This project presents an explanatory and predictive analysis of airline passenger satisfaction using a large-scale customer dataset.
The objective is to identify the key factors influencing customer satisfaction and to develop classification models capable of predicting whether a passenger is satisfied or neutral/dissatisfied.
The project was implemented through a structured KNIME workflow integrating data preprocessing, feature engineering, dimensionality reduction, and multiple supervised learning models.
In the aviation industry, customer experience is a strategic asset.
Passenger satisfaction directly impacts:
- customer loyalty
- brand perception
- competitive positioning
- long-term revenue stability
Understanding which variables most strongly influence satisfaction allows airlines to implement targeted service improvements.
The dataset is sourced from Kaggle (Airlines Customer Satisfaction Dataset) and contains approximately 130,000 records.
The data were collected by a company with a fictitious name (“Invistico”) to preserve passenger privacy.
Flight-related variables
- Seat Comfort
- Departure/Arrival Time Convenience
- Gate Location
- Baggage Handling
- Check-in Service
- Online Boarding
- Ease of Online Booking
Service-related variables
- Food and Drink
- Inflight Wi-Fi Service
- Inflight Entertainment
- Inflight Service
- On-board Service
- Leg Room
- Cleanliness
Timing-related variables
- Departure Delay
- Arrival Delay
- Flight Distance
Passenger characteristics
- Gender
- Customer Type (Loyal / Disloyal)
- Age
- Type of Travel (Business / Personal)
- Class (Business / Eco / Eco Plus)
Target variable
- Satisfaction (Satisfied / Neutral-Dissatisfied)
Key preprocessing steps included:
- Replacement of “0” values in ordinal features with missing values
- Conditional mean imputation based on Satisfaction class
- Feature engineering:
- In-flight Delay (Arrival - Departure delay)
- Arrival Delay per Distance ratio
- Removal of Arrival Delay due to high correlation with Departure Delay (ρ ≈ 0.97)
- Train/Test split (75% / 25%)
- Principal Component Analysis (PCA) for feature structure evaluation
Despite PCA explaining 86% of variance with 8 components, original features were retained to preserve interpretability.
The following supervised classification models were trained and compared:
- Multilayer Perceptron (MLP)
- K-Nearest Neighbors (KNN)
- Logistic Regression
- Random Forest
| Model | Accuracy |
|---|---|
| Multilayer Perceptron | 0.925 |
| K-Nearest Neighbors | 0.914 |
| Logistic Regression | 0.892 |
| Random Forest | 0.965 |
The Random Forest classifier achieved the best overall performance, with:
- Accuracy: 0.965
- AUC: 0.995
- High robustness and computational efficiency
Permutation-based feature importance analysis highlighted the most influential variables:
- Customer Type
- In-flight Wi-Fi Service
- Online Boarding
- Type of Travel
Key insight: Passengers traveling for personal reasons and first-time customers show significantly higher dissatisfaction rates.
From a strategic standpoint, digital service quality (Wi-Fi, online boarding, booking experience) appears to be a decisive factor.
The KNIME workflow includes:
- Data cleaning and preprocessing
- Feature engineering
- Train/test splitting
- Model training and evaluation
- Cross-validation comparison
- Global feature importance analysis
- Random Forest provides the best trade-off between interpretability, accuracy, and computational efficiency.
- KNN shows competitive accuracy but high computational cost.
- Logistic Regression underperforms compared to tree-based models.
- Digital service-related variables are stronger predictors than many physical service variables.
Airlines should prioritize:
- Investment in high-quality in-flight Wi-Fi
- Optimization of online boarding and booking systems
- Targeted loyalty strategies for first-time customers
- Segmented service design based on type of travel
Daniele Lepre
Alice Anna Maria Brunazzi
Alessandro Della Beffa