DATA HANDLING AND PROCESSING IN MACHINE LEARNING

Since data preprocessing forms an important aspect of the process of developing machine learning models, the quality and representation of the data have a direct impact on the performance of the model. Real-world datasets usually include missing values, duplicate records, inconsistent entries, categorical variables, numerical features on different scales, irrelevant attributes, and variables which could cause data leakage. It is therefore necessary to deal with these problems if reliable and efficient predictive models are to be built. The study looks at the way a series of preprocessing steps affect the performance of machine learning models, using the IBM Telco Customer Churn dataset and considering customer churn as a binary classification problem. It involves the creation of four preprocessing pipelines, starting with data cleaning, then moving on to categorical encoding, numerical feature scaling, and finally feature selection. In order to compare the impact of these preprocessing stages under the same experimental conditions, Logistic Regression, Decision Tree, and Random Forest classifiers are employed. The evaluation of the models includes the use of Accuracy, Precision, Recall, and the F1-score, as well as the preprocessing time and the time taken to train the models. The study also places special emphasis on spotting the attributes which might lead to data leakage and on investigating whether reducing the number of features has an impact on model performance. Instead of assuming that a single preprocessing method is appropriate for all machine learning algorithms, the study compares the three models in order to see how their reactions to different preprocessing techniques differ.