An Intelligent Government Complaint Prediction Approach

Abstract

Recent advances in machine learning (ML) bring more opportunities for greater implementation of smart government construction. However, there are many challenges in terms of government data application due to the previous nonstandard records and man-made errors. In this paper, we propose a practical intelligent government complaint prediction (IGCP) framework that helps governments quickly respond to citizens' consultations and complaints via ML technologies. In addition, we put forward an automatic label correction method and demonstrate its effectiveness on the performance improvement of intelligent government complaint prediction task. Specifically, the central server collects the interaction records from users and departments and automatically integrates them by the label correction approach which is designed to evaluate the similarity between different labels in data, and merge highly similar labels and corresponding samples into their most similar category. Based on those refined data, the central server quickly generates accurate solutions to complaints through text classification algorithms. The main innovation of our approach is that we turn the task of government complaint distribution into a text classification problem which is uniformly coordinated by the central server, and employ the label correction approach to correct redundant labels for training better models based on limited complaint records. To explore the influences of our approach, we evaluate its performance on real-world government service records provided by our collaborator. The experimental results demonstrate the prediction task which uses the label correction algorithm achieves significant improvements on almost all metrics of the classifier.