Prediction of consumers refill frequency of LPG: A study using explainable machine learning

Area of Science:

• Data Science
• Public Policy
• Machine Learning

Background:

• The Pradhan Mantri Ujjwala Yojana (PMUY) aimed to provide LPG connections to rural Indian women, but faced challenges with sustained refill adoption.
• This impacted the program's sustainability and effective targeting of beneficiaries.

Purpose of the Study:

• To develop a predictive model for beneficiary LPG refill frequency using Explainable Machine Learning (XAI).
• To enhance the targeting strategy and sustainability of the PMUY program.

Main Methods:

• An enhanced stacked Support Vector Machine (SVM) model (ISS) was proposed and compared against Random Forest, SVM-RBF, Naive Bayes, and Decision Tree models.
• Performance was evaluated using accuracy, sensitivity, specificity, Cohen's Kappa, ROC, and AUC, with 10-fold cross-validation.
• Explainable AI (XAI) techniques were employed to understand feature importance and model interactions.

Main Results:

• The proposed ISS model achieved the best overall accuracies across various data splits (50-50, 66-34, 80-20).
• The XAI model provided insights into feature contributions, aiding in understanding prediction drivers.
• The study demonstrated the effectiveness of XAI in analyzing user behavior for policy interventions.

Conclusions:

• The developed XAI-powered prediction model can effectively anticipate LPG refill frequency among PMUY beneficiaries.
• This approach offers a valuable tool for optimizing beneficiary targeting and informing policy interventions for program sustainability.
• The findings support the use of advanced machine learning and XAI in public policy for improved social program outcomes.