Communications on Applied Nonlinear Analysis ISSN: 1074-133X Vol 32 No. 9s (2025) 1125 https://internationalpubls.com Rank based RELIEFG Method for Spam Mail Detection and Classification B. Aruna Kumari1, C. Nagaraju2 1Research Scholar, YSR Engineering College of Yogi Vemana University, Computer Science & Engineering, Proddatur, arunakumarib1421@gmail.com 2Professor, YSR Engineering College of Yogi Vemana University, Computer Science & Engineering, Proddatur, nagaraju.c@yvu.edu.in Article History: Received: 12-01-2025 Revised: 15-02-2025 Accepted: 01-03-2025 Abstract: Email plays a crucial role in day-to-day communication. An increase in spam emails poses a significant threat, leading to financial and resource losses. It's not easy to differentiate spam emails from legitimate emails due to large no.of features in datasets. Developing effective spam detection method is crucial for email security with missing values, outliers and noise. In the literature many methods have been existed among those rank-based Relief method is suitable for detection and classification. However, the relief method considers one hit and one miss. Due to this reason, it fails for multiclass classification. Due to the importance of Relief method, researchers have made many extensions to Relief. Among those ReliefF is the one of the latest variants of Relief method which improves accuracy by reducing the computational cost in the presence of missing values and noise for multiclass classifications. However, it fails in presentce of outliers. In this paper, a new variant ReliefG is proposed. This method reduces the computational cost by finding the highest rank of attributes and generating crisp set values by integrating efficient fuzzy membership function into a sigmoid function to improve the accuracy in the presence of outliers. Keywords: Crisp set, Rank-based methods, Relief, Iterative Relief, ReliefF. 1. Introduction Communication has undergone significant evolution in recent years, presenting various methods for individuals to share ideas and thoughts. From traditional in-person conversations, phone calls, and letters to modern online communication, methods have become more diverse and cost – effective. The widespread use of mobile phones has also drawn the attention of cyber-criminals, who misuse technology for malicious intent [1]. An often-used strategy involves sending unsolicited electronic messages, commonly known as spam. Attackers can use these messages to deceive recipients and lead them to respond and potentially face additional charges or simply promote offers and products. Spam messages are sent as text, images, and multimedia through mobile text messages, emails, and social media. Email or electronic mail, is an essential and widely used communication tool on the internet. Harness the power to exchange messages with anyone possessing an email address, no matter their location across the globe. Email relies on various protocols within the TCP/IP suite and has long been a vital medium of communication, offering near instant global connectivity. Email spam is a major issue in today’s digital age, creating challenges for organizations, businesses and individuals [2]. Spam emails are unwanted messages that fill up inboxes, wasting precious time and resources, all while risking the exposure of users to harmful content or scams. Machine learning techniques have become mailto:arunakumarib1421@gmail.com mailto:cnrcse@yahoo.com Communications on Applied Nonlinear Analysis ISSN: 1074-133X Vol 32 No. 9s (2025) 1126 https://internationalpubls.com powerful tools for detecting email spam, addressing this issue. Email spam detection aims to classify incoming emails as legitimate or spam. Traditional rule-based methods are less effective due to evolving spam. Machine learning offers a more adaptable approach using patterns and features from large email datasets. 