Machine Learning Models and Applications for Early Detection

1. Introduction

Machine learning (ML) has become a discipline that automates repetitive and complex tasks through its algorithms, thereby increasing operational efficiency across various organizations. ML analyzes large amounts of data to identify patterns and trends, aiming to improve decision-making in different contexts [1,2,3,4,5,6,7]. This discipline of artificial intelligence trains models based on data analysis to make automatic predictions, allowing the models to deduce correct labels based on the learning acquired from historical data.

Although ML has advantages over classical methods used in multiple areas, each ML model is unique. Each model is trained with data of different characteristics that must be identified to make correct predictions with a high degree of accuracy and precision (e.g., numerical, alphanumeric, and discrete data). Additionally, ML training takes long time due to large volumes of data that the models require [3].

Another associated disadvantage is the interpretability of the data [8], which limits the understanding of the model to make high-quality predictions. There is also an ongoing need for high-quality data, which must be sufficiently reliable and that effectively allows expressing the problem statement. In addition to this, the need to have a sufficient amount of data to train the models can be costly and time-consuming to collect [9]. There is also the risk of bias. Poor data quality during the training, validation, and testing can lead the model to make inaccurate decisions [10].

In general, despite the multiple disadvantages mentioned earlier, ML continues to have advantages in automating processes, improving decision-making, and personalizing services by analyzing user preferences to enhance customer experience. ML can revolutionize businesses and society in general. Currently, ML is a trending tool that every organization must constantly monitor to remain relevant and increasingly competitive [11].

Early Detection (ED) is an important field in ML. The implementation of accurate predictive models for Early Detection (ED) improves outcomes by optimizing time and resources, enhancing problem prevention, and contributing to the improvement of policies for early decision-making [9]. ML-based Early Detection (ED) is currently actively used in socially impactful areas such as production processes, agronomy, energy efficiency, fraud detection, and medicine.

ML-based Early Detection (ED) in production processes can improve aspects such as quick intervention in automated processes, limiting negative impacts on operations, and preventing potential damages. It also allows cost savings by reducing downtime and cutting maintenance expenses. Additionally, it enables greater reliability in automated processes, ensuring consistent and accurate results, improving yields, and maximizing productivity and production [4,9].

ML-based Early Detection (ED) in agronomic processes is used to improve crop health, maximize resource management, increase yields, mitigate climate change, and promote sustainable agricultural practices. Farmers can ensure food security and environmental sustainability by proactively addressing challenges and making better decisions through the use of ML and new technologies [6,12].

In energy systems, ML-based Early Detection (ED) helps optimizing energy consumption, reducing costs, and improving sustainability in various sectors such as electric grids and microgrids. ML algorithms are used to detect faults or anomalies at an early stage, aiding in fault prevention, improving reliability, and prolonging the lifespan of critical assets, such as power transformers [13]. Additionally, ED through ML can enable the implementation of preventive and proactive maintenance strategies, leading to greater operational efficiency and minimizing downtime in industrial energy systems [3].

ML-based Early Detection (ED) in medicine is used in public health programs to reduce healthcare costs and improve procedures with the aim of achieving more satisfactory outcomes in patients [14,15,16]. ML-based Early Detection (ED) has shown promising results in the prediction and early identification of diseases such as diabetes [17,18,19,20], cardiovascular diseases [21], breast cancer [10,18,22,23,24] and dementia [25].

Currently, medical personnel can detect diseases in early stages, facilitating rapid interventions and the application of preventive measures more tailored to the symptoms of the diseases [26]. Additionally, improvements have been found in patient care, optimizing treatment plans [27] and achieving more accurate diagnoses. Continuously, ML promotes proactive healthcare management [15], which improves the sustainability of healthcare systems and contributes to health improvement initiatives.

Also, ML-based Early Detection (ED) in fraud detection is used for several reasons. It allows timely intervention, preventing further fraudulent activities and minimizing potential financial losses and damages [28]. It helps reduce negative impacts on organizations and financial systems by detecting fraudulent activities before they escalate into larger issues. Additionally, it allows for saving resources by preventing larger losses that may occur if fraud is detected at a later stage [29]. ML-based fraud ED helps maintain the integrity of data and financial records by restricting unauthorized access and manipulation of information [30]. ML-based ED provides valuable information about fraudulent patterns and trends to enhance fraud detection strategies. In general, ML-based ED allows for risk mitigation, asset protection, maintaining trust between entities, ensuring security and stability in financial organizations and government fiscal surveillance systems.

