It is important to note that many churn prediction studies, including this one, rely on benchmark datasets such as the publicly available Telco Customer Churn dataset from Kaggle. While this facilitates reproducibility and direct comparison with prior work, it also highlights the need for future research to validate findings across multiple datasets from different contexts, geographies, and service portfolios. Churn behavior may vary based on cultural, regulatory, and market-specific factors (Verbraken et al., 2012), and our study acknowledges this limitation while using the IBM dataset as a well-established benchmark for methodological development.

To address these opportunities, this study pursues four interconnected objectives. First, we design and validate a Soft-voting ensemble framework that combines diverse machine learning algorithms for robust churn prediction. Second, we integrate SHAP-based explainability to provide both global and local interpretability of model decisions. Third, we apply autoencoder-based representation learning to discover latent customer segments with distinct churn risk profiles. Fourth, we evaluate model performance using both statistical metrics and cost-sensitive business measures to ensure practical utility.

This study employs the IBM Telco Customer Churn dataset as it represents a well-established benchmark in churn prediction literature, enabling direct comparison with prior work. The dataset's comprehensive feature set and public availability support reproducibility and methodological evaluation. While acknowledging that findings from a single dataset may not generalize to all telecom contexts, this dataset provides a standardized foundation for developing and comparing analytical approaches that can be subsequently validated on additional datasets.

The original dataset comprises 7, 043 customer records with 21 features that capture demographic information, account details, service subscriptions, and customer behavior patterns. The feature set can be mathematically represented as:

where xi∈ℝ21 represents the feature vector for customer i, and y_i∈{0, 1} denotes the binary churn status (0: No Churn, 1: Churn).

The feature space X encompasses three distinct types of variables:

The target variable distribution exhibits significant class imbalance, with approximately 73.5% of customers retaining services and 26.5% churning, which necessitates specialized handling strategies discussed in Section 3.3.

A rigorous data quality assessment was conducted to identify and address potential issues that could compromise model performance:

Missing values analysis: Systematic examination revealed missing values predominantly in the TotalCharges feature, which were identified as blank entries rather than explicit null values. The missing pattern was determined to be Missing Completely at Random (MCAR) through statistical testing.

Data type validation: Initial exploration identified data type inconsistencies, particularly with the TotalCharges feature being stored as string type due to the presence of whitespace characters in missing values. This required type conversion to numerical format for analytical processing. For example, ContractDuration_i is encoded numerically (Month-to-month = 1, 1 year = 12, 2 year = 24).

Missing value imputation: For the TotalCharges feature with missing values (11 instances, 0.16%), median imputation was selected:

Domain justification: In telecommunications billing data, charge distributions often exhibit right skewness due to high-value outliers. Median imputation preserves the central tendency while minimizing distortion from extreme values, which is critical for maintaining the economic interpretability of spending patterns. This approach aligns with industry practices where billing anomalies are handled conservatively to avoid artificial inflation of customer value metrics.

Contract duration encoding: Contract types were encoded numerically (Month-to-month = 1, 1 year = 12, 2 year = 24) rather than using one-hot encoding alone. Domain justification: This ordinal encoding captures the inherent hierarchy in contract commitment, which directly correlates with churn risk in telecommunications. The numerical representation preserves the business intuition that longer contracts indicate higher commitment, enabling models to learn this progressive relationship more effectively than treating contract types as purely categorical.

Feature scaling: Numerical features were standardized using StandardScaler. Domain justification: In telecom datasets, features like MonthlyCharges (range: $18–118) and tenure (range: 0–72 months) operate on vastly different scales. Standardization prevents algorithm bias toward features with larger numerical ranges, particularly important for distance-based algorithms and neural networks. This ensures that all customer attributes contribute proportionally to the model's learning process.

The addition of 1 in the denominator prevents division by zero for new customers with zero tenure.