2. Literature Survey Author, Year Aim Method Results Future Scope Verónica Bolón- Canedo, et.al., [3], 2012 To assess feature selection methods with increasing irrelevant features, data noise, and attribute interactions. Correlation-based Feature Selection, INTERACT algorithm, Consistency-based Filter, ReliefF, Information gain. SVM – RFE, FS-P Embedded methods and WrapperSubsetEval Methods are used and compared. ReliefF filters offer independence from the induction algorithm and outpace embedded and wrapper methods in terms of speed. In future work, this procedure could be further developed to tackle regression problems or be utilized for image analysis. Girish Chandrashekar, Ferat Sahin, [4], 2014 To offer a general introduction to variable elimination that can be used for a wide range of machine learning problems. Filter methods, Wrapper methods, Embedded methods are used and compared. Feature selection provides benefits such as gaining insight into the data, improving classifier models, enhancing generalization, and identifying irrelevant variables. Future research must prioritize hybrid methods that combine multiple feature selection methods to maximize accuracy, computational efficiency, and stability. This approach is essential for overcoming individual weaknesses through the combination of algorithms. Newton Spolaor, et.al., [5], 2012. By evaluating each feature separately, we can gain a comprehensive understanding of their RF and IG-ML are used The two feature selection methods were assessed across ten benchmark datasets, demonstrating In the future, it will be possible to expand the concepts of ReliefF for multi-label data. Communications on Applied Nonlinear Analysis ISSN: 1074-133X Vol 32 No. 9s (2025) 1127 https://internationalpubls.com individual impact. highly promising results. Hidayet Tacki, Fatema Nusrat [6], 2023. To introduce a machine learning – based model that will filters spam emails. C4.5, ID3, RndTree, C-SVC, Naïve Bayes algorithms are used. Various feature selection algorithms like Forward selection, ReliefF and Fisher Filtering was used and compared. In the comparison of studies for accuracy, the C-SVC algorithm proved to be the most successful. It achieved the highest correct recognition rate of 93.13%, outperforming the RndTree algorithm. In experiments involving feature selection, the forward selection method yielded superior results. Additionally, binary transformation demonstrated significantly better success in data transformation. Forward selection produced good results; however, future research should explore advanced feature selection methods like RFE, L1 regularization (Lasso), or genetic algorithms to optimize processing time and accuracy, especially with larger datasets. Panem Charanarur, et.al., [7], 2023. The focal point of this project is to significantly enhance the efficiency of KNN, Naïve Bayes, ETC, Random Forest, SVC, AdaBoost, xgboost, LR, GBDT, BgC, The KN and Naive Bayes algorithms outperform than other algorithms, by Consider deep learning models like RNNs or Transformers, especially pre-trained ones like BERT, for Communications on Applied Nonlinear Analysis ISSN: 1074-133X Vol 32 No. 9s (2025) 1128 https://internationalpubls.com spam email detection and management. Decision Tree are used and compared. achieving the highest accuracy and precision scores. more accurate email spam detection. Dr. Purva Mange, Aditi Lule, Rohini Savant, [8], 2023. The aim is to detect and mitigate spam emails precisely. PSO (Particle Swarm Optimisation), Genetic algorithm and the hybrid Genetic algorithm + PSO methods was used and compared. The hybrid model produced lower false positive rates and superior accuracy compared to conventional spam detection techniques. In the future, this hybrid model could be expanded to cover additional domains such as phishing attempts and fraudulent transactions. 3. Relief The original Relief algorithm is based on the principles of instance-based learning, making it a powerful and effective approach. Relief assesses each feature by calculating a proxy statistic, making it an effective method for individual feature evaluation and selection. This statistic is crucial for determining the feature's 'quality' or 'relevance' to the target concept [9]. Remember, these statistics represent feature weights or feature scores, which range from worst to best i.e., from -1 to +1. The Relief algorithm could only handle binary classification problems and did not account for missing data. Due to this, it was limited in addressing real-world datasets with missing values or multiclass classification problems. Updating the weight vector by using following formula: Wi = Wi – (xi – nHi) 2 + (xi – nMi) 2 Eq (1) In the above equation, i represents the components, nH denotes nearHit, and nM specifies nearMiss 3.1 Variants of Relief The extended version, Relief – A, averages the contribution of k nearest hits/misses for the random instance. So that noise and redundancy in