Source-linked AI summary
Customer churn prediction in telecom using machine learning and social network analysis in big data platform
Abdelrahim Kasem Ahmad, Assef Jafar, Kadan Aljoumaa
TL;DR
Telecom operators need to predict customer churn because churn directly affects revenue and retention is economically important. The paper builds a Spark-based churn model using engineered statistical and social-network features with tree-based algorithms. Adding SNA features raises AUC from 84% to 93.3%, with XGBoost selected as the classification algorithm, while performance on a later-period dataset falls to 89% AUC.
Problem
Telecom customer churn directly affects company revenue, motivating prediction of customers likely to leave and identification of factors associated with churn.
Method
The paper uses Spark, engineered statistical and social-network features, and four tree-based algorithms to predict churn from SyriaTel customer data.
Results
93.3% AUC was achieved after adding SNA features, compared with 84% using statistical features alone; XGBoost outperformed the other tested algorithms.
Takeaways & Limitations
Social Network Analysis features enhance telecom churn prediction by providing different information about customers from the social perspective.
Takeaways & Limitations
On a new dataset from different periods without proactive marketing action, XGBoost achieved 89% AUC, and the authors attribute the decrease possibly to non-stationary data requiring periodic retraining.
Abstract
from arXiv · showhide
Customer churn is a major problem and one of the most important concerns for large companies. Due to the direct effect on the revenues of the companies, especially in the telecom field, companies are seeking to develop means to predict potential customer to churn. Therefore, finding factors that increase customer churn is important to take necessary actions to reduce this churn. The main contribution of our work is to develop a churn prediction model which assists telecom operators to predict customers who are most likely subject to churn. The model developed in this work uses machine learning techniques on big data platform and builds a new way of features' engineering and selection. In order to measure the performance of the model, the Area Under Curve (AUC) standard measure is adopted, and the AUC value obtained is 93.3%. Another main contribution is to use customer social network in the prediction model by extracting Social Network Analysis (SNA) features. The use of SNA enhanced the performance of the model from 84 to 93.3% against AUC standard. The model was prepared and tested through Spark environment by working on a large dataset created by transforming big raw data provided by SyriaTel telecom company. The dataset contained all customers' information over 9 months, and was used to train, test, and evaluate the system at SyriaTel. The model experimented four algorithms: Decision Tree, Random Forest, Gradient Boosted Machine Tree "GBM" and Extreme Gradient Boosting "XGBOOST". However, the best results were obtained by applying XGBOOST algorithm. This algorithm was used for classification in this churn predictive model.
Introduction
Telecom churn threatens revenue in a highly competitive market, motivating predictive models that identify likely-to-leave customers early. This work combines big-data processing, engineered behavioral and social-network features, and tree-based machine learning for SyriaTel.
- Motivation: Retaining existing customers is presented as more profitable and less costly than acquiring new ones, making churn reduction a central telecom strategy.The paper frames churn as customer movement from one provider to another.
- Motivation: Early prediction of likely churners could provide a potentially large additional revenue source.
- Approach: The study processes SyriaTel data on a big-data platform to engineer features and build a churn prediction model.The dataset includes customer information over nine months and combines structured, semi-structured, and unstructured data.
- Approach: Customer social-network features, including centrality, similarity, and connectivity, add information about customers and enhance AUC results.
- Approach: The study compares Decision Tree, Random Forest, GBM, and XGBoost methods after data preparation, feature engineering, and feature selection.
- Evaluation: The model was evaluated on a new dataset, tested for its impact on churn decisions, and deployed to production.
Related work
Prior churn studies used varied machine-learning, data-mining, sampling, and classification methods, but this paper identifies limited Syrian telecom research and limited raw-data feature engineering.
- Existing approaches: Earlier work applied single data-mining methods or compared several strategies for telecom churn prediction.
- Existing approaches: Studies explored neural networks, genetic programming with AdaBoost, Random Forest, rough-set classification, and sampling techniques across telecom datasets.
- Class imbalance: Undersampling outperformed the other tested techniques in one comparison of churn-model balancing methods using AUC and Lift.
- Research gap: The authors report no prior research addressing this problem in a Syrian telecommunications company.
- Research gap: The paper emphasizes feature engineering from raw data, big-data preparation, and comparison of four tree-based machine-learning algorithms.
Data set
The SyriaTel dataset combines customer, location, complaint, network, transaction, and mobile-device information to represent telecom behavior for churn modeling.
- Data sources: Customer data covers services, contracts, offers, packages, subscriptions, and CRM attributes such as GSMs, demographics, and location.
- Data sources: Tower and complaint data provide transaction locations and records of coverage, offer, package, and other telecom problems.
- Data sources: Network logs record internal internet, call, and SMS sessions, including session duration and call-ending status.
- Data sources: Call-detail records contain charging information for calls, SMS, MMS, and internet transactions.
