About

Processing

Our analysis began with a structured data preparation pipeline. The first step involved cleaning the dataset by removing unnecessary or redundant features. For example, we dropped “Departure Delay” because it almost perfectly correlated with “Arrival Delay,” which helped reduce noise and improve model performance. Categorical variables such as “Class,” “Customer Type,” and “Type of Travel” were encoded using label encoders so they could be interpreted numerically by machine learning algorithms. Continuous variables were standardized using z-score normalization to ensure equal weight across features during training.

We then split the data into training, validation, and test sets to establish a reliable framework for model evaluation. This allowed us to fine-tune hyperparameters and compare performance across models with minimal risk of overfitting. Our team tested several classification algorithms including Decision Trees, AdaBoost, XGBoost, and Random Forest. After comparing performance metrics, Random Forest was selected as the final model due to its balance of high accuracy and interpretability. The model also provided a ranked list of feature importances, which allowed us to identify the most influential factors contributing to passenger satisfaction. These included Onboard Service, Seat Comfort, Leg Room, Cleanliness, and Inflight Wi-Fi. Each of these variables consistently appeared as top contributors across multiple runs.

During development, we encountered challenges related to class imbalance, as a majority of passengers in the dataset were labeled “Satisfied.” To address this, we used stratified sampling during data splitting to preserve the original class proportions. We also observed overlapping or collinear service categories that risked inflating the influence of related features.

We resolved this by conducting exploratory analysis and removing or consolidating highly correlated inputs. All preprocessing and modeling were completed in Python using pandas, sci-kit learn, and other supporting libraries. The result was a reproducible, efficient pipeline that enabled us to translate raw survey data from airlines into meaningful, interpretable insights about what matters most in the flight passenger experience.