the data are reduced, leading to improved accuracy. The Relief – B method is an extension to Relief -A, it can handle incomplete data, specifically missing values. This method is limited to handling binary class problems. It is designed to manage data with missing values for a particular attribute of instances. When calculating the diff function, it does not take into account instances with missing values. The Relief-C method is same as Relief-B. It can handle data with missing values for a given attribute of instances by ignoring the instances with missing values for evaluation. This method can also handle two-class problems. The Relief – D is an extension to Relief – C method and it evaluates neighbours only if their attribute values differ from those of a random instance. Three innovative strategies proposed to tackle incomplete data in Relief such as Communications on Applied Nonlinear Analysis ISSN: 1074-133X Vol 32 No. 9s (2025) 1129 https://internationalpubls.com ReliefB, ReliefC, and ReliefD. When missing value is encountered, Relief – D sets the "diff" function to the class-conditional probability of different feature values between two instances, but these three strategies are limited to binary class problems. To overcome this limitation, the Relief – E method was proposed to handle multi-class endpoints. It employs membership generalization over multi-class neighbours and randomly selects 'k' hits and 'k' misses for each instance. Recently two more powerful methods iterative relief and ReliefF extended and which are described below. 3.2 Iterative relief An iterative approach for Relief, denoted as Iterative Relief. It is an extension to relief and it’s variant algorithms. At the outset, all features are assigned equal weights. Any negative feature weights are treated as outliers and are rounded up to zero. The algorithm iterates until the weights converge. Iterative Relief requires multiple iterations for convergence, resulting in significantly higher computational costs are increased when dealing with large datasets. So to overcome this drawback than ReliefF method is derived. 3.3 ReliefF ReliefF is a significant improvement of the original Relief and iterative relief algorithms. ReliefF selects multiple neighbors to enhance robustness and identifies discriminative features between spam and non-spam emails [10]. ReliefF seeks the k nearest hits and misses, averaging their impact on feature weights. The value of k can be adjusted to suit each specific problem. The weight measure of a feature ‘A’ is represented as W[A] as follows: w[A] = w[A] − 1 m−k ∑ (𝑘 𝑗=1 diff (A, xi , Hitj ) ) + 1 m−k ∑ (𝑘 𝑗=1 𝑃(𝐶)− 𝑑𝑖𝑓𝑓(𝐴,xi,Missj,c) 1−𝑃(𝐶) ) Eq (2) In Eq(2), diff (A, 𝑥, Hitj) denotes distance function calculating the difference between the values of feature A with hit or miss operations. The j-th nearest hit is represented by Hitj, j-th nearest miss from class 𝐶 is denoted by Miss𝑗, 𝐶 and prior probability of class 𝐶 is shown by P(C). ReliefF's effectiveness is compromised when dealing with highly correlated or redundant features. ReliefF may not perform well with outliers. ReliefF method handles multiclass classification, multi label classification very effectively, when the number of attributes is very low (for example, just one or two), the algorithm may struggle to identify meaningful patterns because it lacks sufficient discriminative power. On the other hand, when the number of attributes is very high, ReliefF can become computationally expensive and less reliable due to the curse of dimensionality, which may lead to inaccurate results. So, in this paper, we are proposing a new variant i.e., ReliefG which will work effectively with large datasets as well as outliers. 4. Proposed Method The ReliefG is proposed. The no. of hits and no.of misses are automatically selected by using fuzzy membership function. The crisp set values are evaluated by integrating trapezoidal membership function into sigmoid function such that we can eliminate the outliers. Because the sigmoid function can work by enhancing lower values and reducing the highest value to lower value. Simultaneously it reduces the computational costs. It gives the highest accurate results evem in presence of more attributes and it automatically selects the k value. Communications on Applied Nonlinear Analysis ISSN: 1074-133X Vol 32 No. 9s (2025) 1130 https://internationalpubls.com 4.1 Sigmoid function and its characteristics The sigmoid membership function is an essential tool for smoothly transitioning between different degrees of membership. Its versatile nature makes it a valuable asset in various applications. 