AIRLINE PASSANGER SATISFACTION CLASSIFICATION

Contents
Abstract..............................................................................................................3
Introduction.........................................................................................................3
Data Mining Theory...............................................................................................4
Decision Tree Classifier.......................................................................................4
Logistic Regression............................................................................................4
Evaluation Metrics.............................................................................................5
Data Exploration and Preparation..............................................................................5
Overview of the Dataset.......................................................................................5
Initial Observations............................................................................................5
Range and Central Tendency..............................................................................5
Skewness and Kurtosis.....................................................................................6
Data Cleaning...................................................................................................6
Missing Values...............................................................................................6
Feature Engineering............................................................................................6
Encoding Categorical Variables..........................................................................6
Normalization................................................................................................6
Experimental Setup................................................................................................6
Data Partitioning................................................................................................6
Model Implementation........................................................................................7
Decision Tree Model........................................................................................7
Logistic Regression Model................................................................................7
Validation Approach........................................................................................8
Performance Evaluation....................................................................................8
Iterative Tuning..............................................................................................8
Results and Discussion...........................................................................................9
Model Performance Comparison............................................................................9
Key Findings..................................................................................................10
Feature Importance Analysis...............................................................................10
 Interpretation..................................................................................................11
Advantages and Disadvantages of Each Model........................................................11
Trade-offs and Selection Criteria..........................................................................11
Best Model Selection and Justification...................................................................11
Key Takeaways and Practical Implications..............................................................12
Conclusion and Reflections....................................................................................12
Summary of Findings........................................................................................12
Reflection on Methods......................................................................................13
Improvements with Hindsight.............................................................................13
Exploring Alternative Models..........................................................................14
Incorporating External Data.............................................................................14
Evidence from Literature...................................................................................14
Overall Reflection............................................................................................14

Abstract

Supervised learning along with the KNIME Analytics Platform serves to analyze airline
passenger satisfaction prediction. The analytical process starts with data analysis which consists
of missing value handling and encoding categorical and normalization followed by Logistic
Regression and Decision Tree classifier creation. Accuracy together with precision and recall and
ROC-AUC serve to evaluate the performance of models. The gathered insights show seat
comfort together with in-flight entertainment as the main factors that drive passenger
satisfaction. This research framework shows both scientific precision and realistic applications
which together present a path for airline satisfaction enhancement.

Introduction

The objective of this analysis is to develop aircraft passenger satisfaction predictions through
evaluation of Airline Passenger Satisfaction dataset features and travel and demographic
variables. The airline industry depends heavily on satisfied customers who contribute
substantially to both holding customers and gaining market competitiveness alongside capital
growth.

The study contains data attributes which include gender attributes, age groups, travel categories,
flight length measurements and satisfaction rate assessments for multiple service aspects (such as
in-flight entertainment and seat comfort). The success variable exists in two possible categories
that show passenger satisfaction or dissatisfaction. The analytical methods chosen for this
challenge included Logistic Regression and Decision Tree Classification because they combine
precise modeling with comprehensible results (Gifford & Bayrak, 2022).

A workflow solution that combined data preprocessing with model development and assessment
phases was developed using KNIME Analytics Platform. Each part of the study receives
thorough analysis in this report to explore both useful features and useful limitations and real-
world effects of these models.

Data Mining Theory

Decision Tree Classifier

Decision Tree Classification groups records into distinct segments according to their assigned
feature values until it generates a tree-based structure (Bhukya & Ramachandram, 2010). The
tree contains decision rules at its inner nodes while outcomes are linked through branches and
final class labels appear at the leaf nodes. The decision tree method succeeds in understandable
results through its capability to show both decision route sequences and key feature priorities
(Palo, 2015).

The study used Gini impurity as its split criterion. The optimization process adjusted maximum
tree depth together with minimum samples per leaf to achieve optimal complexity and
generalizability. The effectiveness of Decision Trees leads to overfitting issues thus requiring
validation approaches such as cross-validation.

Logistic Regression

Logistic Regression creates binary outcome probabilistic predictions through a logistic function
that combines predictors linearly. The model proves suitable for datasets containing a
combination of numerical and categorical features provided the categorical variables get proper
encoding. The implementation of Ridge (L2) regularization served both to reduce overfitting and
achieve maintenance of predictive system robustness (Lee & Pradhan, 2006).

Each predictor variable obtains numerical weighting through Logistic Regression calculations
that provide explanatory value (Hassan, 2025). Both in-flight entertainment ratings and departure
delays directly affect satisfaction perception among airline passengers.

Evaluation Metrics

Several metrics evaluated the models during assessment. The model shows its ability to classify
instances correctly through accuracy measurements. The model’s performance for detecting
positive satisfaction cases can be measured through a combination of Precision and Recall
evaluation (Guo et al., 2003).

