From Network Experience to Subscriber Retention: An Explainable AI Framework for Mobile Operators

Abdol Saleh; Faris B. Mismar; Ivan Maxmillian Putra Pasaribu; Suhelmy Syaifuddin

arxiv: 2606.04838 · v1 · pith:Z6XKN7TXnew · submitted 2026-06-03 · 💻 cs.NI

From Network Experience to Subscriber Retention: An Explainable AI Framework for Mobile Operators

Faris B. Mismar , Abdol Saleh , Ivan Maxmillian Putra Pasaribu , Suhelmy Syaifuddin This is my paper

Pith reviewed 2026-06-28 04:01 UTC · model grok-4.3

classification 💻 cs.NI

keywords churn predictionquality of experienceexplainable AImobile operatorssubscriber retentionmachine learning

0 comments

The pith

Subscriber quality of experience indicators predict churn more effectively than traditional network counters in mobile operators.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper develops a framework for predicting subscriber churn using machine learning and explainable AI on data from a large telco. It demonstrates that quality of experience metrics serve as stronger signals for churn than standard network performance counters. This approach provides actionable insights for operators to improve retention through better analytics focused on user experience rather than just network metrics. The framework is tested on real-world data with tens of millions of subscribers to show its robustness.

Core claim

Subscriber quality of experience (QoE) indicators provide stronger churn signals than traditional network counters alone, which reinforces the need for QoE-centric analytics in modern telco operations.

What carries the argument

An explainable AI and machine learning framework applied to subscriber QoE data for churn prediction.

If this is right

Operators can shift from network-counter based monitoring to QoE-focused models for better churn prediction.
Actionable insights from the model can guide interventions to retain subscribers.
The framework's longevity is supported by results on large-scale real data.
Future work can focus on improving predictability and operational deployment.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

Similar frameworks might apply to other service industries where experience metrics matter more than raw performance.
Without retraining, the model may not generalize if network conditions change significantly.
Integrating real-time QoE data could enable proactive churn prevention.

Load-bearing premise

Patterns learned from one large telco's data will identify reliable churn signals that apply to other operators and future periods without major adjustments.

What would settle it

Applying the trained model to data from a different mobile operator and checking if the QoE indicators remain the strongest churn predictors.

Figures

Figures reproduced from arXiv: 2606.04838 by Abdol Saleh, Faris B. Mismar, Ivan Maxmillian Putra Pasaribu, Suhelmy Syaifuddin.

**Figure 2.** Figure 2: Microsegments and dynamics of loyalty and value in the telco [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗

**Figure 3.** Figure 3: Churn rate increases as value contribution declines. [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗

read the original abstract

This article presents a framework for the prediction of subscriber churn in mobile operators also known as telecommunication operators (or telcos). This framework covers relevant aspects of data-driven approaches using explainable artificial intelligence and machine learning. To demonstrate the robustness of the framework, we implement it on real data from one of the globally leading telcos with tens of millions of subscribers and show results and actionable insights confirming the usefulness and longevity of the framework. Our results suggest that subscriber quality of experience (QoE) indicators provide stronger churn signals than traditional network counters alone, reinforcing the need for QoE-centric analytics in modern operations in telcos. We conclude with future research directions for improving churn predictability and operational deployment.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

Applies standard XAI and ML to churn prediction on one large telco's data and reports QoE metrics outperform network counters, but provides almost no method details or validation.

read the letter

This paper takes existing machine learning and explainable AI methods and runs them on churn prediction for a mobile operator. The main result is that quality-of-experience indicators appear to give stronger signals than the usual network counters, based on data from one global telco with tens of millions of subscribers.

What it does is show a working implementation at real scale and pull out some actionable insights for operators. That part has practical value for people who already work in telecom analytics and want to see how these tools behave on actual large datasets.

The paper does not introduce new techniques or derivations. It stays inside established XAI and ML practice and does not benchmark against specific prior churn models, so the advance over earlier work is hard to judge.