𝜇(𝑥) = 1 1+𝑒^(−𝑎(𝑥−𝑐)) Eq (3) In Eq (3), the 𝜇(𝑥) represents the membership degree of x in the fuzzy set, ranging from 0 to 1. Here, x is the input variable, ‘a’ controls the slope of the function, and c is the threshold value. The sigmoid function creates an 'S' curve that moves from 0 to 1. The slope of the curve is controlled by the parameter "a". When a > 0, the function increases from 0 to 1 as x increases; when a < 0, the function decreases from 1 to 0 as x increases. The sigmoid function's non-linear nature empowers it to adeptly navigate intricate relationships between variables within fuzzy systems. To effectively eliminate outliers, we are integrating the power of the sigmoid function with the trapezoidal function. This innovative approach will ensure a more robust and accurate analysis. 5. Dataset Description The SPAM dataset consists of 1813 spam mails and 2788 ham emails. Among 4601 instances, there are 57 attributes, consisting of frequently occurring words or characters. The first 48 attributes are continuous real numbers, ranging from 0 to 100, labeled as "WORD." This indicates the percentage of words present in the email. The next six attributes are continuous real numbers, ranging from 0 to 100, labeled as "char_freq_CHAR," indicating the percentage of characters in the email. The remaining attributes include a mix of continuous real numbers and integers. The final column indicates whether the email is spam (denoted as "1") or ham (denoted as "0"). Fig 1. Actual SPAM dataset Communications on Applied Nonlinear Analysis ISSN: 1074-133X Vol 32 No. 9s (2025) 1131 https://internationalpubls.com Fig2. SPAM dataset with noisy data Fig 3. SPAM dataset with outliers Communications on Applied Nonlinear Analysis ISSN: 1074-133X Vol 32 No. 9s (2025) 1132 https://internationalpubls.com Fig 4. SPAM dataset with data preprocessing For experimentation, here we considered SPAM dataset as shown in Fig 1, SPAM dataset with noisy data is represented in Fig 2. The SPAM dataset with outliers is denoted in Fig 3. Finally, the SPAM dataset with data preprocessing is shown in Fig 4. 6. Experimental Results The proposed method has been implemented and tested on the SPAM dataset. The database contains 57 features utilized to differentiate between two types of emails, comprising specific words or characters that frequently appear in an email. In this paper, a confusion matrix is used to evaluate the method's performance. The matrix consists of four outcomes: TN, TP, FP, and FN [12] Statistical parameters such as precision, recall, accuracy, and f1-score are computed from the confusion matrix to analyze performance using tables and graphs. • Accuracy: Accuracy specifies no.of accurate predictions were made regarding the total number of values. Accuracy = truepositive+truenegative (truepositive+falsepositive+truenegative+falsenegative) Eq (4) • Precision: Precision is the key to accurately anticipating true positive outcomes. Precision = truepositive (truepositive+falsepositive) Eq (5) • Recall: Utilized for obtaining accurate negative values. Recall = truenegative (truepositive+falsenegative) Eq (6) Communications on Applied Nonlinear Analysis ISSN: 1074-133X Vol 32 No. 9s (2025) 1133 https://internationalpubls.com • F1 Score: It is used for classifying dataset values as positive or negative. 𝐹1 − 𝑠𝑐𝑜𝑟𝑒 = 2∗𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛∗𝑅𝑒𝑐𝑎𝑙𝑙 𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛∗ 𝑟𝑒𝑐𝑎𝑙𝑙 Eq (7) Table 1. Accuracy Datasets Relief ReliefF Iterative Relief ReliefG Quality data 94.17897 94.52649 94.72649 95.04778 data filling with mean 94.61337 94.09209 94.17897 94.9609 noisy data 93.83145 94.30026 94.71833 94.87402 data with outliers 94.0609 94.23961 94.46585 94.52649 data with missing values elimination 95.13466 95.12218 95.13466 95.44778 Fig 5. Accuracy Fig 6. Accuracy In above Fig 5 and Fig 6, thick blue bar & line represents accuracy of quality data for all variants of Relief. Orange colored bar & line denotes accuracy of data after preprocessing for all relief methods. Grey colored results shown that the accuracy of noisy data for relief and it’s variants. Accuracy of dataset with outliers is represented by yellow colored bar and line. Finally