Domain rationale: Traditional metrics like TotalCharges and MonthlyCharges provide incomplete pictures of customer value. In telecommunications, a customer with high total charges over long tenure represents stable loyalty, whereas similar spending over a short period may indicate premium but potentially unstable service usage. This normalized metric captures spending intensity relative to relationship duration, addressing a key business insight: customers who spend more per month relative to their tenure may have higher perceived value or, conversely, may experience “bill shock” leading to churn. This feature incorporates principles from behavioral economics, particularly the sunk cost effect, where customers perceive greater investment in services used intensively over time.

where k = 9 represents the total service categories: phone lines, internet services, online security, device protection, technical support, streaming TV, streaming movies, and associated add-ons.

Domain rationale: Service bundling is a fundamental strategy in telecommunications, yet its impact on churn is complex. While theory suggests multiple services create higher switching costs and loyalty, industry experience shows complex bundles can lead to confusion and “bill shock.” This feature quantifies service engagement depth, capturing the dual nature of bundling: customers with 3+ premium services showed 67% lower churn risk in our analysis, validating the switching costs principle, yet those with only basic services but high monthly charges exhibited elevated risk. The feature directly measures a customer's integration into the service ecosystem, a key indicator of retention potential in telecom CRM strategies.

The binary nature of churn prediction requires careful examination of class balance, as standard classification algorithms often underperform on imbalanced datasets. Analysis of the Telco Customer Churn dataset revealed a significant skew: churned customers constitute the minority class at approximately 26.5% of the dataset, while retained customers comprise the majority 73.5%. This corresponds to a class distribution ratio of approximately 1:2.77, meaning for every churned customer, there are nearly three retained ones.

Such imbalance presents three critical challenges:

Predictive bias: Algorithms tend to optimize overall accuracy by favoring the majority class, potentially achieving high accuracy while failing to identify churners—the customers of greatest business interest.

Evaluation metric distortion: Traditional accuracy metrics become misleading, necessitating alternative measures like precision, recall, F1-score, and AUC-ROC that better capture minority-class performance.

Cost-sensitive considerations: From a business perspective, the cost of false negatives (missed churners) typically exceeds that of false positives, further emphasizing the need for specialized imbalance handling techniques.

A diverse set of machine learning algorithms was implemented to leverage their complementary strengths and provide comprehensive coverage of different modeling paradigms.

To ensure full reproducibility, Table 4 provides the complete hyperparameter configurations for all models used in this study. These parameters were determined through systematic grid search with 5-fold cross-validation, with final values selected to optimize F1-score.

All models were implemented using scikit-learn (v1.3.0), except for XGBoost (v1.7.0) and LightGBM (v4.1.0). To ensure full reproducibility, the random seed was consistently set to 42 across all components of the pipeline, including data splitting (train_test_split), cross-validation (StratifiedKFold), model initialization (random_state in all classifiers), and oversampling with SMOTE. This setup guarantees that results can be reliably reproduced under the same software environment and data preprocessing steps.

Seven machine learning models were implemented to ensure comprehensive evaluation across distinct algorithmic paradigms:

This selection provides coverage of the primary algorithmic families used in classification tasks, from linear discriminants to complex ensemble methods.

Global feature importance was quantified via mean absolute SHAP values:

Table 5 summarizes key global insights alongside business interpretation and retention actions.

Local explanations decompose individual predictions:

Example: Customer 7590-VHVEG (predicted churn probability 0.87): Month-to-month contract (+0.42), electronic check (+0.18), tenure < 6 months (+0.15) increased risk; mitigated by multiple services (-0.08) and auto-pay (–0.03). Actions: contract upgrade, payment migration, personalized onboarding.

Counterfactual analysis indicated that converting to a one-year contract would reduce churn by 0.35, with technical support providing an additional 0.08 reduction, enabling precise ROI calculation for interventions.

To bridge technical explanations and business understanding, we implemented multi-level visualizations:

These visualizations, combined with global and local SHAP insights, transformed predictions into transparent decision-support tools, reducing predicted churn by 18%–25% in validation scenarios while improving stakeholder trust and adoption.

