Abstract: Because of fierce competition in telecommunication (telco) industry,it is crucial to reduce customer complaint rate and improve customer services to improve competitive advantages for telecommunication companies.Thus,accurately predicting the complaint behaviors to reduce the complaint rates becomes one of the most important tasks for telco operators.Traditional complaint prediction models only focus on classification algorithms and artificial feature selection and do not release the full power of telco big data.In this paper,we proposed a big data-driven complaint prediction model on the Hadoop/Spark platform using efficient parallel random forests.To better explore the performance of the proposed method,we performed feature engineering not only on all data from business support system (BSS) and operations support system (OSS),but also on those from the customer service records (CSR).Moreover,several useful graph-based features and second-order features between the relationship of users were designed and used to enhance the predictive performance.Experiment results based on the practical data of the telco operator in Shanghai show that using more data sources and high dimension data to train the complaint prediction models make the prediction accuracy higher than the state-of-the-art algorithms.Based on the result,we took comfort measures on target users,which can make the lo-wer complaint rate of users and bring significant business value to the operator.

Key words: Big data,Complaint prediction model,Feature engineering,Second-order feature,Graph-based features,Random forest