the light blue colored bar, line represents the accuracy of dataset with missing values elimination. Table 2. Precision Type of data Relief ReliefF Iterative Relief ReliefG Quality data 94.54148 94.7826 94.88874 96.44444 data with Preprocessing 94.9891 94.14316 94.54148 96.43652 noisy data 94.69026 94.4206 93.92624 96.42857 93 93.5 94 94.5 95 95.5 96 Relief ReliefF Iterative Relief ReliefG 93 93.5 94 94.5 95 95.5 96 Relief ReliefF Iterative Relief ReliefG Communications on Applied Nonlinear Analysis ISSN: 1074-133X Vol 32 No. 9s (2025) 1134 https://internationalpubls.com data with outliers 94.85393 94.77124 94.55337 96.18834 data with missing values elimination 95.84245 95.87852 95.84245 95.95478 Fig 7. Precision Fig 8. Precision In above Fig 7 and Fig 8, thick blue bar & line represents precision values of quality data for all variants of Relief. Orange colored bar & line denotes precision values of data after preprocessing for all relief methods. Grey colored results shown that the precision values of noisy data for relief and it’s variants. precision values of dataset with outliers is represented by yellow colored bar and line. Finally the light blue colored bar, line represents the precision values of dataset with missing values elimination. Table 3. Recall Type of data Relief ReliefF Iterative Relief ReliefG Quality data 91.15789 91.78947 92 92.36842 data with Preprocessing 91.08947 91.36842 91.15789 92.15789 noisy data 90.10526 91.03157 91.15789 91.94736 data with outliers 90.73684 91.57894 91.36842 91.61578 data with missing values elimination 90.21052 91.05263 92.21052 92.42105 92.5 93 93.5 94 94.5 95 95.5 96 96.5 97 Relief ReliefF Iterative Relief ReliefG 92.5 93 93.5 94 94.5 95 95.5 96 96.5 97 Relief ReliefF Iterative Relief ReliefG Communications on Applied Nonlinear Analysis ISSN: 1074-133X Vol 32 No. 9s (2025) 1135 https://internationalpubls.com Fig 9. Recall Fig 10. Recall In above Fig 9 and Fig 10, thick blue bar & line represents recall values of quality data for all variants of Relief. Orange colored bar & line denotes recall values of data after preprocessing for all relief methods. Grey colored results shown that the recall values of noisy data for relief and it’s variants. recall values of dataset with outliers is represented by yellow colored bar and line. Finally the light blue colored bar, line represents the recall values of dataset with missing values elimination. Table 4. F1-Score Type of data Relief ReliefF Iterative Relief ReliefG Quality data 92.81886 93.26203 93.27641 93.83783 data with Preprocessing 93.36188 92.73504 92.81886 93.72294 noisy data 92.34088 93.51753 92.52136 93.6078 data with outliers 93.69565 93.14775 92.93361 93.1596 data with missing values elimination 93.99141 94.44444 93.99141 94.90374 Fig 11. F1-Score Fig 12. F1-Score 88.5 89 89.5 90 90.5 91 91.5 92 92.5 93 Relief ReliefF Iterative Relief ReliefG 88.5 89 89.5 90 90.5 91 91.5 92 92.5 93 Relief ReliefF Iterative Relief ReliefG 91 92 93 94 95 96 Relief ReliefF Iterative Relief ReliefG 91 92 93 94 95 96 Relief ReliefF Iterative Relief ReliefG Communications on Applied Nonlinear Analysis ISSN: 1074-133X Vol 32 No. 9s (2025) 1136 https://internationalpubls.com In above Fig 11 and Fig 12, thick blue bar & line represents F1-Score of quality data for all variants of Relief. Orange colored bar & line denotes F1-Score of data after preprocessing for all relief methods. Grey colored results shown that the F1-Score of noisy data for relief and it’s variants. F1-Score of dataset with outliers is represented by yellow colored bar and line. Finally the light blue colored bar, line represents the F1-Score of dataset with missing values elimination. Table 5. Mean Absolute Error Type of data Relief ReliefF Iterative Relief ReliefG Quality data 0.058 0.054 0.054 0.049 data with Preprocessing 0.053 0.059 0.058 0.05 noisy data 0.061 0.052 0.06 0.051 data with outliers 0.05 0.055 0.057 0.054 data with missing values elimination 0.048 0.045 0.048 0.049 Fig 13. Mean Absolute Error Fig 14. Mean Absolute Error In above Fig 13 and Fig 14, thick blue bar & line represents MAE values of quality data for all variants of Relief. Orange colored bar & line denotes MAE values of data after preprocessing for all relief methods. Grey colored results shown that the MAE values of noisy data for relief and it’s variants. MAE values of dataset with outliers is represented by yellow colored bar and line. Finally