Predicting Passenger Satisfaction

Introduction

Airline passenger satisfaction is one of the most consequential metrics an airline can track: it shapes repeat business, brand reputation, and customer lifetime value in an industry with notoriously thin margins. But satisfaction is shaped by dozens of interacting factors, from seat comfort to boarding process to flight delays, making it difficult to know where to focus improvement efforts without a data-driven approach.

This project analyzes a survey of 129,880 airline passengers to build a classification model that predicts whether a passenger will report themselves as satisfied or neutral/dissatisfied based on their demographic profile, flight details, and service experience ratings. The goal was twofold: build the most accurate predictive model possible, and demonstrate a complete, production-style machine learning workflow from raw data to a stakeholder-ready report.

The dataset includes 25 features per passenger, covering demographics (age, gender, customer type), flight details (travel class, distance, type of travel), delay times, and 14 individual service ratings on a 0-5 scale covering everything from inflight WiFi to baggage handling to cleanliness.

The Approach: Data Preparation

Cleaning a Real-World Survey Dataset

Survey data is rarely analysis-ready out of the box, and this dataset was no exception. The raw data included unnecessary index columns, missing values scattered across both numeric and categorical fields, and a small number of extreme outliers in the flight delay columns that could have skewed model training.

Preprocessing Pipeline

The preprocessing pipeline addressed each of these issues systematically:

• Outlier removal: Departure and arrival delay values above the 99th percentile were removed, since a handful of extreme delays (multi-day irregular operations, not typical travel experiences) would otherwise distort the model's understanding of a "normal" delayed flight.
• Missing value imputation: Numeric fields were filled with the median to preserve the underlying distribution, while categorical fields were filled with the most frequent category.
• Categorical encoding: Gender, Customer Type, Type of Travel, Class, and the satisfaction target were converted to numeric form using label encoding.
• Feature scaling: All features were standardized so that no single feature (e.g., flight distance, measured in hundreds of miles) would dominate model training simply due to its scale relative to features like individual service ratings (measured 0-5).
• Stratified splitting: The data was split 80/20 into training and test sets with stratification on the target variable, ensuring both sets preserved the same satisfied/dissatisfied ratio found in the full dataset.

Key Finding: Model Training & Tuning

Two Ensemble Approaches, Head to Head

Two ensemble learning algorithms were selected for this task: Random Forest, which builds many decision trees and averages their predictions, and XGBoost, a gradient boosting framework that builds trees sequentially, with each new tree correcting the errors of the ones before it. Both are strong choices for structured, tabular data like this survey dataset.

Hyperparameter Optimization

Rather than using default settings, both models were tuned using GridSearchCV with 3-fold cross-validation, systematically testing 27 hyperparameter combinations per model (81 fits each, 162 total):

• Random Forest: number of trees (50, 100, 200), maximum tree depth (10, 20, unrestricted), and minimum samples required to split a node (2, 5, 10)
• XGBoost: number of trees (50, 100, 200), learning rate (0.01, 0.1, 0.2), and maximum tree depth (3, 6, 9)

The best-performing configurations were 100 trees with unrestricted depth for Random Forest, and 200 trees with a learning rate of 0.1 and max depth of 9 for XGBoost.

What This Represents

On a held-out test set of 25,609 passengers the model had never seen during training, the tuned XGBoost model correctly classified 96% of passengers as satisfied or dissatisfied, with balanced precision and recall across both classes (0.97 precision / 0.94 recall for satisfied passengers). Random Forest performed nearly identically, trailing XGBoost by only a handful of misclassifications out of over 25,000 predictions. This kind of accuracy on real survey data, rather than a clean textbook dataset, reflects both the quality of the underlying signal in service ratings and the value of systematic hyperparameter tuning over default model settings.

Why This Matters: From Model to Business Value

A 96% accurate satisfaction classifier is more than an academic exercise. In practice, a model like this could be used to flag at-risk passengers in near real time based on their trip characteristics and in-flight service delivery, allowing an airline to intervene proactively (through service recovery gestures, loyalty outreach, or targeted follow-up) rather than only learning about dissatisfaction after the fact through churn or negative reviews.

Because the underlying model also exposes which features most influence its predictions, this type of analysis can help airlines prioritize operational investment: is it more valuable to reduce delays, improve boarding processes, or upgrade inflight entertainment? Quantifying passenger experience this precisely turns a subjective question into a data-driven resource allocation decision.

Limitations & Transparency

Like all analyses, this project has limitations worth understanding when interpreting the findings:

The Bottom Line

The Core Finding

Starting from a raw 130,000-row passenger survey with missing values, outliers, and mixed data types, a systematic preprocessing and hyperparameter tuning pipeline produced a model that correctly predicts passenger satisfaction 96% of the time. XGBoost narrowly outperformed Random Forest, but both ensemble approaches proved highly effective on this structured, tabular dataset.

Why the Process Matters as Much as the Result

The specific accuracy number is only part of the story. Equally important is the reproducible pipeline behind it: documented preprocessing decisions, systematic (rather than manual) hyperparameter search, stratified evaluation to avoid misleading results from class imbalance, and automated reporting so results can be communicated to non-technical stakeholders without re-running code.

Methodology & Data Notes