Inferring hidden structure in mobile network performance data with noisy net promoter scores using a probabilistic graphical model

Understanding customer satisfaction in the context of mobile network performance is helpful when designing reliable cellular networks to retain customers and drive customer loyalty. Using Infer.NET, we propose a probabilistic graphical model that infers hidden structure in network key performance indicators using noisy customer survey responses. Our model uses real-world net promoter score survey data, network session data consisting of sites visited by respondents, and network performance data from active sessions. The model learns hidden structure in the network performance data that represent good and bad quality of experience. The discovered properties are consistent with industry-recommended signal strength and quality levels for UMTS and LTE standards. Furthermore, our methodology allows us to estimate a daily network performance for each site, which helps to identify problem areas in the network. Due to the subjective nature of survey data, our model also estimates the overall asymmetric noise associated with the surveys.