PaperlayerView exact revision on arXiv

Source-linked AI summary

A Differentially Private Federated Proximal Optimization Framework for Customer Churn Prediction in Heterogeneous Federated Telecom Networks

Joydeb Kumar Sana, Subrata Chakraborty, M M Manjurul Islam

arXiv:2609.12470v1cs.LG

TL;DR

Centralized churn prediction requires aggregating sensitive customer data, while federated training can still expose information through model updates and standard FedAvg may struggle with non-IID clients. The paper combines differential privacy with FedProx for telecom churn prediction, finding competitive performance with a small reduction while protecting privacy. It concludes that the framework offers a practical balance for collaborative prediction without raw-data exchange.

  • Problem

    Centralized churn prediction raises privacy and regulatory concerns, while heterogeneous clients and information leakage from federated model updates remain challenges.

    Source: §1 · pp. 1–2

  • Method

    The framework combines FedProx with differential privacy to train a shared churn model across telecom operators without exchanging raw customer data.

    Source: §2.4.1 · p. 5, §8 · p. 17

  • Results

    FedProx-based models consistently outperform FedAvg-based models, and DP-FedProx achieves competitive prediction performance with only a small reduction while providing privacy protection.

    Source: §7 · pp. 16–17, §8 · p. 17

  • Takeaways & Limitations

    The framework offers a practical approach to collaborative telecom churn prediction where operators cannot share raw customer data.

    Source: §8 · p. 17, §7 · p. 17

    Under tenure-based partitioning with mild client heterogeneity, the proximal correction may provide only limited practical benefit after DP-SGD noise is added.

    Source: §7 · p. 17

Abstract

from arXiv · show

Customer churn is one of the major issues in the telecommunication industry. To predict customer churn, conventional centralized machine learning approaches have been widely used. This centralized approach requires customer data to be stored in a central repository, which raises privacy concerns and may violate data protection regulations. Federated learning addresses this problem by allowing multiple telecom operators to collaboratively train a global model without transferring their raw customer data. However, real-world customer data are often heterogeneous (non-IID), which may negatively affect the performance of standard federated learning. Trained models can also suffer from privacy attacks. To address those issues, we propose a Differentially Private (DP) based Federated Proximal optimization (FedProx) framework. All experiments were performed on two publicly available telecom churn datasets. We trained Federated Averaging (FedAvg), DP-FedAvg, FedProx, and the proposed DP-FedProx framework. For baseline comparison, we also used several centralized and local models. To evaluate the models, we employed seven widely used evaluation metrics. The experimental results show that the FedProx based models consistently outperform the FedAvg based models. Compared with the best centralized model, the proposed DP-FedProx framework achieves competitive prediction performance with only a small reduction in accuracy while providing privacy guarantees. To explain our model, we conducted SHAP analysis which shows that DP-FedProx method priorities revenue group features. These results indicate that the proposed DP-FedProx framework provides a practical balance between prediction performance and data privacy protection.

1. Introduction

The paper addresses privacy risks in centralized churn prediction and the limits of standard federated learning on heterogeneous telecom data. It proposes combining differential privacy with FedProx and evaluates this approach for telecom churn prediction.

1. Introduction

  • Centralized churn models aggregate customer records, raising privacy and regulatory concerns even as telecom providers need to identify customers likely to leave.The introduction links customer retention to provider revenue and notes that retaining customers is less costly than acquiring new ones.Source: §1 · p. 1
  • Federated learning keeps customer data at its branch while training a shared model, but FedAvg assumes IID clients, unlike heterogeneous telecom settings.FedProx adds proximal regularization and is presented as a potential way to improve robustness and convergence in non-IID settings.Source: §1 · p. 2
  • The paper identifies a gap in evaluating differential privacy with non-IID federated learning for telecom churn prediction and proposes a DP-FedProx framework.The related-work discussion notes that prior telecom churn studies either transferred raw data or used FedAvg without differential privacy.Source: §1 · p. 2, §1.1 · p. 3

1.1. Literature Review

  • Prior telecom churn work included only two federated studies: one transferred raw customer data, while the other used FedAvg without differential privacy.The review also describes privacy-preserving data transformation and notes that no common framework manages heterogeneous data effectively.Source: §1.1 · pp. 2–3

1.2. Our Contributions

  • The contributions combine FedProx and differential privacy for churn prediction without raw-data sharing, addressing heterogeneity and privacy leakage.The paper describes this as the first use of differential privacy in federated telecom churn prediction.Source: §1.2 · p. 3
  • The study simulates heterogeneity with tenure/month-based client partitions and evaluates centralized, local, FedAvg, FedProx, and DP variants on two benchmarks using seven metrics.It reports global and client-level performance assessments.Source: §1.2 · p. 3

2. Methodology

The methodology evaluates churn prediction across two telecom datasets using tenure-based client partitions and client-specific feature encoding. Its preprocessing includes an adaptive Weight-of-Evidence scheme intended to avoid cross-client information leakage.

