One-Sided Selection (OSS) achieve the highest performance (0.984), followed by Data Augmentation and DCGAN (0.981). This indicates that the specific choice of method becomes less relevant as the class imbalance decreases.
With a 50% balance, the performance differences among the evaluated methods become minimal. ROS and One-Sided Selection (OSS) achieve the highest g-mean value (0.984), closely followed by Data Augmentation and DCGAN (0.981). This convergence suggests that, as class distributions become more balanced, the influence of the balancing strategy diminishes, and performance is increasingly determined by the intrinsic characteristics of the dataset and the classifier architecture.
The strong performance of OSS can be attributed to its hybrid mechanism, which combines the Condensed Nearest Neighbor rule with Tomek Link removal. By retaining informative majority-class samples while eliminating noisy and redundant instances near the decision boundary, OSS improves class separability without the information loss often associated with random undersampling. These results indicate that, under moderate imbalance conditions, refining the decision boundary becomes more important than increasing minority-class representation, making hybrid selection methods particularly effective.
In the most favorable balance scenario evaluated (75%), all methods achieve g-mean values above 0.96, and the performance gaps between techniques become negligible (Figure 12). DCGAN (g-mean = 0.989) and DeepSMOTE (g-mean = 0.988) achieve the highest performance, though with minimal differences compared to ROS and Data Augmentation (g-mean = 0.987). Even Near Miss attains an acceptable g-mean (0.960), confirming that performance disparities between methods decrease significantly under conditions of greater balance. The slight advantage of generative models in this scenario can be explained by the fact that, with a sufficiently large minority sample, these architectures can learn a more faithful and diverse representation of the minority distribution, producing synthetic instances that enrich feature variability without the redundancy inherent to ROS.
This result aligns with the theoretical expectation that the benefits of generative oversampling are contingent on the availability of a minimally representative minority sample [12]. However, the marginal differences between techniques (less than 0.002 in g-mean) suggest that, under near-balanced conditions, the choice of balancing method has a negligible impact on classifier performance, and other factors such as hyperparameter tuning, data quality, or architectural choices may be more critical.
The nonparametric Friedman rank test was applied to each imbalance scenario, since each balance level represents a distinct experimental condition with different class distributions and classification difficulty. Consequently, it would be inappropriate to perform a single global comparison across all scenarios, since differences in class proportions could bias the results. Thus, each Friedman test was designed to determine whether significant performance differences existed among the evaluated balancing techniques within a specific imbalance condition. No inferential comparisons were conducted between different balance percentages. When statistically significant differences were detected, post hoc analyses were performed using the Holm and Shaffer procedures to control the family-wise error rate for multiple comparisons. This condition-specific statistical framework is consistent with recent studies in imbalanced learning, where algorithm performance is analyzed separately to account for the contextual dependence of classification results [54,55,56].
Based on the results shown in Figure 13 and obtained through Holm and Shaffer post-hoc tests, it is observed that the number of statistically significant differences between techniques varies considerably according to the degree of imbalance. In highly asymmetric scenarios (1% and 5%), multiple pairs with significant differences are identified, highlighting greater sensitivity of performance to the selected balancing technique. As the proportion of the minority class increases, this disparity tends to decrease, with a convergence in technique performance observed, particularly in the 50% and 75% scenarios, where the matrices show a lower density of significant comparisons. This behavior suggests that under more balanced conditions, the choice of balancing method has a less decisive impact on model performance. Additionally, the analysis reveals that certain techniques, such as ROS, DCGAN, and SMOTE, more frequently appear in statistically superior comparisons, especially in extreme imbalance scenarios. Conversely, methods such as NM and RUS exhibit more limited performance, being outperformed in multiple configurations. This empirical evidence underscores the necessity of selecting balancing strategies based on the degree of asymmetry in the data, avoiding methodological generalizations that could compromise the model's sensitivity to less prevalent classes.
6. Conclusions
This study evaluated different balancing techniques applied to the classification of images depicting physical violence. The results show that the effectiveness of each method varies depending on the degree of class imbalance. In general, classes with low prevalence caused biases in deep neural network models, which affected the identification of samples from the minority class. When evaluating the strategies primarily using the g-mean metric, improvements in model generalization were observed, particularly in scenarios with more asymmetric distributions.
ROS was the most effective strategy under extreme imbalance conditions (1% and 5%), achieving the highest average g-mean values (0.750 and 0.922, respectively). This indicates that, when the minority class is severely underrepresented, duplicating existing samples provides the classifier with a more stable learning signal than generating synthetic instances from a limited distribution. ROS also has practical advantages: simplicity, low computational cost, and preservation of all majority class samples while increasing minority class representation. However, its main limitation is the risk of overfitting, as duplicated samples may cause the model to learn the same visual patterns repeatedly without actually increasing feature diversity.
