Explainable AI Based Machine Learning Framework for Service Demand and Stress Prediction

An effective process for handling massive numbers of citizens' service requests in metropolitan cities necessitates efficient predictive systems. This work examines NYC 311 service request data from 2022 through 2024, evaluating the model using 2025 data to test its generalizability. Unlike traditional methods, which only consider the volume of requests, the approach considers past trends and operational stress. The time-series analysis accounts for weekly and seasonality trends in the demand for services. The SBPI (Service Backlog Pressure Index) is a new consideration. It measures the connection between unprocessed requests and the speed of service, making it an indicator of the system's operational stress. Periods of surges are determined based on a statistically derived threshold, and it is modeled as a binary classification problem. Machine learning models, including XGBoost and TabNet, are trained using temporal and backlog-based features. Experimental results show that XGBoost outperforms TabNet, achieving an accuracy of 92.73%, demonstrating its effectiveness in predicting service demand surges.