International Journal of Artificial Intelligence, Machine Learning and Intelligent Systems

Data-Driven Insights into Customer Churn: A Predictive Analytics Approach

Shreya Sarkate — 2026-01-21

Customer churn remains one of the most pressing challenges for the telecommunications industry, where intense competition and low switching barriers make customer loyalty increasingly fragile. Since acquiring new subscribers is substantially more expensive than retaining existing ones, early identification of customers likely to leave is vital for sustaining profitability, improving service quality, and ensuring long-term business resilience. Predictive analytics, particularly Machine Learning (ML), offers a powerful means of modeling customer behavior and uncovering the complex patterns that precede churn. A thorough data-driven framework for churn prediction utilizing several machine learning models, such as Logistic Regression, Support Vector Machine (SVM), Random Forest, and Gradient Boosting Machine (GBM), is presented in this study. A telecommunications dataset that includes important elements of customer demographics, service usage, billing, contractual features, and tenure characteristics is used to train and assess the models. Rigorous preprocessing—including feature scaling, handling of class imbalance, and encoding of categorical variables—is undertaken to ensure model robustness. Comparative performance analysis demonstrates that ensemble-based approaches, particularly Random Forest and GBM, consistently outperform linear and margin-based models in terms of accuracy, precision–recall balance, and F1-score. Feature importance interpretation reveals that variables such as contract type, tenure, payment method, total charges, and monthly expenditure exert the strongest influence on churn behavior. These insights highlight the interplay between service affordability, customer engagement duration, and perceived value—factors that heavily shape customer retention. The findings of this study not only validate the efficacy of ML-driven churn analytics but also provide actionable intelligence for telecom providers. By integrating predictive modeling with targeted retention interventions, such as personalized offers or contract optimization, companies can significantly reduce churn, enhance user satisfaction, and minimize revenue leakage. The proposed framework thus supports data-driven decision-making, aligning technical accuracy with strategic business impact.

A Full-stack Generative AI-powered Platform for Automated Voice-based Candidate Evaluation

Archana Kale — 2026-01-22

The evolution of artificial intelligence (AI) and natural language processing (NLP) has led to significant advancements in real-time human-computer interactions. One promising domain is AI-driven real-time voice interview platforms, which integrate automatic speech recognition (ASR), natural language understanding (NLU), and analysis of emotion or cognitive state to evaluate human speech. This review paper presents a critical synthesis of three selected research works: AI-enhanced interview simulation, transforming language education using AI-based speaking practice, and AI-based voice biomarker models for cognitive assessment. Together, these papers demonstrate the transformative potential of AI voice systems in assessing human communication, emotion, cognition, and performance. The review focuses on how these methods contribute to designing a real-time voice-based interview platform for automated evaluation and personalized feedback.

Balancing Data Utility and Individual Privacy: A Comparative Analysis of Zero-knowledge Proofs and Differential Privacy in AI-driven Ecosystems

Mohamed Ibrahim — 2026-02-04

Artificial intelligence (AI) ecosystems increasingly depend on large-scale data collection and analytics, but this intensifies the privacy-utility tension and exposes individuals to inference and re-identification risks. This paper comparatively analyzes two leading privacy-preserving technologies Zero-knowledge proofs (ZKPs) and differential privacy (DP) to clarify where each is most effective and what trade-offs they impose in AI-driven environments. ZKPs (including zk-SNARKs and zk-STARKs) enable parties to verify claims about data integrity or computational correctness without revealing the underlying sensitive data, making them suitable for verification-heavy workflows such as identity assertions, transactional compliance checks, and blockchain-adjacent systems. DP, in contrast, provides a formal privacy guarantee by adding calibrated noise to datasets or query outputs, thereby limiting privacy leakage while preserving useful aggregate patterns for analytics and model development. Using a comparative analytical approach across representative contexts (healthcare analytics, financial services, and AI model training), the results indicate that DP is more effective for statistical reporting, data sharing, and federated learning scenarios, whereas ZKPs provide stronger guarantees when the primary requirement is trustless verification and auditability. The analysis further finds that neither technique fully addresses the complete spectrum of privacy requirements alone. Accordingly, the paper recommends a hybrid privacy architecture that combines ZKPs and DP (optionally alongside complementary cryptographic methods such as homomorphic encryption) to deliver scalable, regulation-aligned privacy governance for modern AI ecosystems.

Developing an Accent Recognition System for Nigerian Native Languages Using Convolutional Neural Networks and Long Short-term Memory Models

Adeyemi Michael Oduwale — 2026-03-26

Automatic speech recognition (ASR) is a component of what is known as human/computer interaction, which allows voice assistants, transcription services, and even smart communication systems. Yet, the problem of limited recognition and accommodation to the accent variations can be ranked among the greatest issues of ASR systems. This issue is highly evident in a multilingual nation such as Nigeria, where there is a linguistic diversity, and numerous native accents have arisen that the systems in place struggle to categorize appropriately. These restrictions make speech technologies less inclusive and less useful, particularly to populations which are underrepresented.

In this study, it was filled in by constructing an accent detection system for the three main indigenous languages in Nigeria (Yoruba, Hausa, and Igbo) using deep learning methods.

Three models were proposed and compared: a convolutional neural network (CNN), a long short-term memory network (LSTM), and a CNN-LSTM hybrid architecture. The hybrid model was presented and tested by the SautiDB-Naija dataset containing audio of native speakers. The experimental findings revealed that CNN and LSTM had similar performance, whereas the CNN-LSTM hybrid model had a higher performance with the validation accuracy of 92.3% and F1-score of 91.3, which indicates the benefit of using convolutional layers in extracting spectral features and applying the recurrent layers in the sequencing of time features. This study brings about the development of inclusive speech technology and sets the stage for future studies which could assist more African accents in ASR systems.