Given the importance of ML-based ED, this article aims to conduct a systematic literature review to identify the most commonly used ML models (MLMs) in ML-based ED in the aforementioned areas. As a second objective, this article aims to identify new methodologies for using these MLMs to improve classification or prediction in ED. This literature review is motivated by the exploration of MLMs currently used in ED for fraud detection. The intention of this article is to identify how ML-based ED models have impacted the field of fraud detection by analyzing their advantages and disadvantages, and to present a discussion on the benefits that the fraud domain can obtain as part of Fiscal Surveillance and Control.

This review aims to provide academics and professionals with guidance in their work, facilitating the quick identification of current algorithms and methodologies used in ML for the application of ED in the referenced areas and in the field of fraud. The main contributions of this article are as follows:

• Presentation of the most used MLMs for ED.
• Division of MLMs for ED into two main categories: Single Base Model (SBM) and Stacking Ensemble Model (SEM).
• Identification of SBM or SEM in ED for fraud.
• Discussion on how ML-based ED can improve processes in fraud.

The article is structured in the following sections: Section 2 describes the research article selection process for a systematic literature review on MLMs for ED. Section 3 gives an overview of data balancing and model validation metrics currently used in machine learning. Section 4 gives an overview of the machine learning models found and their performance in multiple applications and specifically in fraud detection. Section 5 discusses the performance of the machine learning models found for early detection in multiple areas and the importance of using these models in early fraud detection. Finally, conclusions are presented.

2. Article Selection Process

A systematic literature review is a research approach that examines information and findings regarding a research topic [31]. This approach aims to locate the largest possible number of relevant studies on the subject of study and, through referenced or proprietary methodologies, determine what can be confidently asserted from these studies [32,33]. This section provides an overview of the literature to help understand the MLMs for ED used in the present literature.

In this article, the process of searching and selecting articles consists of two stages, aiming to answer the following two questions:

• RQ1: Which MLMs are currently used in the literature for early detection in multiple areas?
• RQ2: How have these MLMs for ED been implemented in the context of fraud?

In stage 1, based on RQ1 and RQ2, Scopus was selected as the search engine, and the following keywords were used for the search equation: "machine learning model", "data analytics" or "data analysis", "early detection", and finally "fraud detection". These five keywords formed the first search equation within the time frame between 2018 and 2024. As a result of this search, a total of 55 scientific articles were obtained, included in electronic databases such as Springer Link, Elsevier, IEEE, MDPI, Taylor and Francis.

Figure 1 presents an analysis of the occurrence of words in the selected articles. The color map in Figure 1 illustrates the frequency of recurring concepts in the literature found. It is inferred, then, that models such as Logistic Regression (LR), Support Vector Machines (SVM), Decision Trees (DT), Data analytics (DA) and Random Forests (RF) are base models used in ED or early diagnosis.

Having reviewed the information from this preliminary search, in stage 2, it was identified that currently in ML, from a methodological aspect, with the aim of improving the prediction of base models, stacking ensemble has been implemented in recent years. This allowed extending the initial search equation by including the keyword "Stacking Ensemble" with a focus on ED. As a result, a total of 57 articles related to ML-based ED were obtained. Additionally, a search matrix was synthesized to identify the contribution of each article, the base MLMs (44 articles) and the ensemble MLMs (12 articles), and their application areas. Finally, the necessary information was extracted regarding applications and MLMs applied to ED in multiple areas (45 articles), particularly in fraud detection (11 articles).

3. MLM Data Balancing and Performance Metrics

In the literature review consulted, various procedures were evident for conducting proper validations of MLMs, such as data balancing and the application of performance metrics. Validating MLMs is crucial to ensure their reliability and effectiveness in decision-making. Validation allows for evaluating the predictive capacity of models, identifying potential issues like overfitting, and ensuring that the results obtained are generalizable and consistent with new data [34]. In model validation, it can be verified its performance, accuracy, and capability to handle different scenarios, which helps ensure that they are useful and reliable in real-world applications [17].