We employed a deep autoencoder for nonlinear dimensionality reduction, transforming the 46-dimensional feature space into a compressed 16-dimensional latent representation. This process captures essential customer behavior patterns while mitigating the curse of dimensionality. The symmetric encoder-decoder architecture (detailed in Appendix A1) learns to reconstruct input features through a bottleneck layer, forcing the model to retain only the most salient information for customer differentiation.

The autoencoder was trained to minimize reconstruction error with L2 regularization, using the Adam optimizer with early stopping to prevent overfitting. The resulting latent representations provide a denoised, lower-dimensional space optimized for clustering.

K-means clustering was applied to the learned latent representations, leveraging K-means++ initialization and multiple restarts to avoid local minima. We standardized latent features prior to clustering and evaluated cluster counts from 2 to 6 using multiple validation metrics. The clustering objective minimizes within-cluster variance while maximizing between-cluster separation (mathematical formulation in Appendix A2).

Cluster quality was assessed through three complementary internal validation metrics:

These metrics (defined in Appendix A3) were computed for K∈{2, 3, 4, 5, 6} across multiple clustering algorithms. The autoencoder+K-means combination with K = 3 demonstrated optimal performance, balancing interpretability with statistical validity.

Systematic cluster interpretation employed multidimensional profiling across four domains:

Statistical testing (ANOVA for continuous, chi-square for categorical variables) identified significant inter-cluster differences. Effect sizes (Cohen's d, Cramér's V) quantified practical significance, while visual analytics (parallel coordinates, radar charts) facilitated intuitive interpretation.

This segmentation framework enables identification of distinct customer archetypes with unique churn propensity profiles, providing actionable insights for targeted retention strategies and personalized customer engagement.

Model performance is assessed using standard classification metrics computed via stratified 5-fold cross-validation: accuracy, precision, recall, F1-score, and AUC-ROC. Due to class imbalance, F1-score serves as the primary optimization metric, balancing precision and recall for the minority churn class.

To align model assessment with business objectives, we implement a cost-sensitive framework where false negatives (missed churners) incur substantially higher costs than false positives. The cost function is defined as:

where C_FP = 1 and C_FN = 5. This 5:1 cost ratio is derived from multiple considerations:

Business justification:

Statistical validation:

This cost matrix enables model selection and threshold optimization that prioritizes business impact over purely statistical metrics.

Our evaluation employs a multi-tiered statistical validation approach:

The integrated framework balances technical performance with business relevance, supporting informed model selection for operational deployment.

The dataset (n = 7,043) was split into training (80%, n = 5,634) and test (20%, n = 1,409) sets using stratified sampling with random_state=42 to maintain class distribution. SMOTE was applied exclusively to the training set to prevent data leakage. All models were evaluated using stratified 5-fold cross-validation on the training set. Each fold maintained the original churn class distribution (73.5% non-churn, 26.5% churn). Performance metrics were averaged across folds, with standard deviations calculated to assess variability.

The optimal decision threshold was determined through a two-step process:

All experiments used fixed random seeds (42) for data splitting, SMOTE oversampling, and base model initialization to ensure reproducibility. Hyperparameter optimization was performed using Optuna with 5-fold stratified cross-validation, as documented in Appendix B. The optimization objective was evaluated consistently across folds, and the best-performing hyperparameter configurations were retained for final model training.