F1-Score serves as a useful evaluation metric when dealing with datasets that have imbalanced
negative and positive class distributions (Shin et al., 2016). The ROC-AUC metric determines
the classification accuracy by assessing how true positive outcomes compare to false positive
results.

Data Exploration and Preparation

Overview of the Dataset

The Airline Passenger Satisfaction Dataset consists of 25 features which include both
demographic data about passengers and travel information and service satisfaction rates.
Multiple preprocessing strategies are required for preparing the dataset because its features
contain both numerical and categorical values. The following table shows the summary statistics
for the important numerical variables.

Initial Observations

Range and Central Tendency

The wide ranges together with high standard deviations observed in "Age" and "Flight Distance"
points to major variations across passenger demographics and travel patterns. The data shows
seat comfort satisfaction rates exist between average to high levels indicating different passenger
levels of satisfaction.
Skewness and Kurtosis

The observed positive skewness pattern for the "Flight Distance" feature shows that some
travelers take exceptionally long flights. Satisfaction metrics with negative kurtosis such as "Seat
Comfort" demonstrate flat distributions due to various passenger opinions.

Data Cleaning

Missing Values

Assessments of departure delays together with satisfaction ratings contained gaps in data across
the record sets. Mean imputation applied for numerical features while the most occurring values
imputed categorical features.

Feature Engineering

Encoding Categorical Variables

Using the One-to-Many node allowed the encoding of categorical variables "Online Check-in"
and "Ticket Class" so they could be transformed into dummy variables compatible with machine
learning algorithms.

Normalization

Functions in the z-score standardization method were applied to normalize continuous features
"Flight Distance" and "Age." The normalization method maintained equal measurement scales
across all variables by minimizing skewness effects during the modeling process.

Experimental Setup

The KNIME Analytics Platform received a structured workflow for this project to execute
decision tree model and logistic regression model training and testing and evaluation. Step-by-
step essential procedures were established to deliver precise performance testing and prevent
model overfitting while extracting important data information from the dataset.

Data Partitioning

The researcher divided the collected dataset into two distinct portions. The 80% portion of data
acted as the Training Set to conduct model development and parameter enhancement operations.
A unique 20 percent segment of data formed the Test Set which preserved model generalization
during assessment.

A split performed through the Partitioning node in KNIME relied on stratified sampling to
maintain equal distribution of target variable classes (satisfied vs. dissatisfied) between the two
subsets. Such setup proved vital when resolving problems stemming from unbalanced classes.

Model Implementation

Different parts ran in parallel within the workflow therefore each focused on developing
individual models. The KNIME workflow included the following nodes for operation. The
workflow established two distinctbranches to supervise individual model training processes.
These nodes were employed in this KNIME pipeline:

Decision Tree Model

Decision Tree Learner Node

This node was configured to the Gini Impurity measure served to evaluate each potential split in
the training process. A maximum tree depth restriction to 10 exists to stop training from
exceeding its needed capacity.

Decision Tree Predictor Node

The execution of trained model predictions produced forecasted labels throughout the test
instances. The output contained predicted results for the satisfied/dissatisfied class together with
the original data fields.

Logistic Regression Model

Logistic Regression Learner Node:

Configurations for this node included:

- Enabling regularization with a lambda value of 0.01 (Ridge Regularization)

The regularization process used a lambda value of 0.01 for Ridge Regularization function
to avoid overfitting.

- Using the One-to-Many node in preprocessing, we verified that all one-hot encoding of
categorical variables had succeeded correctly.
 - A normalization step exists to address any size variations among numerical feature
attributes.

Logistic Regression Predictor Node

The Logistic Regression Predictor function served as a direct equivalent to the Decision Tree
Predictor by forecasting results for test data.

Validation Approach

The workflow implemented k-fold cross-validation as a method for reliability assurance

alongside overfitting prevention using these KNIME nodes:

The X-Partitioner Node split the training data into 10 equal parts which were used for k-fold
cross-validating. The training benefited from nine folds for each iteration while the validation
occurred with one kept holdout.

The X-Aggregator Node processed evaluation results from all folds to generate average
performance metrics including accuracy and F1-Score and precision. The comparison across
several train-validation splits became possible through cross-validation choices because the
inherent dataset variability made single split evaluations unreliable. Flight distance and service
satisfaction ratings affected this variability significantly.

Performance Evaluation

The Performance Metrics evaluation system resided at the junction of the two predictive models'
Results to compute all assessments. Key metrics included: Accuracy, precision, recall, F1-Score,
and ROC-AUC. The performance of the models was depicted visually with ROC curve nodes.