3.1. Data Balancing

Data balancing in MLMs addresses the problem of class imbalance. One class may have significantly more instances than another in a dataset. When classes are imbalanced, models tend to favor the class with more instances, which can lead to poor performance in predicting the class with fewer instances.

Using data balancing techniques in the preprocessing of information to train a ML model improves the model's ability to learn patterns from all classes equitably, resulting in more accurate, precise, and generalizable predictions. The purpose of balancing is to achieve an equilibrium where the detection of both minority and majority classes is of interest.

Data balancing can be achieved through techniques such as oversampling (duplicating instances of the minority class), undersampling (removing instances of the majority class), or more advanced methods like SMOTE (Synthetic Minority Over-sampling Technique) [35]. These techniques help improve the predictive capability of models by ensuring that all classes are treated equitably during training.

3.2. Performance Metrics

Performance metrics in ML are measures used to evaluate the performance and effectiveness of machine learning models in data prediction and classification. In the literature review consulted, metrics such as accuracy, precision, recall (sensitivity), specificity, F1-score, AU-ROC (Area under receiver operating characteristics curve), AU-PRC (Area under the precision-recall curve), MCC(Matthews correlation coefficient) and confusion matrix [34] are used. These metrics provide information about the predictive capability, accuracy, and overall effectiveness of MLMs. All the mentioned metrics are related to the confusion matrix. The confusion matrix (Figure 2) is a tool that allows visualizing the performance of a classification model by showing the number of true positives (TP), true negatives (TN), false positives (FP), and false negatives (FN) that the model has produced on a test dataset [22]. Table 1 presents descriptions of the evaluation metrics.

The confusion matrix can also be extended to multiclass problems. It is not necessarily intended only for binary problems. In a multiclass classification context, the confusion matrix is expanded to include all classes present in the problem. Thus, its construction will have a length of N, corresponding to the number of classes.

4. Machine Learning Models for ED and Applications

According to the literature review, two methodologies for applying MLMs in multiple applications were identified. These methodologies describe the use of MLMs from the following perspectives.

Single Base Models (SBM). These base models serve as individual classifiers or regressors that make predictions based on input data. Among the most common base models are LR (Logistic Regression), SVM (Support Vector Machine), DT (Decision Tree), RF (Random Forest), NB (Naive Bayes), K-Nearest Neighbors (KNN), and Neural Networks (NN). When using a single model, the choice depends on the characteristics, distribution, and properties of the datasets [16]. SBMs are also used as a basis for ensemble methods and more complex stacking models [36]. Table 2 presents a brief description of SBMs.

Stacking Ensemble Model (SEM). They involve combining multiple base models to improve the predictive performance of SBM. These models use a two-level stacking approach; base models make predictions at the first level, and meta-learning combines these predictions at the second level [13,16]. The purpose of this ensemble methodology is to combine two or more models, each with its strengths and weaknesses, to construct a more robust model. Stacking ensemble models have proven to be promising in various applications by offering advantages such as improved accuracy, reduced overfitting, and enhanced performance compared to individual models [37]. SEM models use boosting, bagging, and stacking schemes [30]. Each SEM operates within its own domain space, showing varying levels of performance based on the aggregated selection of base models and the distribution, non-linearity, and class imbalance present in the dataset. Some of the most popular boosting algorithms are AdaBoost (Adaptive Boosting), Gradient Boosting, XGBoost, and LightGBM [13]. These algorithms have variations in how they adjust weights and combine weak models to form the final model, but they follow the general scheme of boosting. An example of bagging is RF, where multiple decision trees are trained on training datasets generated by bootstrap sampling (sampling with replacement), and predictions from individual trees are averaged to produce the final prediction.

Within the literature review, different types of SBMs and SEMs were found. Under the search parameters in the matrix of sintering literature review, for ED, different configurations of these models were obtained in areas such as Medicine (37 articles), fraud detection (11 articles), agronomy (2), energy efficiency (2 articles), industrial processes (2 articles), education (1 article), and telecommunications (1 article). Table 3 and Table 4 present the SBMs and SEMs, respectively, for multiple areas excluding articles related to fraud detection, which are analyzed later.

¹Polynomial Regression; ²Gradient Boosting; ³AdaBoost; ⁴Gaussian naive bayes; ⁵Stochastic Gradient Descent; ⁶Gaussian process; ⁷Insolation Forest; ⁸DenseStream.