The analysis commenced with the Telco Customer Churn dataset, which comprises records for 7,043 customers, described by 21 original features spanning customer demographics, service subscriptions, account information, and the target churn indicator. An initial examination of the target variable, Churn, confirmed a significant class imbalance, which is a common challenge in churn prediction. Specifically, only 1,869 customers (26.5%) churned (“Yes”), while the majority, 5,174 customers (73.5%), were retained (“No”). This imbalance necessitated specialized sampling techniques during model training to prevent algorithmic bias toward the majority class. Data quality checks revealed a minimal presence of missing values, confined exclusively to the TotalCharges column, where 11 instances (0.16% of the dataset) were blank. These missing values were logically imputed using the median of the TotalCharges variable. This approach was selected over mean imputation to maintain robustness against potential skewness in the charge distribution and to preserve the integrity of the dataset without introducing significant bias. The class distribution of churn across key categorical variables is shown in Figure 2, which reveals substantial imbalance between customers who stayed (73.5%) and those who churned (26.5%). Month-to-month contract holders exhibited a notably higher churn proportion, highlighting the transient nature of short-term subscriptions. Similarly, customers using electronic check payments and those subscribing to fiber-optic internet services demonstrated elevated churn rates, indicating that both pricing sensitivity and perceived service reliability are strong behavioral signals. This imbalance justified the later application of Synthetic Minority Oversampling (SMOTE) to ensure fair model learning and mitigate bias toward the majority class. Such visualization-driven diagnostics strengthen the data understanding stage, which is crucial for robust feature engineering and subsequent predictive modeling (Huang et al., 2012).

To enhance the predictive power of the model and capture more nuanced customer behaviors, two new features were engineered from the existing variables.

The first engineered feature, AvgMonthlyCharge.

This feature normalizes a customer's total spending by their lifetime with the company, effectively capturing their average monthly expenditure. The utility of this feature was substantiated through correlation analysis, which revealed that AvgMonthlyCharge exhibited a stronger association with the churn outcome (correlation coefficient, r≈0.23) than the raw MonthlyCharges (r≈0.19). This indicates that the normalized spending metric provides a more discriminative signal for identifying churn-prone customers.

The second engineered feature, HasMultipleServices, was created by summing the number of active services a customer subscribes to, including PhoneService, MultipleLines, InternetService, and various premium add-ons such as OnlineSecurity and StreamingTV.

The correlation heatmap in Figure 3 visualizes the relationships between three key numerical features: tenure, MonthlyCharges, and TotalCharges. These specific features were selected for analysis as they were consistently ranked among the ten most important by the SHAP analysis. The plot reveals expected strong correlations, particularly between tenure and TotalCharges, which is logical as total charges accumulate over time. This pattern supports the hypothesis that bundled services often co-occur among loyal customers, reinforcing their retention likelihood. The weak correlations between financial indicators and tenure also confirm that spending behavior alone cannot fully explain churn dynamics—contextual and service quality features play complementary roles. This aligns with recent findings emphasizing the importance of multi-dimensional feature representation in churn modeling (Óskarsdóttir et al., 2017).

To directly address the class imbalance identified in Section 4.1.1, the SMOTE was applied to the training data. The application of SMOTE successfully balanced the dataset by adjusting the class distribution. Initially, the dataset exhibited an imbalanced split between churn and non-churn customers:

After applying SMOTE, the distribution was modified to achieve perfect balance:

This balancing ensures that both classes contribute equally during model training, reducing bias toward the majority class and improving generalization.

The effectiveness of SMOTE was quantitatively demonstrated in the subsequent model evaluation phase. Post-SMOTE, all classification models showed a marked improvement in sensitivity (recall for the churn class). This was particularly evident in complex ensemble methods like LightGBM and XGBoost, which leverage multiple base learners. These models benefited greatly from the more balanced data, as it provided a richer and more varied set of minority class examples to learn from, thereby reducing their inherent bias toward predicting the majority class. For instance, without SMOTE, models tended to achieve high accuracy by simply predicting “no churn” for most cases, but with SMOTE, their ability to correctly identify true churners (True Positives) increased substantially without a proportional rise in false alarms, as reflected in the significantly higher F1-scores for the churn class across the board.

SHAP (SHapley Additive exPlanations) assigns each feature an importance value for a specific prediction based on cooperative game theory. A higher positive SHAP value indicates the feature increases the likelihood of churn, while a lower negative SHAP value decreases the likelihood of churn (promoting customer retention). The mean absolute SHAP value for each feature provides its average importance across all customers.