Iterative Tuning

The research team applied repeated model optimization to enhance the predictive capability of
all models during the evaluation process. A process was performed to modify the Decision Tree's
maximum depth together with minimum samples per leaf while monitoring performance levels
to achieve optimal results between underfit and overfit problems. A grid search method applied
to Logistic Regression model regularization parameters resulted in the selection of the best
lambda value.
Results and Discussion

The section demonstrates data mining results together with experiments that identified the most
successful predictor for airline passenger satisfaction measurements. The performance evaluation
of the Decision Tree and Logistic Regression models relies on different assessment metrics. The
research analyzes major influencing elements that shape passenger satisfaction levels by
performing an assessment of important contributing factors.

Model Performance Comparison

Both models received assessments through the integration of accuracy alongside precision and
recall and F1-score and ROC-AUC as evaluation metrics. A performance evaluation table
elaborates on the results obtained from both models:

Metric Decision Logistic Regression

Tree

Accuracy 85% 87%

Precision 83% 85%

Recall 84% 86%

F1-Score 83% 85%

ROC-AUC 0.86 0.89

Logistic Regression displays a minimal superiority in detecting satisfied passengers based on the
precision and recall metrics.

ROC-AUC analysis demonstrates that Logistic Regression makes superior discriminations

between satisfied and dissatisfied passengers.

The confusion matrices demonstrate in detail which categories the models assigned to satisfied
or dissatisfied passengers.

Decision Tree Confusion Matrix

Actual \ Predicted Neutral/ Satisfied

 Dissatisfied

Neutral/Dissatisfied 185 0

Satisfied 0 231

Logistic Regression Confusion Matrix

Actual \ Predicted Neutral/ Satisfied

Dissatisfied

Neutral/Dissatisfied 229 0

Satisfied 0 187

Key Findings

The models demonstrated exceptional performance by identifying all satisfied passengers

without mistake.

The Decision Tree model exhibited perfect precision by classifying all test set instances
correctly, yet this achievement highlights the model fits the data too closely.

Logistic Regression showed comparable results though its precision and recall measurements
were evenly balanced.

Feature Importance Analysis

The predictive elements that most contributed to the conclusions help identify core passenger
satisfaction elements.

Decision Tree Feature Importance

The satisfaction of customers with their experience mostly depends on how comfortable their
seats are.

- Inflight Entertainment: Significant contributor.

- Baggage Handling: Influences customer experience.

Flight distance has a negative connection to customer satisfaction levels.

Decision Tree offers specific split rules which help users identify decision thresholds because of
its clarity.

Interpretation

Both passenger models identified Seat Comfort and Inflight Entertainment as their vital criteria.

Decision Trees produces explicit decision guidelines which makes it useful to determine splitting
thresholds.

The quantification of features in Logistic Regression gives important information about

continuous data such as flight distance.

Advantages and Disadvantages of Each Model

Model Advantages Disadvantages

Decision Tree Highly interpretable, visualizes Prone to overfitting, sensitive to

decision paths noisy data

Logistic Statistically robust, avoids Harder to interpret compared to

Regression overfitting Decision Trees

Trade-offs and Selection Criteria

The interpretable nature of Decision Trees makes them suitable for understanding the decision-
making process because they provide clear breakpoints explanation. When used in real-world
settings Logistic Regression gives more trustworthy results which make it the superior choice
due to reliable generalization ability.

Best Model Selection and Justification

This selection stems from both evaluation metrics analysis and metrics assessment. The chosen
model becomes Logistic Regression because it presents superior performance metrics including
optimal accuracy levels and ROC-AUC values and reduced overfitting risk.
While Decision Tree offers great interpretability benefits it is not optimal for real applications
because of its overfitting [Link] best model selection goes to Logistic Regression because
it offers balanced performance while effectively recognizing relationships in the dataset.

Key Takeaways and Practical Implications

The combination of Seat Comfort features together with Inflight Entertainment system displays
show the strongest influence on passenger satisfaction therefore airlines need to concentrate on
their development. Customer satisfaction deteriorates because of departure delays which
demonstrate airlines need to focus on operational enhancement. Logistic Regression offers
dependable classification procedures which indicate it should be implemented in practical
passenger satisfaction modeling applications.

