Random Undersampling Research Articles

Vehicle load identification based on bridge response using deep convolutional neural network

ABSTRACT Ensuring bridge safety and longevity through precise vehicle load characterization is crucial for maintaining structural integrity and efficiency. This research highlights the efficacy of deep convolutional neural networks (DCNNs) in analyzing bridge responses. The proposed methodology incorporates abridge model developed in ANSYS APDL, integrated subsequently with Universal Mechanism (UM) software for simulating vehicle-bridge interactions, considering factors such as vehicle load, speed, road surface roughness, and ambient noise. To validate the accuracy and robustness of the DCNN models, the study involved training and testing them on diverse datasets, each comprising 250 randomized samples for various parameters such as load, speed, and surface roughness. Two popular deep learning frameworks, ResNet and Inception models, were employed to assess their capability to identify these critical vehicle-bridge interaction characteristics accurately. The results reveal that the Inception model excelled in vehicle load identification, achieving an impressive validation accuracy of 99.9%. In contrast, ResNet outperformed in vehicle speed identification with a99% validation accuracy, surpassing Inception’s 97.9%. Both models exhibited remarkable resilience to noise, with road surface roughness identification demonstrating exceptional performance, including aperfect 100% training accuracy for ResNet and 99.9% for Inception. Furthermore, to validate the noise robustness of the DCNN model, the load identification is conducted under the effects of measurement noise, and the identification results show that the DCNN model can accurately identify vehicle loads with accuracies of 97.9% at 10 dB and 99.7% at 20 dB, respectively. These findings firmly establish the DCNNs’ capability to accurately identify vehicle load, speed, and road surface roughness, emphasizing their potential to significantly improve bridge safety and longevity.

Early prediction models and crucial factor extraction for first-year undergraduate student dropouts

PurposeThis study aimed to use enrollment information including demographic, family background and financial status, which can be gathered before the first semester starts, to construct early prediction models (EPMs) and extract crucial factors associated with first-year student dropout probability.Design/methodology/approachThe real-world samples comprised the enrolled records of 2,412 first-year students of a private university (UNI) in Taiwan. This work utilized decision trees (DT), multilayer perceptron (MLP) and logistic regression (LR) algorithms for constructing EPMs; under-sampling, random oversampling and synthetic minority over sampling technique (SMOTE) methods for solving data imbalance problems; accuracy, precision, recall, F1-score, receiver operator characteristic (ROC) curve and area under ROC curve (AUC) for evaluating constructed EPMs.FindingsDT outperformed MLP and LR with accuracy (97.59%), precision (98%), recall (97%), F1_score (97%), and ROC-AUC (98%). The top-ranking factors comprised “student loan,” “dad occupations,” “mom educational level,” “department,” “mom occupations,” “admission type,” “school fee waiver” and “main sources of living.”Practical implicationsThis work only used enrollment information to identify dropout students and crucial factors associated with dropout probability as soon as students enter universities. The extracted rules could be utilized to enhance student retention.Originality/valueAlthough first-year student dropouts have gained non-stop attention from researchers in educational practices and theories worldwide, diverse previous studies utilized while-and/or post-semester factors, and/or questionnaires for predicting. These methods failed to offer universities early warning systems (EWS) and/or assist them in providing in-time assistance to dropouts, who face economic difficulties. This work provided universities with an EWS and extracted rules for early dropout prevention and intervention.

Resampling to Classify Rare Attack Tactics in UWF-ZeekData22

One of the major problems in classifying network attack tactics is the imbalanced nature of data. Typical network datasets have an extremely high percentage of normal or benign traffic and machine learners are skewed toward classes with more data; hence, attack data remain incorrectly classified. This paper addresses the class imbalance problem using resampling techniques on a newly created dataset, UWF-ZeekData22. This is the first dataset with tactic labels, labeled as per the MITRE ATT&CK framework. This dataset contains about half benign data and half attack tactic data, but specific tactics have a meager number of occurrences within the attack tactics. Our objective in this paper was to use resampling techniques to classify two rare tactics, privilege escalation and credential access, never before classified. The study also looks at the order of oversampling and undersampling. Varying resampling ratios were used with oversampling techniques such as BSMOTE and SVM-SMOTE and random undersampling without replacement was used. Based on the results, it can be observed that the order of oversampling and undersampling matters and, in many cases, even an oversampling ratio of 10% of the majority data is enough to obtain the best results.