The SHAP analysis revealed distinct patterns between churned and retained customers:

Churned customers tended to be on Monthly Charges, pay higher monthly fees, have shorter tenures, and often lack security or support add-ons. They predominantly used electronic check payment methods, suggesting a preference for flexibility over automated billing.

Non-churned customers typically had long-term contracts, multiple services (security, backup, streaming), and automatic payment setups. Their monthly costs were either lower or justified by a comprehensive service bundle, indicating higher perceived value and commitment.

The clear dominance of contract-related features in the SHAP analysis underscores the critical importance of customer commitment in retention strategies, while the strong showing of tenure highlights the vulnerability of newer customers.

Contract type analysis: Month-to-month contracts exhibited SHAP values 3.2 × higher than one-year contracts and 4.8 × higher than two-year contracts, quantitatively indicating that longer contractual commitments were associated with lower churn risk in the model's predictions. This finding underscores the critical importance of contract structure in customer retention strategies illustrated in Figure 5.

Tenure effects: SHAP dependence plots revealed a non-linear relationship where predicted churn risk decreases exponentially with tenure in the model, stabilizing after approximately 24 months. The analysis showed that customers with less than 12 months tenure had 3.7 × higher churn probability compared to those with tenure exceeding 24 months illustrated in Figure 6.

Service bundle impact: In the model's predictions, customers with multiple premium services (OnlineSecurity, TechSupport, DeviceProtection) showed 67% lower average SHAP values, suggesting service bundling is associated with lower predicted churn risk. The presence of 3+ premium services reduced churn probability by 58% compared to customers with only basic services illustrated in Figure 7.

The SHAP analysis enabled the identification of distinct customer profiles with varying predicted churn risk, providing actionable segmentation for targeted retention strategies:

Strategic implications: The SHAP-based risk profiling presented in Table 6 can inform resource allocation, with the high-risk segment (18% of customers) representing the primary focus for retention efforts. By targeting interventions based on feature contributions identified through SHAP analysis, organizations can achieve estimated cost savings of 35%–45% in retention marketing expenditures while improving overall retention rates by 18%–25%.

The explainable AI approach bridges the gap between predictive accuracy and business actionability, transforming black-box model outputs into interpretable insights that can inform strategic customer retention initiatives aligned with organizational objectives.

Comprehensive evaluation of six machine learning algorithms and one deep learning model revealed significant performance variations, as summarized in Table 7. The models were trained on SMOTE-balanced data and evaluated using multiple metrics to assess their predictive capability for customer churn.

Critical finding: The comprehensive evaluation revealed distinct performance tiers among the tested models. Tree-based ensemble algorithms, particularly gradient boosting variants, consistently outperformed others. XGBoost, LightGBM, and Gradient Boosting achieved the highest balanced performance, each attaining an accuracy, precision, recall, and F1-score of 0.84. XGBoost obtained the highest discriminative ability with an AUC-ROC of 0.932, as further illustrated in the ROC curve comparison (Figure 8). The Soft-Voting Ensemble of the top models matched this high F1-score (0.84) while maintaining a robust AUC of 0.918, demonstrating effective model consolidation. In contrast, Random Forest showed solid but lower performance (F1-score: 0.81, AUC: 0.887), with no immediate signs of overfitting indicated by the presented metrics, followed by AdaBoost, Logistic Regression, and the MLP.

The probability calibration process significantly improved model reliability, addressing the well-documented tendency of complex ensemble methods to produce overconfident probability estimates (Niculescu-Mizil and Caruana, 2005; Guo et al., 2017). Platt scaling is specifically designed to correct miscalibration by fitting a sigmoid function to a model's outputs, making it most suitable for classifiers whose raw scores are approximately sigmoidal, such as Logistic Regression. For tree-based ensembles (Random Forest, XGBoost, LightGBM, Gradient Boosting, and AdaBoost) and neural networks (MLP), the predicted probabilities often exhibit complex, non-linear miscalibration patterns that a simple sigmoid cannot adequately correct. Consequently, isotonic regression, a non-parametric, monotonic mapping, is generally preferred for these models, as it can flexibly adjust for overconfident or irregular probability estimates, which explains why Platt scaling is applied primarily to Logistic Regression while isotonic regression improves calibration more effectively for the other techniques. Table 8 summarizes calibration performance across all evaluated models, measured by Brier score reduction following established evaluation protocols (Glenn et al., 1950; Shafer and Vovk, 2008).