Traditional undersampling methods exhibited more variable performance. Techniques such as RUS, NM, TL, ENN, and OSS can help reduce majority class dominance and refine the decision boundary. However, their effectiveness depends on the degree of imbalance and the amount of information removed from the majority class. In extreme imbalance scenarios, NM and RUS showed limited performance, suggesting that aggressive removal of majority class samples may discard useful information and reduce generalization ability. In contrast, TL, ENN, and OSS proved more suitable as data cleaning methods, as they remove noisy, redundant, or decision-boundary-near samples rather than randomly reducing the majority class. In this study, OSS achieved competitive performance in moderately imbalanced settings, particularly in the 50% balance scenario, where it attained an average g-mean value of 0.984.
Generative methods demonstrated increasing competitiveness as class balance improved. In the 75% scenario, DCGAN and DeepSMOTE achieved the highest performance (g-mean of 0.989 and 0.988, respectively), indicating that these methods benefit from a sufficient number of minority samples to learn a stable latent distribution. However, their effectiveness under extreme imbalance is limited by the scarcity of minority samples, which hinders the ability of generative architectures to capture the underlying distribution. Additionally, these methods may incur higher computational costs than traditional oversampling, which must be considered in resource-constrained applications.
Data augmentation provided a competitive alternative across various scenarios, increasing visual variability through transformations such as horizontal flipping, rotation, brightness adjustment, contrast modification, saturation changes, and rescaling, without altering the semantic content of the original images. This is particularly important in physical violence classification, where variations in lighting, perspective, scale, resolution, and capture conditions are common. However, its effectiveness depends on the appropriateness of the transformations for the specific domain; overly limited transformations may not provide sufficient diversity, while overly aggressive ones may distort visual patterns critical for violence detection.
Overall, statistical analysis using the nonparametric Friedman test confirms that there is no universally optimal balancing technique. The performance of each method depends on the class ratio, dataset characteristics, the amount of information available for the minority class, computational cost, and the risk of overfitting or information loss. Accordingly, the selection of a balancing strategy should not be generalized but must be tailored to the degree of imbalance and the characteristics of the application domain.
6.1. Limitations and Scope
We acknowledge that the use of a single backbone (ResNet-18) constitutes a limitation of this work, as it prevents assessing whether the observed rankings of balancing techniques generalize to more recent or deeper architectures. The deliberate choice of a stable, widely used backbone was motivated by the need to isolate the effects of the balancing strategies under controlled conditions, but this implies that the conclusions should be interpreted within this specific experimental setting. Accordingly, the findings of this study should not be read as universal claims about the superiority of any particular technique, but rather as context-dependent evidence of how balancing strategies behave under controlled imbalance conditions, with the classifier held constant.
It is also important to consider the computational implications of the evaluated techniques. Traditional methods such as ROS and SMOTE usually have low computational costs and can be applied efficiently even in resource-constrained environments. In contrast, generative models require significant computational resources for training, including GPU acceleration and extended training times (300–500 epochs, compared to 2 epochs for classifier fine-tuning). This trade-off between performance and computational cost should be carefully considered when selecting a balancing strategy for practical applications, particularly in real-time surveillance systems where computational efficiency is critical.
6.2. Future Work
Several directions emerge from this study for future research. First, extending this empirical evaluation to architecturally diverse backbones, including modern efficient architectures such as EfficientNet-B0, ConvNeXt-Tiny, or lightweight Vision Transformers, represents a natural and necessary continuation of this work, as it would allow assessing whether the observed rankings of balancing techniques generalize across different model capacities. Second, we recommend studying hybrid strategies that combine the advantages of different approaches. One option is to first apply an oversampling method, such as ROS, to increase the representation of the minority class, especially in scenarios of severe imbalance. Subsequently, cleaning or selection techniques, such as TL or ENN, could be incorporated to remove noisy, redundant, or out-of-class samples near the decision boundary. This combination would allow taking advantage of the increased representation offered by oversampling while simultaneously reducing noise and refining the decision boundary using subsampling-based methods. Third, it would be worthwhile to explore adaptive hybrid methods that automatically select the most appropriate balancing strategy based on the level of imbalance, the complexity of the dataset, and the performance achieved during validation. Finally, the evaluation of model robustness across multiple random seeds constitutes a valuable extension for future research, as it would provide deeper insights into the stochastic variability of balancing techniques under different experimental conditions.