The soft spots are clear. The abstract and description give no information on model architecture, validation approach, error bars, or statistical tests. The claim that QoE metrics are stronger rests on a single operator's historical data with no cross-operator test, temporal hold-out, or domain adaptation check. That leaves the broader recommendation for QoE-centric analytics resting on an untested transfer assumption.

This is for readers already inside telecom operations or retention analytics who are looking for case studies rather than new methods. It shows straightforward engagement with the problem but does not move the technical frontier.

I would send it to peer review. The scale of the data makes the empirical demonstration worth referee time, provided the full paper supplies the missing method details and validation steps.

Referee Report

2 major / 2 minor

Summary. The paper presents an explainable AI and machine learning framework for predicting subscriber churn in mobile telecom operators. It implements the framework on real historical data from one large global telco with tens of millions of subscribers, reports actionable insights, and concludes that quality-of-experience (QoE) indicators supply stronger churn signals than traditional network counters alone.

Significance. If the empirical claims hold under proper validation, the work could encourage telcos to shift from counter-based to QoE-centric analytics for retention. The explicit use of explainable methods is a practical strength for operational trust and deployment.

major comments (2)

[Abstract and §4] Abstract and §4 (results): the claim that QoE indicators provide stronger churn signals than network counters is stated without any reported model architecture, cross-validation procedure, temporal hold-out, error bars, statistical significance tests, or data-exclusion criteria, so the strength of evidence cannot be assessed.
[§5 and conclusion] §5 (discussion) and conclusion: the recommendation that operators adopt QoE-centric analytics rests on patterns learned from a single telco's historical dataset; no cross-operator validation, domain-adaptation experiment, or multi-period temporal test is described to support transfer to other operators or future periods.

minor comments (2)

[§3] Notation for QoE features and network counters should be defined consistently in a table or appendix before first use.
[Figures 3-5] Figure captions should explicitly state the performance metric (e.g., AUC, F1) and the baseline comparator used in each panel.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments on validation and generalizability. We address each major comment below.

read point-by-point responses

Referee: [Abstract and §4] Abstract and §4 (results): the claim that QoE indicators provide stronger churn signals than network counters is stated without any reported model architecture, cross-validation procedure, temporal hold-out, error bars, statistical significance tests, or data-exclusion criteria, so the strength of evidence cannot be assessed.

Authors: Section 3 details the model architectures (gradient boosting ensembles and neural networks with hyperparameters) and the XAI component. Section 4 reports comparative results via performance metrics and SHAP values. To enable full assessment of evidence strength, we will revise §4 and the methods to explicitly describe the temporal hold-out cross-validation, error bars from repeated runs, statistical significance tests, and data exclusion criteria. revision: yes
Referee: [§5 and conclusion] §5 (discussion) and conclusion: the recommendation that operators adopt QoE-centric analytics rests on patterns learned from a single telco's historical dataset; no cross-operator validation, domain-adaptation experiment, or multi-period temporal test is described to support transfer to other operators or future periods.

Authors: The results are from one large operator's dataset, a common constraint due to data access. The framework is presented as general, but we agree transferability claims require caution. We will revise §5 and the conclusion to explicitly state this limitation, avoid overgeneralizing the recommendation, and identify cross-operator validation as future work. revision: yes

Circularity Check

0 steps flagged

No circularity; empirical framework self-contained

full rationale

The paper presents a data-driven ML/XAI framework for churn prediction and applies it to one telco's historical dataset to produce empirical results on QoE vs. network-counter signals. No equations, derivations, or uniqueness theorems are described that reduce any claimed prediction or result to a fitted parameter or self-referential definition by construction. The central claim is an observable outcome from the data rather than a tautological renaming or self-citation chain, making the derivation chain independent of its own outputs.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review provides no information on free parameters, axioms, or invented entities used in the framework.

pith-pipeline@v0.9.1-grok · 5660 in / 1068 out tokens · 38081 ms · 2026-06-28T04:01:55.281753+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