Implementation of K-Nearest Neighbor using the oversampling technique on mixed data for the classification of household welfare status

Welfare is closely related to poverty and the socio-economic disparities in a society. Based on data from the Central Bureau of Statistics, Kulon Progo in Indonesia had the highest poverty rate in the province of the Special Region of Yogyakarta; an increasing trend was observed every year from 2019 to 2021; Kulon Progo also had a low poverty line (after Gunung Kidul) compared to other regencies/cities in this province. This study aimed to classify the household welfare status in Kulon Progo in March 2021 using the K-Nearest Neighbor (KNN) method. Since imbalance was found between the poor and non-poor classes, an oversampling technique was employed. Imbalanced data affect classification, particularly when predicting the results of the classification. The following oversampling techniques were employed in this study: Random Oversampling (RO), the Adaptive Synthetic (ADASYN) and the Synthetic Minority Oversampling Technique (SMOTE). It was found that, of the three techniques, RO was the most efficient with k = 5, which yielded the best performance in terms of sensitivity, specificity, the G-mean, and accuracy reaching 0.643, 0.805, 0.719, and 78.873%, respectively. Therefore, it can be concluded that the classification model performed well enough to classify household welfare status, especially among the poor (minority class).

You must be myths-taken: Examining belief in falsehoods during the COVID-19 health crisis.

The prevalence of health myths is increasing with the rise of Internet use. Left unaddressed, online falsehoods can lead to harmful behaviours. In times of crisis, such as the recent COVID-19 pandemic, the circulation of many myths is exacerbated, often to varying degrees among different cultures. Singapore is a multicultural hub in Asia with Western and Asian influences. Although several studies have examined health myths from a Western or Eastern perspective, little research has investigated online health falsehoods in a population that is culturally exposed to both. Furthermore, most studies examined myths cross-sectionally instead of capturing trends in myth prevalence over time, particularly during crisis situations. Given these literature gaps, we investigated popular myths surrounding the recent COVID-19 pandemic within the multicultural setting of Singapore, by examining its general population. We further examined changes in myth beliefs over the two-year period during the pandemic, and population demographic differences in myth beliefs. Using randomised sampling, two online surveys of nationally representative samples of adults (aged 21-70 years) residing in Singapore were conducted, the first between October 2020 and February 2021 (N = 949), and the second between March and April 2022 (N = 1084). Results showed that 12.7% to 57.5% of the population were unable to identify various myths, such as COVID-19 was manmade, and that three of these myths persisted significantly over time (increases ranging from 3.9% to 9.8%). However, belief in myths varied across population demographics, with ethnic minorities (Indians and Malays), females, young adults and those with lower education levels being more susceptible to myths than their counterparts (p < 0.05). Our findings suggest that current debunking efforts are insufficient to effectively counter misinformation beliefs during health crises. Instead, a post-COVID-19 landscape will require targeted approaches aimed at vulnerable population sub-groups, that also focus on the erroneous beliefs with long staying power.

Effectiveness of chemical inactivation of infectious liquid biological waste: A randomized sample study of research laboratories in Switzerland

Laboratory wastewater has been suggested as an important escape route for microorganisms from research environments. Likely reasons for the unintentional release of laboratory organisms are shortcomings in the handling of infectious liquid biological waste (LBW) and inadequate inactivation procedures. We developed an analytical approach to investigate the use of chemical inactivation (CI) procedures in Swiss research laboratories by on-site random sampling of presumably inactivated infectious LBW and testing it for the presence of infectious lentiviruses (HIV-1) and adenoviruses (AdV). In addition, standard operating procedures (SOPs) for CI were collected and evaluated, and laboratory-staff knowledge of CI processes was assessed using a questionnaire. Although we found several deficiencies in the technical knowledge and training of laboratory staff on the CI of LBW, as documented by 27 returned questionnaires, no infectious viruses were detected in the eight LBW samples collected. Whilst we acknowledge that the number of LBW samples and SOPs is small, we conclude that CI of LBW containing infectious lentiviruses and adenoviruses does not result in the systematic release of considerable amounts of infectious viruses into the environment from research laboratories in Switzerland.

How Does ChatGPT Use Source Information Compared With Google? A Text Network Analysis of Online Health Information.

How Does ChatGPT Use Source Information Compared With Google? A Text Network Analysis of Online Health Information.

Sampling pattern discrepancy in the application of compressed sensing hyperpolarized xenon-129 lung MRI.