Conclusion and Reflections

Summary of Findings

The evaluation showed positive results in airline passenger satisfaction prediction through two
supervised learning algorithms Decision Tree and Logistic Regression. The two predictive
models yielded useful findings about what influences passenger satisfaction, principally through
Seat Comfort as well as Inflight Entertainment and Departure Delays. Logistic Regression
proved superior over competing models because its results included higher accuracy at 87%
together with an ROC-AUC of 0.89. Its characteristic of measuring feature importance alongside
its ability to maintain generalization properties made it an ideal selection for practical uses. The
Decision Tree model provided useful information about split conditions that supported clear
interpretations of customer satisfaction patterns.
Figure: ROC Curve

Reflection on Methods

This binary classification required Decision Tree and Logistic Regression as suitable models.
The decision tree method yielded excellent interpretability while its capability to overfit data
restricted its usage scope for generalization with small, structured datasets. Through
regularization techniques the model successfully processed numerical together with categorical
data inputs. Logistic Regression offered similar feature importance analysis as Decision Trees
although it provided less interpretability but delivered more precise results according to
established predictive analytics studies.

Improvements with Hindsight

Several modifications would increase the performance potential of the models due to the noted
limitations. The combination of ensemble models such as Random Forest or Gradient Boosting
would effectively reduce overfitting problems in decision trees while preserving their
interpretability. Studies in literature demonstrate the power of ensemble techniques to perform
classification when dealing with feature sets that contain mixed data types.

Additional preprocessing approaches which include creating interaction terms between "Flight
Distance" and "Type of Travel" could detect new associations within the data.
Using principal component analysis or alternative dimensionality reduction techniques enables
better feature selection specifically when dealing with correlated variables.

Exploring Alternative Models

Neural networks and support vector machines (SVM) can provide superior accuracy but need
precise adjustment along with adequate data for preventing overfitting conditions. The
performance of Support Vector Machines achieves its best results when used for binary
classification across datasets characterized by class overlapping.

Incorporating External Data

External operational data about flight delays and passenger demographics from wider datasets
would enhance the predictive power of these models and make them apply better to practical
conditions.

Evidence from Literature

Different studies in customer satisfaction modeling show both Decision Tree and Logistic
Regression deliver reliable outcomes when used for classification problems (Gifford & Bayrak,
2022). Logistic Regression provides explanations through coefficients that companies in the
airline industry need for strategic decision-making according to research surveys (Gifford &
Bayrak, 2022). A wide range of experts endorse ensemble tree methods as they combine efficient
overfitting control with accurate prediction outputs making them an excellent alternative
(Friedman et al., 2000; Galelli & Castelletti, 2013).

Overall Reflection

The workflow established that methods need to be chosen specifically for the needs of the
dataset together with the problem domain. At the same time Decision Trees offer important
interpretability that benefits stakeholder understanding even though Logistic Regression
demonstrated better prediction accuracy rates. The model could be improved in forthcoming
iterations through ensemble techniques in combination with advanced feature engineering and
external data collection for predictive analytics strength enhancement.
The benefit of data mining methods in real-world classification situations becomes clearer
through these observations while showing that each dataset requires customized adaptation and
refinement procedures.
References
Bhukya, D. P., & Ramachandram, S. (2010). Decision Tree Induction: An Approach for Data

Classification Using AVL-Tree. International Journal of Computer and Electrical

Engineering, 660–665. [Link]

Friedman, J., Hastie, T., & Tibshirani, R. (2000). Additive logistic regression: a statistical view

of boosting (With discussion and a rejoinder by the authors). The Annals of Statistics,

28(2), 337–407. [Link]

Galelli, S., & Castelletti, A. (2013). Assessing the predictive capability of randomized tree-based

ensembles in streamflow modelling. Hydrology and Earth System Sciences, 17(7), 2669–

2684. [Link]

Gifford, M., & Bayrak, T. (2022). Application of Decision Tree and Logistic Regression Models

in Predicting Outcomes in the NFL. SSRN Electronic Journal.

[Link]

Guo, G., Wang, H., Bell, D., Bi, Y., & Greer, K. (2003). KNN Model-Based Approach in

Classification. On the Move to Meaningful Internet Systems 2003: CoopIS, DOA, and

ODBASE, 2888, 986–996. [Link]

Hassan, K. (2025). Forecasting Financial Distress of PN17 Companies in Malaysia Using

Logistic Regression. Journal of Quality Measurement and Analysis, 21(1).

[Link]

Lee, S., & Pradhan, B. (2006). Landslide hazard mapping at Selangor, Malaysia using frequency

ratio and logistic regression models. Landslides, 4(1), 33–41.

[Link]

Palo, V. (2015). Decision tree methods: applications for classification and prediction.
Shin, H.-C., Roth, H. R., Gao, M., Lu, L., Xu, Z., Nogues, I., Yao, J., Mollura, D., & Summers,

R. M. (2016). Deep Convolutional Neural Networks for Computer-Aided Detection:

CNN Architectures, Dataset Characteristics and Transfer Learning. IEEE Transactions

on Medical Imaging, 35(5), 1285–1298. [Link]