According to Table 3, the most used SBM for early detection, based on the reviewed literature, are Random Forest (RF), Support Vector Machine (SVM) and K-Nearest Neighbor (KNN). These models are applied in the medical field, achieving average performances above 80% according to the metrics reported by various authors. In other fields, RF also stands out as a frequently used model, with performance levels similarly approaching 80%.

On the other hand, in the SEMs presented in Table 4, it is identified that RF is the most commonly used base model, followed by DT, then LR, SVM, and KNN with the same frequency, followed by ANN and XGB also at the same level. NB is the least used base model for early detection in the literature studied. As the best Metalerner, LR is identified as the most used due to its simplicity in computation and inference for decision-making. It is also identified that the SEM methodology achieves performances in some cases exceeding 90%.

According to Table 4, SEM yields superior results compared with SBM. It is important to note that the SEM approaches found are implemented in the medical field, which is a critical area for decision-making, as a misdiagnosis can have severe repercussions in various health contexts. This justifies why most of the literature on SBM and SEM consulted focuses on the field of medicine, in comparison with other areas for early detection.

Hybrid models acting as MLM and SEM were also found. The selection of base, ensemble, or hybrid models will depend on the working context and the characteristics of the data.

4.1. Machine Learning Models for ED in Fraud Detection

MLMs are used in fraud detection to analyze patterns and behaviors in data in order to identify fraudulent activities. These models are important for early fraud detection (ED) as they can quickly process large amounts of data to detect and classify suspicious transactions or activities, helping to prevent financial losses.

ML-based models for ED offer several advantages over traditional fraud detection methods. MLMs have the ability to adapt and improve over time as more data is analyzed, thereby increasing their accuracy in fraud detection [64]. They can also analyze complex and diverse data sources, allowing them to detect sophisticated and evolving fraud schemes that may go unnoticed by traditional rule-based systems or human intervention [65]. Table 5 presents the MLMs found in the literature for ED for fraud detection.

Table 5 presents the MLMs found in the literature consulted for ED in fraud detection. Only the use of SBM was found for fraud detection. RF persists as the most frequently used model, followed by KNN. LR, SVM and XGB are used at the same level for fraud detection. RF continues to be the most used and reliable model for fraud detection according to Table 5, with two authors reporting better performance with this model. Although RF is based on bagging and XGB utilizes boosting, they are considered base models grounded in DT. Table 5 reflects that in most cases, each MLM achieved a high number of accurate fraud detections in real fraud cases, with authors reporting accuracy metrics exceeding 90%.

All the MLMs found in Table 3, Table 4, and Table 5 were evaluated using the metrics reported in Table 1. Each achieved positive performance metrics in the respective contexts where they were implemented.

5. Discussion

ML algorithms are increasingly used in various fields due to their ability to adapt to new data and identify hidden patterns, enabling decision-making with a higher degree of reliability. Although most models found in the literature work as standalone base models, the use and experimentation with ensemble methodologies to improve the performance of base models is becoming increasingly common.

The information in Table 3 and Table 4 shows that base models are used diversely across multiple areas, unlike SEM. Specifically, Table 4 reports that SEMs work exclusively in the medical field. This is justified because SEMs, by gathering the various decisions from SBM, allow for a more robust acquisition of data variability and a better fit to the data. Consequently, the identification of the problem is more accurate and precise, whether in the context of regression or classification. The unification of these decisions constitutes a more solid knowledge base that serves as input to another model for final decision-making. This is particularly important in the medical field, where decision-making is critical for diagnosing a person, requiring a minimal margin of error.

Within the models in Table 5 for early fraud detection, it was found that the models found in the literature are SBM. Although the authors do not consider SEM models for early fraud detection, the SBM used achieved significant performance with good adaptability to emerging patterns, good training times, and good adaptation to data for fraud detection [1,65,66,68,69,70].

An important aspect to achieve satisfactory results in the training, validation, and testing of SEM and SBM in any context will be the associated data engineering analysis [71]. This refers to the effective selection of the data characteristics that would be supplied as information to the SBM. Additionally, analysis of data balancing techniques to balance major or minor classes [64], such as the SMOTE method [69].