Although hyperpolarized (HP) 129Xe ventilation MRI can be carried out within a breath hold, it is still challenging for many sick patients. Compressed sensing (CS) is a viable alternative to accelerate this approach. However, undersampled images with identical sampling ratios differ from one another. Twenty subjects (n = 10 healthy and n = 10 patients with asthma) were scanned using a GE MR750 3 T scanner, acquiring fully sampled 2D multi-slice HP 129Xe lung ventilation images (10 s breath hold, 128 × 80 (FE × PE-frequency encoding × phase encoding) and 16 slices). Using fully sampled data, 500 variable-density Cartesian random undersampling patterns were generated, each at eight different sampling ratios from 10% to 80%. The parallel imaging and compressed sensing (PICS) command from BART was employed to reconstruct undersampled data. The signal to noise ratio (SNR), structural similarity index measurement (SSIM) and sidelobe to peak ratio of each were subsequently compared. There was a high degree of variation in both SNR and SSIM results from each of the 500 masks of each sampling rate. As the undersampling increases, there is more variation in the quantifying metrics, for both healthy and asthmatic individuals. Our study shows that random undersampling poses a significant challenge when applied at sampling ratios less than 60%, despite fulfilling CS's incoherency criteria. Such low sampling ratios will result in a large variety of undersampling patterns. Therefore, skipped segments of k-space cannot be allowed to happen randomly at low sampling rates. By optimizing the sampling pattern, CS will reach its full potential and be able to be applied to a highly undersampled 129Xe lung dataset.

Evaluating Random Forest Algorithm in Educational Data Mining: Optimizing Graduation on-time prediction using Imbalance Methods

The study aims to evaluate the performance of Random Forest algorithms in data mining education by optimizing graduation on-time (GOT) predictions using imbalanced data methods. Methods used to handle imbalanced data include random under-sampling (RUS), random over-sampling (ROS), hybrids of RUS and ROS, synthetic minority over-sampling techniques for nominal classes (SMOTE-NC), and hybrids of SMOTE-NC and RUS. After applying these methods, studies analyze their performance on training and testing data. The research findings show that on training data, the RUS-ROS hybrid showed the best performance compared to other methods, while the SMOTENC and RUS hybrid techniques showed the best performance on testing data based on AUC values. The research showed that the use of an imbalanced data method significantly improved the ability of Random Forest algorithms to predict graduation on time (GOT) in the context of educational data. We discuss the implications for educational data mining applications and provide suggestions for future research.

High-performance imbalanced learning ensembles of decision trees for detecting mineralization anomalies from geochemical exploration data

How to effectively detect geochemical anomalies associated with mineralization is a challenging task due to the extreme class-imbalance of geochemical exploration data. To address this challenge, various machine learning techniques have been employed to detect geochemical anomalies associated with mineralization. However, almost all of these machine learning techniques have their limitations when it comes to modeling class-imbalanced geochemical exploration data. To establish efficient robust high-performance models for detecting geochemical anomalies associated mineralization, a case study was carried out in the Helong area, Jilin Province, China. Decision tree (C4.5) classifiers were used as the base or weak classifiers, four imbalanced learning ensemble models, including self-paced ensemble model, under-bagging ensemble model, synthetic minority oversampling technique (SMOTE)-boost ensemble model and random under sampling (RUS)-boost ensemble model were established and compared in the detection of polymetallic mineralization anomalies from the 1:50,000 stream sediment survey data. The output of the base classifier is a vector consisting of the probability that the sample belongs to a mineralization anomaly and the probability that the sample belongs to the background. The performance of each ensemble model was evaluated using the receiver operating characteristic (ROC) curve and the area under the ROC curve (AUC). The ROC curves of the four models are close to the upper left corner of the ROC space. The AUC values of the four models are not lower than 0.9589. The polymetallic mineralization anomalies detected by the four models account for less than 7 % of the entire study area, but contain not less than 93 % of the polymetallic deposits found in the study area. In addition, the polymetallic mineralization anomalies detected by the four models spatially coincide with the regional controlling factors of polymetallic mineralization in the study area. To summarize, the four models have very high performance in detecting polymetallic mineralization anomalies, and the polymetallic mineralization anomaly detection results are consistent with the regional geological and metallogenic characteristics in the study area. Therefore, the imbalanced learning ensemble techniques are powerful tools for the establishment of high-performance ensemble classification models for detecting mineralization anomalies from geochemical exploration data.

