Data Science Salary Prediction: A Comparative Study of Standard Regressors and Leakage Testing

This study presents a comparative analysis of standard regression models for predicting data science salaries, with a specific focus on identifying and testing for data leakage—a frequent cause of over-optimistic results in applied data mining.The dataset comprises 3,761 records describing data science positions, featuring variables such as work year, experience level (entry, mid, senior, executive), employment type, job title, company location, company size, and salary converted to US dollars.The methodology involves preprocessing the data through one-hot encoding of categorical features and evaluating the models under different experimental settings to explicitly test for target leakage. The study highlights that the dataset is heavily skewed toward senior, full-time, and US-based positions, which may affect the reliability of predictions for other demographics.Key contributions include:A side-by-side comparison of standard regressors using default hyperparameters.An explicit evaluation of data leakage in the prediction pipeline.An exploratory analysis of the dataset's characteristics and feature distributions.The study also outlines future work, which includes grouping similar job titles into role families, applying a log transformation to the skewed target variable, tuning hyperparameters with cross-validation, evaluating gradient boosting libraries such as XGBoost, and adding features such as remote-work ratio, industry, and required skills.