the light blue colored bar, line represents the MAE values of dataset with missing values elimination. Table 6. Mean Squared Error Type of data Relief ReliefF Iterative Relief ReliefG Quality data 0.058 0.054 0.054 0.049 0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 Relief ReliefF Iterative Relief ReliefG 0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 Relief ReliefF Iterative Relief ReliefG Communications on Applied Nonlinear Analysis ISSN: 1074-133X Vol 32 No. 9s (2025) 1137 https://internationalpubls.com data with Preprocessing 0.053 0.059 0.058 0.05 noisy data 0.061 0.052 0.06 0.051 data with outliers 0.05 0.055 0.057 0.054 data with missing values elimination 0.048 0.045 0.048 0.049 Fig 15. Mean Squared Error Fig 16. Mean Squared Error In above Fig 15 and Fig 16, thick blue bar & line represents MSE values of quality data for all variants of Relief. Orange colored bar & line denotes MSE values of data after preprocessing for all relief methods. Grey colored results shown that the MSE values of noisy data for relief and it’s variants. MSE values of dataset with outliers is represented by yellow colored bar and line. Finally the light blue colored bar, line represents the MSE values of dataset with missing values elimination. Table 7. Root Mean Squared Error Type of data Relief ReliefF Iterative Relief ReliefG Quality data 0.241 0.233 0.233 0.222 data with Preprocessing 0.232 0.243 0.241 0.224 noisy data 0.248 0.23 0.246 0.226 data with outliers 0.224 0.235 0.239 0.233 data with missing values elimination 0.22 0.212 0.22 0.222 0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 Relief ReliefF Iterative Relief ReliefG 0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 Relief ReliefF Iterative Relief ReliefG Communications on Applied Nonlinear Analysis ISSN: 1074-133X Vol 32 No. 9s (2025) 1138 https://internationalpubls.com Fig 17. Root Mean Squared Error Fig 18. Root Mean Squared Error In above Fig 17 and Fig 18, thick blue bar & line represents RMSE values of quality data for all variants of Relief. Orange colored bar & line denotes RMSE values of data after preprocessing for all relief methods. Grey colored results shown that the RMSE values of noisy data for relief and it’s variants. RMSE values of dataset with outliers is represented by yellow colored bar and line. Finally the light blue colored bar, line represents the RMSE values of dataset with missing values elimination. Table 8. t-test values Type of data Relief ReliefF Iterative Relief ReliefG Quality data 0.72 0.63 0.55 1.06 data with Preprocessing 0.67 0.59 0.72 1.1 noisy data 0.97 0.38 0.59 1.14 data with outliers 1.07 0.67 0.67 1.23 data with missing values elimination 0.76 0.59 0.76 1.43 Fig 19. t-test values Fig 20. t-test values In above Fig 19 and Fig 20, thick blue bar & line represents t-test values of quality data for all variants of Relief. Orange colored bar & line denotes t-test values of data after preprocessing for all relief methods. Grey colored results shown that the t-test values of noisy data for relief and it’s variants. t- 0.19 0.2 0.21 0.22 0.23 0.24 0.25 0.26 Relief ReliefF Iterative Relief ReliefG 0.19 0.2 0.21 0.22 0.23 0.24 0.25 0.26 Relief ReliefF Iterative Relief ReliefG 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 Relief ReliefF Iterative Relief ReliefG 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 Relief ReliefF Iterative Relief ReliefG Communications on Applied Nonlinear Analysis ISSN: 1074-133X Vol 32 No. 9s (2025) 1139 https://internationalpubls.com test values of dataset with outliers is represented by yellow colored bar and line. Finally the light blue colored bar, line represents the t-test values of dataset with missing values elimination. Table 9. p-test values Type of data Relief ReliefF Iterative Relief ReliefG Quality data 0.47 0.52 0.58 0.28 data with Preprocessing 0.49 0.55 0.47 0.26 noisy data 0.32 0.7 0.55 0.25 data with outliers 0.2 0.49 0.49 0.21 data with missing values elimination 0.44 0.55 0.44 0.52 Fig 21. p-test values Fig 22. p-test values In above Fig 21 and Fig 22, thick blue bar & line represents p-test of quality data for all variants of Relief. Orange colored bar & line denotes p-test values of data after preprocessing for all relief methods. Grey colored results shown that the p-test values of noisy data for relief and it’s variants. p- test values of dataset with outliers is represented by yellow colored bar