Key Ffindings:

Figure 9 illustrates reliability curves for XGBoost before and after isotonic calibration, demonstrating improved alignment between predicted probabilities and observed event frequencies across all probability bins. The calibrated model shows near-diagonal alignment, indicating well-calibrated probability estimates suitable for business decision-making (Shafer and Vovk, 2008).

To complement the SHAP-based feature importance findings, we constructed a comprehensive multi-panel analytical visualization examining churn rates across four critical attributes identified by the machine learning models: contract type, payment method, internet service, and tenure group (Figure 10). These descriptive results provide intuitive confirmation of the primary churn drivers and validate the model's feature importance rankings through direct observational evidence.

The analytical visualization provides the following critical insights that align with and validate the SHAP analysis:

Strategic integration with model insights:

This multi-faceted analysis bridges the gap between predictive modeling and business intelligence, providing both statistical validation of the machine learning insights and intuitive visual evidence that stakeholders can readily understand and act upon. The convergence of SHAP-based explanations and descriptive analytics creates a robust foundation for data-driven retention strategies.

A comprehensive evaluation of clustering approaches was conducted using multiple internal validation metrics to identify the optimal customer segmentation methodology. Table 9 and Figure 11 presents the comparative performance of various clustering algorithms across silhouette scores, Calinski-Harabasz indices, and Davies-Bouldin indices.

Algorithm performance analysis:

Optimal solution rationale: The Autoencoder + K-means combination with k = 3 clusters was selected as the optimal segmentation approach based on its superior performance across all internal validation metrics. The autoencoder's dimensionality reduction capability effectively addressed the high-dimensionality challenges, learning meaningful latent representations that enabled more coherent cluster formation in the reduced space.

The methodological superiority of AE+KMeans stems from its ability to learn non-linear feature representations through the autoencoder's deep architecture, effectively capturing complex customer behavior patterns in a lower-dimensional latent space. This approach mitigated the challenges of high-dimensional sparse data that plagued traditional clustering algorithms, enabling the discovery of more meaningful and actionable customer segments.

The Autoencoder-based KMeans clustering identified three distinct customer segments with dramatically different churn risks, providing a strategic lens for resource allocation in retention programs. The segmentation, primarily driven by tenure and spending patterns, reveals a clear customer lifecycle trajectory.

Overall, the analysis quantifies the critical relationship between tenure and churn, demonstrating that the initial customer lifecycle phase carries the highest attrition risk. This segmentation moves beyond simple prediction, enabling proactive and differentiated customer management by pinpointing which customers are at risk and why, based on their fundamental behavioral profile.

The revised performance metrics (Random Forest: AUC = 0.887, F1 = 0.81) demonstrate robust generalization without evidence of overfitting that perfect training metrics might indicate. We further validated model stability through:

These validation measures address concerns about model reliability and overfitting, and indicate reliable deployment potential despite the cross-sectional data limitations.

However, several limitations must be acknowledged. Single dataset limitation: This study utilizes a single publicly available dataset (IBM Telco), which while valuable for benchmarking and reproducibility, limits our ability to assess the generalizability of findings across different telecommunications markets, regulatory environments, and customer populations. Future research should validate the proposed framework on multiple datasets from diverse operational contexts. The temporal constraint of single-timepoint data prevents analysis of customer behavior evolution and limits causal inference. The absence of external market factors, such as competitor pricing and regional availability, may affect generalizability across different telecommunications markets. Additionally, the focus on structured data omits potential predictive signals from unstructured sources like customer service interactions and social media sentiment.