2.1. Dataset and Non-IID Client Partitioning

  • The experiments use two public telecom churn datasets with 100,000 and 7,043 records, respectively, removing missing-value rows and customer identifiers before encoding labels and categories.Dataset-2 is documented as a fictional California carrier sample; Dataset-1's similarity to Cell2Cell is an inference, not a confirmed source lineage.Source: §2.1 · p. 3
  • Customers are partitioned into four tenure-based clients, creating differences in feature and class distributions within each dataset rather than mixing records across datasets.The partition is intended to simulate heterogeneity across customer lifecycle segments.Source: §2.1 · pp. 3–4

2.2. Adaptive Weight-of-Evidence Feature Encoding

  • Client-specific Weight-of-Evidence encoding is applied independently to prevent cross-client information leakage during preprocessing.Features with more than 100 unique values are discretized into quantile bins, with granularity adapted to the local client sample size.Source: §2.2 · p. 4
  • The encoder maps each feature group using churned versus non-churned counts, assigns neutral WoE to sparse or degenerate groups, and uses a sentinel for missing values.The resulting numeric representation is described as monotonic with respect to the churn outcome.Source: §2.2 · p. 4

2.3. Adaptive Weight-of-Evidence Feature Encoding

The paper encodes each client’s features independently with adaptive Weight-of-Evidence values, then compares federated learning approaches for heterogeneous client data. FedProx adds a proximal constraint to local training to improve stability and performance in non-IID settings.

2.3. Adaptive Weight-of-Evidence Feature Encoding

  • Features with more than 100 unique values are divided into quantile bins, with bin count adapted to the client’s sample size.Features with 100 or fewer unique values are used directly as categorical groups.Source: §2.3 · p. 4
  • Each feature group receives a Weight-of-Evidence value based on churned and non-churned customer counts, replacing the original feature value.Groups below a minimum sample threshold, or lacking either class, receive a neutral value of 0.Source: §2.3 · pp. 4–5

2.4.1. Federated Learning

  • FedAvg broadcasts a global model, averages client updates weighted by client training samples, and can converge slowly with degraded model quality on non-IID data.The section motivates testing FedAvg and FedProx for telecom churn prediction.Source: §2.4.1 · p. 5
  • FedProx modifies each client’s local objective with proximal regularization that keeps its model closer to the global model during training.The coefficient µ controls the constraint; as µ approaches 0, FedProx reduces to FedAvg.Source: §2.4.1 · p. 5

2.5. Differential Privacy

The section describes differential privacy and the classifiers used to evaluate the proposed framework. Its architecture combines client-side DP-SGD with FedProx or FedAvg training and server aggregation.

2.5. Differential Privacy

  • Differential privacy formalizes protection by requiring a randomized mechanism’s outputs to be indistinguishable for datasets differing in one record.The privacy budget ϵ controls the guarantee, with smaller values providing stronger privacy; δ represents the probability of a privacy breach.Source: §2.5 · p. 5
  • DP-SGD is applied independently at each participating client in the federated setting.The section identifies DP-SGD as a widely adopted approach that incorporates noise-addition mechanisms.Source: §2.5 · p. 6

2.5.1. Centralized, Local, and Federated Learning Classifiers

  • Comparisons include centralized models, independent client Logistic Regression models, and federated ChurnNet experiments using the same neural network.The centralized set includes Random Forest, Logistic Regression, XGBoost, Gradient Boosting, AdaBoost, and ChurnNet.Source: §2.5.1 · p. 6
  • ChurnNet is a feed-forward network with two hidden layers, batch normalization, ReLU activations, dropout, and a sigmoid churn-probability output.It is trained with binary cross-entropy loss.Source: §2.5.1 · p. 6

2.6. Model Architecture

  • Four tenure-based non-IID clients train locally with DP-SGD; the server aggregates updates using FedAvg or FedProx over 20 rounds.The global model is evaluated using seven metrics.Source: §2.6 · p. 6
  • The proposed method combines FedProx with differential privacy for the study’s non-IID data setting.The procedure is formalized in Algorithm 1.Source: §2.6 · pp. 6–7

2.7. Evaluation Measures

The section defines seven evaluation metrics from the confusion matrix, whose outcomes distinguish correctly and incorrectly classified churners and non-churners.

