A Min-Max Based Data Normalization Method Robust to The Existence of Abnormal Data

In recent years, the rapid growth of big data utilization in Machine Learning applications has increased the crucial role of data preprocessing in enhancing the accuracy of algorithms. Among these preprocessing steps, data normalization stands out for its ability to standardize data ranges, ensuring all features contribute equally to the learning process. This article compares the effectiveness of two frequently used data normalization methods, Min-Max and Z-score, across various datasets. Recent studies have shown that Min-Max normalization often outperforms Z-score in specific applications. However, a significant drawback of the Min-Max method is its susceptibility to abnormal data, which can distort the entire normalization process. This is because Min-Max normalization relies solely on the minimum and maximum values of the dataset, which may include outliers among other data points. To address this issue, the article proposes an innovative adaptation of the Min-Max normalization method designed to enhance robustness against abnormal data. This approach aims to retain the advantages of Min-Max normalization while mitigating its vulnerabilities in real-world datasets. Results show that this method effectively handles both lower and upper abnormal values while ensuring reliable transformation of data into the specified range. This improvement is set to strengthen the dependability and usefulness of Min-Max normalization, in real-world machine learning scenarios.

کلیدواژه ها:

Abnormal data ، Normalization ، Min-Max ، SCADA