and line. Finally the light blue colored bar, line represents the p-test values of dataset with missing values elimination. Table 10. f-test values Type of data Relief ReliefF Iterative Relief ReliefG Quality data 0.98 0.98 0.99 0.92 data with Preprocessing 0.98 0.99 0.98 0.94 noisy data 0.98 0.99 0.99 0.97 data with outliers 0.97 0.98 0.98 0.97 data with missing values elimination 0.98 0.99 0.98 0.98 0 0.2 0.4 0.6 0.8 Relief ReliefF Iterative Relief ReliefG 0 0.2 0.4 0.6 0.8 Relief ReliefF Iterative Relief ReliefG Communications on Applied Nonlinear Analysis ISSN: 1074-133X Vol 32 No. 9s (2025) 1140 https://internationalpubls.com Fig 23. f-test values Fig 24. f-test values In above Fig 23 and Fig 24, thick blue bar & line represents f-test values of quality data for all variants of Relief. Orange colored bar & line denotes f-test values of data after preprocessing for all relief methods. Grey colored results shown that the f-test values of noisy data for relief and it’s variants. F- test values of dataset with outliers is represented by yellow colored bar and line. Finally the light blue colored bar, line represents the f-test values of dataset with missing values elimination. 7. Conclusions This paper suggests an innovative approach for improving spam detection by utilizing sophisticated feature selection method ReliefG. This method builds on the principles of original Relief algorithm, enhancing its ability to handle various challenges like noise, missing values and outliers. The subsequent phase involves utilizing Random Forest (RF) for classification, leveraging the chosen features. The ReliefG method produced better accuracy over Relief, ReliefF and Iterative relief methods. However, the method reduces computational cost by identifying key attributes and generating crisp set values. This is done by integrating an efficient fuzzy membership function into a sigmoid function to improve accuracy, especially in the presence of outliers. References [1] Ahmed AI – Ajeli, Raaid Alubady, Eman S. AI – Shamery (2020), Improving Spam Email Detection using Hybrid Feature Selection and Sequential Minimal Optimisation, Indonesian Journal of Electrical Engineering and Computer Science, Vol. 19, No. 1, PP: 535-542. [2] Mrs. Anitha Reddy, et.al., (2023), Email Spam Detection Using Machine Learning, Journal of Survey in Fisheries Sciences, Vol. 10, No. 1, PP: 2658-2664. [3] Veronica Bolon-Canedo, et.al., (2012), A Review of Feature Selection Methods on Synthetic data, Knowledge and Information Systems, Springer. [4] Girish Chandrashekar, Ferat Sahin, (2014), A Survey on Feature Selection Methods, Computers and Electrical Engineering, Vol. 40, PP: 16-28. [5] Newton Spolaor, et.al., (2012), Filter Approach Feature Selection Methods to Support Multi- label Learning Based on ReliefF and Information Gain, Springer, PP: 72-81. [6] Hidayet Tacki, Fatema Nusrat (2023), Highly Accurate Spam Detection with the Help of Feature Selection and Data Transformation, The International Arab Journal of Information Technology, Vol. 20, No.1, PP: 29-37. 0.88 0.9 0.92 0.94 0.96 0.98 1 Relief ReliefF Iterative Relief ReliefG 0.88 0.9 0.92 0.94 0.96 0.98 1 Relief ReliefF Iterative Relief ReliefG Communications on Applied Nonlinear Analysis ISSN: 1074-133X Vol 32 No. 9s (2025) 1141 https://internationalpubls.com [7] Panem Charanarur, et.al., (2023), Machine – Learning – Based Spam Mail Detector, SN Computer Science, https://doi.org/10.1007/s42979-023-02330-x. [8] Dr. Purva Mange, Aditi Lule, Rohini Savant, (2024), Advanced Spam Email Detection using Machine Learning and Bio-Inspired Meta-Heuristics Algorithms, International Journal of Intelligent Systems and Applications in Engineering, Vol. 12, No. 4, PP: 122-135. [9] Ryan J. Urbanowicz, et.al., (2018), Relief – Based Feature Selection: Introduction and Reviews, Journal of Biomedical Informatics. [10] B. Aruna Kumari, et.al., (2024), MultiSURF: Optimal Feature Selection Technique for Spam Mail Detection and Classification, Nanotechnology Perceptions, Vol. 20, No. S8, PP: 452-461. [11] Bruce A. Draper, et.al., (2003), Iterative Relief, CVPR, IEEE Computer Society Conference on Computer Vision and Pattern Recognition, IEEE Explore, DOI: 10.1109/CVPRW.2003.10065. [12] C. Nagaraju, et.al., (2024), A Generalized Two-Level Ensemble Method for Spam Mail Detection, Journal of Electrical Systems, Vol. 20, No. 2, PP: 1570-1579. https://doi.org/10.1007/s42979-023-02330-x