A study A Neural Network approach to identify features associated with multidimensional poverty in rural India, featured in MoSPI publication Sarvekshana (April 2026) applies a machine learning approach to analyse rural multidimensional poverty using HCES 2022–23 data. It constructs an Adjusted Multidimensional Poverty Index (MPI) based on nine available indicators and compares the performance of traditional logistic regression with Deep Neural Networks (DNN), which is an advanced AI model .
The analysis, covering over 1.5 lakh rural households, finds that DNN models outperform traditional regression across key metrics such as accuracy, sensitivity, and AUC, indicating a stronger ability to correctly identify households that are MPI poor (deprivation score > 0.33).
Importantly, the two approaches surface different layers of poverty drivers. The neural network model captures interaction effects, identifying low education (below primary), absence of economic activity, and specific household characteristics (such as marital status of the household head) as dominant predictors.
In contrast, logistic regression highlights structural variables like geographic location (high-poverty states), digital exclusion (lack of internet), and education levels.
The findings show that poverty is shaped not just by individual factors, but by overlapping and reinforcing deprivations, which non-linear models like DNN are better equipped to capture. This has implications for improving the precision of poverty identification and targeting.
Key Statistical and Analytical Benchmarks
Model Superiority: Deep Neural Networks (DNN) showed higher accuracy, sensitivity, and AUC compared to logistic regression.
Threshold for Poverty: Households with a weighted deprivation score above 0.33 were classified as MPI Poor.
Primary Drivers (Neural Network): Education below primary, no economic activity, and "never married" status of the household head.
Primary Drivers (Regression): High-poverty state residence, lack of internet, and education levels.
Digital Access Gap: Access to mobile phones and internet remains significantly lower among the MPI Poor.
Geographic Focus: Higher proportions of poor households identified in Jharkhand, Bihar, Chhattisgarh, UP, MP, Odisha, and Rajasthan.
What is Deep Neural Network?
A Deep Neural Network (DNN) is an advanced AI model that mimics how the human brain processes information. Instead of relying on simple, linear relationships, it uses multiple layers of artificial neurons to detect complex, non-linear patterns in data.
In practical terms, a DNN can analyse how different factors interact, not just individually, but in combination. For example, in rural poverty analysis, it can identify how limited education and lack of digital access together deepen deprivation, even if each factor alone seems less significant.
Its key advantage over traditional models like logistic regression is higher predictive accuracy. By learning these layered relationships, a DNN uncovers hidden drivers of outcomes that simpler models often miss, making it especially useful for complex, real-world problems like multidimensional poverty.
Policy Relevance
Enhances Targeting Precision: Transitioning from traditional regression to Neural Network models allows the government to identify vulnerable households with much higher accuracy, reducing inclusion and exclusion errors in social schemes.
Prioritises Digital Inclusion: The strong correlation between lack of internet and poverty suggests that programs like BharatNet are as essential for poverty reduction as traditional infrastructure.
Focuses on Educational Attainment: Since primary education is a top predictor of poverty, policy interventions must focus specifically on ensuring high school completion to break the cycle of intergenerational deprivation.
Addresses Geographic Inequality: Validating the high concentration of poverty in specific states (EAG states) supports the continued need for region-specific funding and development missions.
Leverages AI for Social Good: The study proves that AI/ML techniques can provide more nuanced insights into the interplay of social factors, helping NITI Aayog and MoSPI refine future survey methodologies and policy responses.
Follow the Full News Here (Page 23 to 45): MoSPI - Sarvekshana Journal Issue 120 (April 2026)