Machine learning-based network intrusion detection for big and imbalanced data using oversampling, stacking feature embedding and feature extraction

Cybersecurity has emerged as a critical global concern. Intrusion Detection Systems (IDS) play a critical role in protecting interconnected networks by detecting malicious actors and activities. Machine Learning (ML)-based behavior analysis within the IDS has considerable potential for detecting dynamic cyber threats, identifying abnormalities, and identifying malicious conduct within the network. However, as the number of data grows, dimension reduction becomes an increasingly difficult task when training ML models. Addressing this, our paper introduces a novel ML-based network intrusion detection model that uses Random Oversampling (RO) to address data imbalance and Stacking Feature Embedding based on clustering results, as well as Principal Component Analysis (PCA) for dimension reduction and is specifically designed for large and imbalanced datasets. This model’s performance is carefully evaluated using three cutting-edge benchmark datasets: UNSW-NB15, CIC-IDS-2017, and CIC-IDS-2018. On the UNSW-NB15 dataset, our trials show that the RF and ET models achieve accuracy rates of 99.59% and 99.95%, respectively. Furthermore, using the CIC-IDS2017 dataset, DT, RF, and ET models reach 99.99% accuracy, while DT and RF models obtain 99.94% accuracy on CIC-IDS2018. These performance results continuously outperform the state-of-art, indicating significant progress in the field of network intrusion detection. This achievement demonstrates the efficacy of the suggested methodology, which can be used practically to accurately monitor and identify network traffic intrusions, thereby blocking possible threats.

A hybrid forecasting framework based on MCS and machine learning for higher dimensional and unbalanced systems

Forecasting methods and theories have been widely researched and applied in complex systems and fields such as statistical physics, econophysics, material crystals, etc. However, challenges persist in applying these methods to complex systems characterized by high dimensionality, data imbalance, and single prediction evaluation. To address these issues, we propose a novel hybrid forecasting approach that integrates the model confidence set (MCS) with machine learning (ML) models. We introduce Principal Component Analysis (PCA) to reduce dimensionality of the data, reduce data imbalance through a combination of random undersampling and oversampling, and introduce several metrics to evaluate the machine learning model set. We also introduce the MCS to select the optimal model from the set of ML models and propose a new combinatorial approach, the MCS-ML combinatorial model. An empirical study is conducted using the example of abnormal transactions in the Bitcoin blockchain. The empirical results show that the proposed MCS-ML combinatorial model has better predictive performance than the models in the ML model set under different data structures.

Column-Randomized Linear Programs: Performance Guarantees and Applications

Solving Large-Scale Linear Programs by Randomly Sampling Columns Large-scale linear programs (LPs) play a pivotal role in various applications, including the classic cutting-stock problem and the vehicle routing problem. The standard solution approach for these LPs, namely, column generation (CG), is often computationally intractable because of the NP-hard nature of the corresponding subproblem in many applications. In “Column-Randomized Linear Programs: Performance Guarantees and Applications,” Akchen and Mišić introduce a randomized method that involves first sampling a set of columns from the original LP and subsequently solving an LP composed of the sampled columns, termed the column-randomized LP. The authors analyze the optimality gap of the column-randomized LPs and establish conditions under which the gap is small. Empirical findings demonstrate the effectiveness of the column-randomized LP approach, showcasing its advantages over the CG approach in two practical applications: the cutting-stock problem and nonparametric choice model estimation.

Research on emotional intelligence among Indian teachers: A Systematic Review and meta-analysis of its correlation with health parameters and impact of gender

Background Emotional intelligence is the self-perception related to identification and regulation of emotions. Several studies have been done among Indian teachers evaluating emotional intelligence in relation to demographic, professional and various psychological parameters, but the variety of scales, teacher types, and conflicting results makes it difficult to draw any meaningful conclusions from this heterogeneous data. The present work aims to synthesize the available data by both qualitative and quantitative analysis and is the first such attempt to include only Indian studies in this field. The main objectives were to determine the correlation of emotional intelligence with teachers’ health parameters and to study the gender difference in emotional intelligence. Methods After a thorough literature search in Google, Google scholar, Scopus, Web of science and Pubmed, fifty-five Indian studies were selected which empirically examined teachers’ emotional intelligence, either alone or in association with another parameter evaluating teachers’ psychological health and performance. After qualitative assessment of major findings, quantitative analysis was performed. Three separate meta-analysis were carried out. The first one with fifteen effect sizes among 3291 participants evaluated correlation with personal health parameters. The second with nineteen effect sizes in 4165 participants evaluated correlation with professional health parameters. The third with twenty-six studies involving 6005 participants assessed effect of gender. Results The results show that almost all studies have used a trait measure, teachers’ emotional intelligence is positively correlated with both personal and professional health parameters and gender has no effect on emotional intelligence. Conclusion Major limitations are a very high degree of heterogeneity of the data, incomplete description of the scales, inadequate randomization and small sample sizes in many studies. The results indicate the importance of emotional intelligence in both personal and professional life of teachers and no effect of gender preparing a solid base for future research.

