SchemeSetu: A Hybrid Explainable AI Framework for Personalised Government Welfare Scheme Recommendation Using Rule-Based Screening, Semantic Search and Random Forest Ranking

Abstract

Many citizens qualify for government welfare schemes but never apply, because scheme information is scattered across departmental websites and eligibility conditions are written in administrative language. Keyword-based portals help only users who already know what to search for. This paper presents SchemeSetu, a full-stack web application that recommends welfare schemes based on a citizen’s profile and a plain-language statement of need. The system uses a four-stage hybrid pipeline: (1) a rule-based eligibility engine that removes schemes whose mandatory criteria are not met, (2) semantic search using the all-MiniLM-L6-v2 sentence encoder and cosine similarity, (3) a 14-feature Random Forest ranker, and (4) an explanation generator that lists the matched criteria, benefits, required documents and application guidance. The three scores are combined as 0.40 × eligibility + 0.35 × semantic + 0.25 × ML score and displayed as a “Recommendation Match”. The prototype is built with Flask, MySQL and Bootstrap, with 24 REST endpoints, an application tracker, saved schemes, feedback and an administrator console. It uses demonstration schemes that are clearly labelled. The Random Forest is trained on synthetic data, so its test metrics only confirm that the pipeline works and are not a measure of real-world accuracy. The paper also covers design rationale, computational complexity, limitations, responsible-use safeguards and a proposed protocol for real-data evaluation. The system gives informational recommendations only. Final eligibility is decided by the concerned government department.