Hybrid machine learning modelling with explainability for predicting case delays and durations in Indian lower courts

This research leverages an automated machine learning (AutoML) framework for predicting case durations and delays, offering valuable assistance and insights to judicial stakeholders and administrators. Utilising a large dataset of cases from lower courts across India, this research employs a tree-based ensemble model enhanced with explainable artificial intelligence (XAI) techniques to forecast case delay and durations with an emphasis on interpretability. Focusing exclusively on data available at the time of case filing, key variables influencing case length and delays, such as case type, existing backlog, and jurisdiction type, are systematically analysed, and multiple predictive models are constructed. We replace a monolithic regression model with two dedicated models, one for delayed cases and another for on-time cases, significantly reducing prediction error, lowering RMSE from 8.33 to 4.90 months for on-time cases and from 23.32 to 13.79 months for delayed cases. A separate classification model augments the regression model for delay prediction, achieving 80% accuracy with a weighted F1 score of 84.1%. Furthermore, a hybrid classification-regression framework demonstrates potential for improved delay and duration prediction in end-to-end simulations. The models demonstrate robust predictive capability, underscoring the potential of AI algorithms in enhancing efficiency and timeliness in judicial processes. Moreover, the application of XAI provides insightful explanations for predictions, ensuring that stakeholders comprehend the reasoning behind predicted durations. The predictive models can support case management, promote equitable judicial processes by identifying areas prone to delays, and inform policies aimed at enhancing the efficiency, timeliness, and transparency of the Indian judiciary. The cleaned dataset of 22.4 million criminal cases and the AutoML model code are made available for further research and reproducibility of results.