Imbalance Dataset in Aspect-Based Sentiment Analysis on Game Genshin Impact Review

Sentiment analysis was commonly used to determine the polarity of the review text. However, there is a problem if some reviews have more than one aspect with different polarities, so the reviews have more than one polarity. That has happened in some reviews on the game Genshin Impact. Not merely that, the number of sentiments contained in a review is not always the same as other reviews will cause imbalanced data. So, this study will handle imbalance data with Random Under-Sampling and Random Over-Sampling on aspect-based-sentiment-analysis of Genshin Impact Review with Multinomial Naïve-Bayes, so that the classification prediction does not ignore the minority class due to the dominance of the majority class. The classification process used K-Fold Cross Validation (k=10) validation method and the Laplace smoothing technique on Multinomial Naïve Bayes. As a result, the conclusion is that Random Oversampling had better accuracy than Random Undersampling in handling imbalanced data on aspect-based sentiment analysis of Genshin Impact game Review in Indonesian with Naïve Bayes Multinomial, with the highest accuracy of 85.55%.

Modality-Specific Feature Selection, Data Augmentation and Temporal Context for Improved Performance in Sleep Staging.

This work attempts to design an effective sleep staging system, making the best use of the available signals, strategies, and features in the literature. It must not only perform well on different datasets comprising healthy and clinical populations but also achieve good accuracy in cross-dataset experiments. Toward this end, we propose a model comprising multiple binary classifiers in a hierarchical fashion, where, at each level, one or more of EEG, EOG, and EMG are selected to best differentiate between two sleep stages. The best set of 100 features is chosen out of all the features derived from selected signals. The class imbalance in data is addressed by random undersampling and boosting techniques with decision trees as weak learners. Temporal context and data augmentation are used to improve the performance. We also evaluate the performance of our model by training and testing on different datasets. We compare the results of five approaches: using only EEG, EEG+EOG, EEG+EMG+EOG, EEG+EMG, and selective modality with a specific combination of EEG, EMG, and/or EOG at each level. The best results are obtained by considering features from EEG+EMG+EOG at each hierarchical level. The proposed model achieves average accuracies of 83.1%, 90.0%, 84.4%, 82.1%, 81.5%, 79.9%, and 73.7% on Sleep-EDF, Exp Sleep-EDF, ISRUC-S1, S2 and S3, DRMS-SUB, and DRMS-PAT datasets, respectively. For all the datasets except DRMS-SUB, the proposed method outperforms all the state-of-the-art approaches. Cross-dataset performance exceeds 80% for all datasets except DRMS-PAT; independent of whether the test data is from normal subjects or patients.

The Childfree Phenomenon in Indonesia: An Analysis of Sentiments on YouTube Video Comments

Childfree is a condition in which a person or couple decides not to have children in marriage. Childfree became popular in Indonesia when YouTuber and influencer Gita Savitri uploaded an Instagram story about it. This brought many pros and cons among the people towards the freedom to have children. Many TV broadcasts and YouTube videos cover this phenomenon. Several YouTube channels that broadcast this phenomenon are Menjadi Manusia and Analisa Channel. We collect YouTube comment data using web scraping techniques. From September 2021 to September 2022, 674 sample data points were obtained from two YouTube videos. Data is labelled (positive, negative, and neutral) using the Indonesian language lexicon approach as well as the Support Vector Machine (SVM) and Random Forest algorithms to determine the best model for classifying YouTube comments. The purpose of this research is to understand the public's perception of childfree and to compare the accuracy and AUC values of the two methods. Based on the results of the analysis, 128 comments are classified as positive, the remaining 39 comments are classified as neutral, and 503 comments are classified as negative. This shows that that the commentators on YouTube do not support or give a negative stigma to people who adhere to childfree. The solution to the balanced data problem for each sentiment class uses the random oversampling (ROS) approach. The RBF kernel SVM classification algorithm is a suitable method for classifying commentary data with an accuracy of 98.01% and an AUC of 98.58%, while the Random Forest algorithm only obtains an accuracy of 94.37% and an AUC of 95.87%.