2.7. Evaluation Measures

  • The seven metrics are calculated from true positives, true negatives, false positives, and false negatives, with churn encoded as 1 and non-churn as 0.Accuracy is the proportion of correctly classified instances.Source: §2.7 · p. 8
  • Precision measures the share of predicted churners who actually churn, while recall measures the share of actual churners correctly identified.F1-score is the harmonic mean of precision and recall.Source: §2.7 · p. 8
  • Specificity measures the fraction of actual non-churners correctly identified, while ROC-AUC measures discrimination across classification thresholds.PR-AUC evaluates how well the model identifies positive cases.Source: §2.7 · p. 8

2.8. Coding and Experimental Environment

The experiments used a fixed software and hardware environment, a tenure-based four-client partition, and stratified train-test splits.

2.8. Coding and Experimental Environment

  • Experiments ran on Windows Server 2022 with an Intel Xeon Silver 4214R, 64 GB RAM, and 1 TB storage.The reported setup also used Python 3.12.Source: §2.8 · p. 8
  • The dataset was partitioned into four non-IID client shards based on customer tenure.The same passage reports an 80% training and 20% testing split using stratified sampling.Source: §2.8 · p. 8
  • Experiments used random seed 42; the listed libraries include PyTorch, Opacus, Flower, scikit-learn, and XGBoost.The passage states that the complete experiment code is available in a GitHub repository.Source: §2.8 · p. 8

3. EXPERIMENTAL RESULTS

Across two datasets, FedProx outperformed FedAvg without differential privacy, while DP-model comparisons differed by dataset. The experiments evaluated models separately using seven metrics.

3. EXPERIMENTAL RESULTS

  • The experiments treated the two datasets independently and evaluated models using seven metrics, with ε set to 1.0 for differentially private federated learning.Table 2 summarizes hyperparameter settings; Table 3 and Table 5 present model comparisons for Dataset-1 and Dataset-2, respectively.Source: §3 · p. 8, Table 2 · p. 9, Table 3 · p. 9

3.1. Results on Dataset-1

  • On Dataset-1, FedProx improved accuracy from 91.69% to 91.95% and F1-score from 91.28% to 91.50%; DP-FedProx remained comparable to non-DP federated models.DP-FedProx achieved 91.61% accuracy, 91.19% F1-score, 97.35% ROC-AUC, and 96.92% PR-AUC; FedProx also had similar multi-metric performance to centralized LR, with higher recall.Source: §3.1 · pp. 9–10, Table 4 · p. 10, Figure 2 · p. 10

3.2. Results on Dataset-2

  • On Dataset-2, the local LR model achieved the highest overall performance, but each local model used only its own data and could not support cross-operator collaboration.The authors attribute the larger centralized-LR versus FedProx gap on Dataset-2, relative to Dataset-1, likely to greater dataset difficulty and non-IIDness.Source: §3.2 · pp. 10–11
  • On Dataset-2, FedProx improved accuracy by 0.78% and F1-score by 3.79% over FedAvg, while DP-FedAvg outperformed DP-FedProx.Table 6 reports DP-FedProx improvements over FedProx of 1% in accuracy and 1.06% in F1-score; the radar comparison says centralized LR outperformed FedProx on every metric.Source: §3.2 · pp. 11–12, Figure 3 · p. 11, Table 6 · p. 12

4. Comparison with Other Studies

The authors identify two prior telecom churn studies using federated learning, but say their differing privacy and evaluation setups prevent direct performance comparisons.

4. Comparison with Other Studies

  • The two identified prior studies differ in privacy or evaluation setup, so the authors compare framework characteristics rather than prediction performance.One prior framework stored raw customer data centrally, while the other did not use differential privacy and relied on a private dataset.Source: §4 · p. 12, Table 7 · p. 12

5. Data Privacy Guarantee

The framework combines federated learning with DP-SGD to protect customer information during training and in shared updates. With ε = 1 and δ = 10^-5, the authors report prediction performance close to non-DP FedProx.

5. Data Privacy Guarantee

  • Federated learning keeps raw customer information local, while DP-SGD adds Gaussian noise to clipped gradients during local training.The resulting model updates are then sent to the central server.Source: §5 · p. 12
  • For DP-FedAvg and DP-FedProx, the authors set ε = 1 and δ = 10^-5 and report performance close to non-DP FedProx.The authors characterize ε = 1 as strong privacy protection, citing prior recommendations that ε ≤ 10 can provide meaningful guarantees.Source: §5 · p. 13

6. Bridging the Gap: Why DP based FedProx and FedAvg Achieve Comparable Performance

DP-FedAvg and DP-FedProx perform comparably in this setting, where tenure-based client groups have mild heterogeneity and the proximal correction is small relative to DP noise. SHAP analysis shows that the privacy-preserving global model can rely on different features than the centralized model, across both datasets.