10 extracted references · 1 canonical work pages · 1 internal anchor

[1]

Key Performance Indicators (KPI) for Evolved Universal Terrestrial Radio Access Network (E-UTRAN)

3GPP. Key Performance Indicators (KPI) for Evolved Universal Terrestrial Radio Access Network (E-UTRAN)
[2]

Radio measurement collection for Minimization of Drive Tests (MDT); Overall description; Stage 2

3GPP. Radio measurement collection for Minimization of Drive Tests (MDT); Overall description; Stage 2
[3]

Autonomous Network Levels Evaluation Methodology

TM Forum. Autonomous Network Levels Evaluation Methodology
[4]

Harald Remmert , year = 2025, month=dec, url =

2025
[5]

, year = 1951, month=aug, Address =

Shapley, Lloyd S. , year = 1951, month=aug, Address =

1951
[6]

Minimizing age of information in vehicular networks , year=

Kaul, Sanjit and Gruteser, Marco and Rai, Vinuth and Kenney, John , booktitle=. Minimizing age of information in vehicular networks , year=
[7]

XGBoost: A Scalable Tree Boosting System

Tianqi Chen and Carlos Guestrin , year=. 1603.02754 , archivePrefix=

work page internal anchor Pith review Pith/arXiv arXiv
[8]

and Alammouri, Ahmad and Alkhateeb, Ahmed and Andrews, Jeffrey G

Mismar, Faris B. and Alammouri, Ahmad and Alkhateeb, Ahmed and Andrews, Jeffrey G. and Evans, Brian L. , journal=. Deep Learning Predictive Band Switching in Wireless Networks , year=
[9]

and Hoydis, Jakob , journal=

Mismar, Faris B. and Hoydis, Jakob , journal=. Unsupervised Learning in Next-Generation Networks: Real-Time Performance Self-Diagnosis , year=
[10]

2016 , Address =

Deep Learning , Author =. 2016 , Address =

2016

[1] [1]

Key Performance Indicators (KPI) for Evolved Universal Terrestrial Radio Access Network (E-UTRAN)

3GPP. Key Performance Indicators (KPI) for Evolved Universal Terrestrial Radio Access Network (E-UTRAN)

[2] [2]

Radio measurement collection for Minimization of Drive Tests (MDT); Overall description; Stage 2

3GPP. Radio measurement collection for Minimization of Drive Tests (MDT); Overall description; Stage 2

[3] [3]

Autonomous Network Levels Evaluation Methodology

TM Forum. Autonomous Network Levels Evaluation Methodology

[4] [4]

Harald Remmert , year = 2025, month=dec, url =

2025

[5] [5]

, year = 1951, month=aug, Address =

Shapley, Lloyd S. , year = 1951, month=aug, Address =

1951

[6] [6]

Minimizing age of information in vehicular networks , year=

Kaul, Sanjit and Gruteser, Marco and Rai, Vinuth and Kenney, John , booktitle=. Minimizing age of information in vehicular networks , year=

[7] [7]

XGBoost: A Scalable Tree Boosting System

Tianqi Chen and Carlos Guestrin , year=. 1603.02754 , archivePrefix=

work page internal anchor Pith review Pith/arXiv arXiv

[8] [8]

and Alammouri, Ahmad and Alkhateeb, Ahmed and Andrews, Jeffrey G

Mismar, Faris B. and Alammouri, Ahmad and Alkhateeb, Ahmed and Andrews, Jeffrey G. and Evans, Brian L. , journal=. Deep Learning Predictive Band Switching in Wireless Networks , year=

[9] [9]

and Hoydis, Jakob , journal=

Mismar, Faris B. and Hoydis, Jakob , journal=. Unsupervised Learning in Next-Generation Networks: Real-Time Performance Self-Diagnosis , year=

[10] [10]

2016 , Address =

Deep Learning , Author =. 2016 , Address =

2016