B-ROS re-balanced learning method for PS-A-RoF FWA communication

The enhanced remote area communication (eRAC) scenario is an important growth point in the communication market. In some remote areas where optical fiber access cannot be realized or the laying cost is too high, fixed wireless access (FWA) is an appropriate supplementary solution for eRAC. Adopting analog radio over fiber (A-RoF) technology to implement FWA can overcome the bandwidth limitation of electronic devices and realize high-frequency carrier communication economically to achieve high-capacity wireless communication. Also, probabilistic shaping (PS) technology can be combined with A-RoF to further improve the flexibility of the network and coverage of service provision. However, in the PS-A-RoF network, the high RF power introduces more undesired nonlinear effects into the network, and it is often necessary to deploy supervised machine learning (ML) compensation modules in wireless receivers (WRs). But the module performances are affected by the uneven probability distribution of PS-QAM constellation points. In this paper, we employ the PS-A-RoF nonlinear model to theoretically investigate the correlation between the distribution of training symbols and the wireless A-RoF system’s performance. Our analysis reveals that reducing the variance of training symbol power contributes to a lower BER in the A-RoF network. We introduce a borderline random over-sampling (B-ROS) that matches with the PS-A-RoF nonlinear model, instead of the mainstream ROS, which is only at the data level. Based on the B-ROS scheme, only the minority examples below the borderline are over-sampled to reach a better variance performance. Introducing the B-ROS method into the supervised complex value nonlinear compensation module can further improve the decision accuracy of WRs with the restoration of phase information, without increasing additional computational resource consumption. The vector noise power, training symbol power variance, and noise factor metrics have been calculated to optimize the borderline value of our ML-based approach. We also present experimental data on the proof-of-concept A-RoF experiment for PS-64QAM. The results demonstrate a promising nonlinear compensation performance of the B-ROS WR, and the optimal borderline agrees well with the one deduced from the theoretical model under certain transmission conditions. Our proposed B-ROS scheme lessens the training size demand and can improve the receiver sensitivity by 0.51 dB compared to the common ML-based WR and by 0.7 dB compared to the conventional ROS scheme.

Impact of Feature Extraction Combined with Data Sampling Methods on Heartbeat Categorization

Dealing with class-imbalanced datasets in data analytics poses challenges, especially when faced with high-dimensional data. In order to handle this issue, researchers often utilize preprocessed methods like feature selection. Feature selection attempts to create a more informative and condensed feature set, while data sampling helps alleviate class imbalance. In our study, aim is to explore the effectiveness of data sampling preprocessed techniques combined with feature extraction using a dataset on ECG Heartbeat. We evaluate ensemble classifiers: Decision Tree; Random Forests (RF), Gradient-Boosted Trees (GBT) for feature extraction. In terms of data sampling, we assess the effectiveness of two methods: Random Under sampling (RUS) and Synthetic Minority Oversampling (SMOTE). The performance of this feature extraction is measured using the sensitivity and the specificity, two important metrics used for accuracy. Our findings depict that the combination of the RUS and GBT method yields the highest performance for ECG Heartbeat detection.

Feature Extraction of Risk Group and Electricity Theft by using Electrical Profiles and Physical Data for Classification in the Power Utilities

Non-technical loss (NTL) is one of the problems that has been a major issue in lost revenue for many years. Electricity distributors have attempted to reduce NTL by detecting electricity theft using various methods. Some events are difficult to detect that conventional meters inspection is inadequate. Moreover, many anomaly patterns found are very complex, confusing in identifying or distinguishing what types of electricity customers are at abnormal risk or energy theft that affects NTL. This paper proposes five key feature extraction methods and six classifying electricity customers using supervised learning. The main problem was studied and collected information, including kilowatt meters, electronic meters, TOU meters, and AMR meters, which cover four customer types that were recorded in the Provincial Electricity Authority (PEA) of Thailand. An electrical profile to be extracted for in-depth analysis of the behavior of each type of electricity customer, combined with the information of physical data to help enhance and increase efficiency. All features examined the relationships in each feature using Pearson correlation and handled unbalanced data using random oversampling (ROS). Then, the extracted data has been trained, validated, and tested to classify three classes: normal, risk, and theft, where we evaluate the results with performance metrics. The results show that random forest (RF) outperforms the rest of the classifiers by achieving a precision-recall area under the curve of 90% and a receiver operating characteristic curve of 78%. Significantly, the results were compared to previous studies and benchmark datasets, which revealed that the proposed method gave better results than other techniques.