- Data sources: Mobile IMEI data describes device brand, model, type, and whether the device uses dual or mono SIM.
Data exploration and challenges with SyriaTel dataset
The SyriaTel data required extensive exploration and big-data processing because of heterogeneous formats, extreme dimensionality, missingness, low-variation variables, and class imbalance.
- Exploration: Spark exploration found many low-variation variables, including 50% of numeric variables with one or two discrete values and 33% of categorical variables with one value.
- Scale and heterogeneity: The raw behavioral dataset spans calls, SMS, MMS, internet, complaints, network data, IMEI, and charging across nine months and exceeds 70 Terabytes.
- Scale and heterogeneity: The data combines structured, semi-structured, and unstructured files from multiple systems, making a big-data platform necessary for processing its size and formats.
- Class imbalance: The generated dataset is unbalanced because its class distribution is not homogeneous, while machine-learning algorithms may not account for class balance.
- Class imbalance: Churn customers represent about 5% of the SyriaTel dataset.
- Dimensionality and missingness: Using all collected sources would produce more than ten thousand columns per customer, with missing values caused by differing subscriptions and all-null configuration columns.
Proposed churn method
The proposed churn method combines a customized big-data platform, Spark-based feature engineering, and tree-based classification with customer social-network features. It selects Parquet for efficient storage and uses SNA measures alongside statistical customer features.
- Big-data platform: Hortonworks Data Platform was customized into the SYTL-BD framework to store, process, manage, develop, and monitor the churn-prediction workflow.The framework included HDFS, Spark, YARN, Zeppelin, Ambari, and Ranger.
- Data acquisition and storing: The system collected nine months of structured, semi-structured, and unstructured customer data and moved it into HDFS for processing.Flume handled CSV, text, JSON, and XML files, while SQOOP transferred structured data from relational databases.
- Feature engineering: Parquet gave the best file-format results, while Parquet with Snappy Compression gave the best space utilization.The comparison transformed 1 terra byte of CSV data into Parquet and Avro using Spark and considered three compression scenarios.
- Feature engineering: Feature engineering aggregated numerical customer data and counted distinct categorical values, producing statistical features from calls, SMS, MMS, and internet records.The process was computationally demanding because of the huge number of columns.
- Social Network Analysis: The social network supplied PageRank, SenderRank, degree, connectivity, clustering, and similarity features for individual customers.Similarity calculations used Jaccard and Cosine similarity, excluding customer pairs without mutual friends to reduce complexity.
- Social Network Analysis: SNA features had a positive effect alongside statistical features and were used to enhance churn prediction.The social graph features were calculated with Spark’s Graph Frame library.
- Classification: The model evaluated Decision Tree, Random Forest, Gradient Boost Machine Tree, and XGBoost using a 70% training split, a 30% test split, and 10-fold cross-validation.Training data were rebalanced through oversampling or random undersampling because the churn class was unbalanced.
Results and discussion
The results show that feature history and type materially affect churn prediction: statistical features benefited from broader recent history, whereas SNA features worked best over the latest four months. Combining both feature types produced the strongest performance, with XGBOOST selected as the final classifier, while the analysis also identified behavioral and network characteristics associated with churn.
- Sliding windows and feature types: 84% was the highest AUC achieved using statistical features alone, with no improvement from adding the oldest three months.Increasing statistical-feature training data improved performance until the oldest three months were included.
- Sliding windows and feature types: 75.3% was the highest AUC achieved using SNA features alone, with the best social-graph window covering the last four months before baseline.Adding older data adversely affected SNA-based model performance.
- Sliding windows and feature types: 93.3% was the maximum AUC after combining statistical and SNA features, a significant improvement over using statistical features alone.The selected feature windows used the last six months for statistical features and the last four months for SNA features.
- Algorithm comparison: 93.3% was XGBOOST’s AUC, outperforming GBM at 90.89%, Random Forest at 87.76%, and Decision Tree at 83%.XGBOOST was therefore chosen as the classification algorithm for the predictive model.
- Feature analysis: Adding SNA features changed feature importance rankings, with MTN Cosine similarity becoming most important and several behavioral and network indicators also ranking highly.Other prominent indicators included time since the last outgoing transaction, total balance, Average Radio Access Type, Local Cluster Coefficient, and transactions involving another operator.
Conclusion
The study concludes that Social Network Analysis features enhance telecom churn prediction, with XGBOOST producing the strongest reported results. Performance decreased on data from different periods, suggesting periodic retraining may be needed.
- XGBOOST achieved the best AUC among the evaluated algorithms, while GBM ranked second and Random Forest and Decision Tree ranked third and fourth.
- 89% AUC was obtained when XGBOOST was evaluated on a new dataset from different periods without proactive marketing action.
- The model may require training during each period because the observed decrease could reflect non-stationary data.
- Social Network Analysis features enhanced telecom churn-prediction results.The conclusion attributes this contribution to the predictive value of customer-network information.