6. Bridging the Gap: Why DP based FedProx and FedAvg Achieve Comparable Performance

  • The small difference between DP-FedAvg and DP-FedProx reflects mild client heterogeneity and a proximal correction small compared with DP-SGD noise.With tenure-based client groups, local and global objectives differ relatively little, limiting the proximal term’s benefit; the paper notes it may be more useful with more heterogeneous clients.Source: §6 · p. 13

6.1. Variable Importance via SHAP Analysis

  • SHAP analysis compared centralized ChurnNet with the FedProx+DP global model (ϵ = 1.0) to examine how training method changes feature importance.SHAP values were estimated on the preprocessed feature space using a 200-sample background reference set and held-out explanations.Source: §6.1 · p. 13

6.1.1. SHAP Analysis for Centralized Model

  • Dataset-1’s centralized model has closely clustered top-10 mean SHAP values (0.0371–0.0439), with usage-volume features dominating rather than one feature standing out.The passage attributes this flat pattern as likely related to the dataset’s 101 attributes spreading predictive information across related usage measures.Source: §6.1.1 · p. 13, Table 8 · p. 14
  • Dataset-2’s centralized model ranks tenure first (mean SHAP = 0.1175), about 34% above the second-ranked feature and over five times the tenth-ranked feature.Billing and contract features follow tenure; the passage links long-term contracts with lower churn risk and their absence with a risk signal.Source: §6.1.1 · p. 13, Table 9 · p. 14

6.1.2. SHAP Analysis for DP-FedProx Global Model

  • For Dataset-1, FedProx+DP retains a flat usage-and-revenue importance pattern (mean SHAP 0.0286–0.0356), but its top features shift toward usage changes and equipment information.The model’s top 10 include eqpdays, change mou, totcalls, and avgrev, replacing adjqty, adjmou, and mou peav Mean from the centralized model.Source: Table 10 · p. 15, §6.1.2 · p. 15
  • For Dataset-2, FedProx+DP ranks MonthlyCharges and TotalCharges highest, moves tenure to third, and replaces centralized contract features in the top 10 with service features.Across both datasets, the strongest predictors can shift between centralized and privacy-preserving federated models, so centralized feature importance may not describe the deployed model’s reliance.Source: Table 11 · p. 15, §6.1.2 · pp. 15–16

7. Discussion

Across two datasets, FedProx outperforms FedAvg, while DP-FedProx remains close to non-private FedProx and competitive with centralized models. Results vary by dataset, and feature rankings shift between centralized and privacy-preserving federated models.

7. Discussion

  • FedProx outperformed FedAvg on both datasets, while DP-FedProx stayed close to FedProx on Dataset-1 and achieved competitive results on Dataset-2.Dataset-1 used ε = 1 for DP-FedProx and DP-FedAvg; the paper reports only a small decrease in accuracy and F1-score for DP-FedProx versus FedProx.Source: §7 · p. 16
  • On Dataset-1, federated models had performance patterns close to the best centralized model, whereas on Dataset-2 they performed lower.The paper describes Dataset-2 as more challenging, with both centralized and federated models achieving lower performance; stronger non-IID effects may make global convergence harder.Source: §7 · pp. 16–17
  • The small DP-FedAvg–DP-FedProx gap is attributed to mild heterogeneity under tenure-based partitioning and proximal correction small relative to DP-SGD noise.The proximal term may offer more practical benefit when client data are more heterogeneous.Source: §7 · p. 17
  • For both datasets, the most important feature rankings differ between centralized and FedProx+DP models, so centralized feature importance may not match the deployed model.The discussion recommends evaluating interpretability on the final deployed federated model itself.Source: §7 · p. 17
  • The authors conclude that DP-FedProx balances prediction performance and privacy while allowing telecom operators to train a shared model without exchanging raw customer data.They report strong privacy protection at ε = 1 with only a small reduction in prediction performance.Source: §7 · p. 17

8. Conclusions

The paper proposes a DP-FedProx framework for collaborative churn prediction across heterogeneous telecom networks. Results on two public datasets show competitive performance and privacy protection, with FedProx outperforming FedAvg and differential privacy causing only a small performance decrease.

8. Conclusions

  • The proposed framework combines FedProx with DP-SGD and partitions clients by customer tenure to address non-IID data while protecting privacy.It is intended to let multiple telecom operators train a shared prediction model.Source: §8 · p. 17
  • Across two public telecom churn datasets, DP-FedProx achieved competitive prediction performance and preserved privacy; FedProx outperformed FedAvg, and adding DP caused only a small decrease in performance.The authors present the framework as a practical option when operators cannot share raw customer data because of privacy regulations, security requirements, or business confidentiality.Source: §8 · p. 17
  • The paper identifies adaptive privacy-budget allocation, personalized federated learning, and transformer-based federated models as future research directions.Source: §8 · p. 17
Loading 2609.12470v1…