This aspect for SEMs is not considered critical since the patterns of training, validation, and testing are responses from the SBM. However, in this case, cross-validation processes must be ensured to avoid overfitting issues, model selection bias, and errors in variance estimation. The use of SEM strictly must consider cross-validation methods such as K-Fold, Hold-Out, Leave-One-Out, Leave-P-Out, Monte Carlo, Stratified K-fold, and Repeated K-fold, Time Series Cross-Validation, and Nested Cross-Validation [25,52]. These methods provide a more reliable estimation of model performance on unseen data, reducing the risk of overfitting and enabling better hyperparameter tuning and model evaluation.

The use of MLMs in ED for fraud detection, whether as SBM or SEM, will continue to be the subject of study in areas such as corporate security, surveillance, and fiscal and financial control, due to its ability to process large amounts of data rapidly and adapt to new information over time.

A determining factor in MLMs training in fraud ED is the limitations in data collection for model training. Data collection for fraud ED is constrained by the availability of labeled data due to the lack of digitalization of this information. Without digitized data, the process of consolidating a labeled historical dataset is slow and costly, which restricts the applicability of MLMs in fraud ED, especially in cases such as tax fraud detection [29]. The lack of digitized data can hinder the effectiveness and accuracy of fraud ED processes, as manual data handling is time-consuming and error-prone. Moreover, the challenge of labeling transactions as fraudulent or non-fraudulent can be complex due to the difficulty in definitively asserting the fraudulent nature of transactions, leading to careful use of labeled examples and the need for verification by expert personnel to identify such specific fraud, which is also subject to ethical considerations [64].

It is important that in the field of fraud detection in ED, strategies for data balancing will be considered based on available information to reduce intrinsic bias that may include human manipulation or unwanted value judgments when labeling data.

Implementing SEM as MLMs for fraud ED offer several advantages over base models and traditional methods. Some of the advantages that this type of models can offer include:

• Improved prediction performance by maximizing fraud detection through effective identification of patterns and anomalies in the data, leading to better prediction performance. Using features based on SBM responses allows the analysis of behaviors that may not have been explored in traditional methods, thereby enhancing a more comprehensive analysis of fraud indicators. Additionally, the adaptability and robustness of SEM enable it to adjust to the strengths and weaknesses of multiple baseline models, improving overall detection performance and robustness in identifying fraudulent activities [16].
• Combining the predictive power of various models enables the identification of fraudulent behaviors at an early stage with greater accuracy, which allows for timely intervention and prevention of fraudulent activities [20].
• SEM provides a more reliable balance between precision and interpretability, making them operationally viable for fraud detection tasks, due to the adoption of features and operating dynamics of SBMs.

6. Conclusions

The literature review under the search equation allowed for the consolidation of two ways of using MLMs for ED: Single-base models (SBM) and Stacking Ensemble Model (SEM). The implementation of SBM was identified in different areas, whereas SEM models were implemented only in the field of medicine due to the high precision required in this area.

The implementation of SEM can favor and strengthen conventional fraud detection efficiently, allowing to improve prediction performance by leveraging the integration of features from different base models. SEM enables better adaptability to data and robust decision-making. Additionally, it provides more accurate early detections in scenarios with high data variability, reduces issues such as overfitting, and allowing handling biases in the data.

Both SBM and SEM have proven to be efficient in early detection across multiple areas, particularly in fraud, with accuracies in some cases exceeding 90%. For the use of MLMs, it will always be relevant to perform data engineering processes to select appropriate features for model training, and to pay special attention to data balancing to achieve adequate results in predictions.

From the analyzed information, it can be inferred that a challenging task in the field of fraud detection is the consolidation of reliable databases for training MLMs for ED, as well as the adoption of new models and cutting-edge methodologies such as Deep Learning.

Future research lines may be oriented to develop further advancements in techniques and technologies like Deep Learning Models as SBM and SEM to enhance the accuracy, efficiency, and scalability of fraud detection systems. Other potential research lines include: Enhanced Data Enrichment (the quality and quantity of data), Advanced Machine Learning Algorithms in fraud detection (Convolutional Neural Networks and ANN- Long Short-Term Memory), Real-time Fraud and Blockchain Technology for secure and transparent transaction verification and enhancing fraud detection capabilities through immutable and decentralized data storage.