An AI-Driven University Recommendation System for Nepalese Study-Abroad Aspirants with Peer Consultation Support

Abstract

Every year, over 100,000 Nepalese students pursue higher education abroad, yet university selection remains manual, fragmented, and heavily reliant on commercial consultancies with potential partner-institution bias. This paper presents UNIFINDER, a web platform integrating machine-learning-based university matching with peer-to-peer consultation. Built on the MERN stack with a Python/Flask microservice, the system hosts two regression models trained on scraped data: an XGBoost Regressor for general international applicants, and a second XGBoost Regressor combined with K-Means clustering for Nepalese applicants, ranking institutions by Nepalese student concentration. Data collection used an automated Puppeteer.js and Cheerio.js pipeline; a Zoom API integration (via Ngrok) enabled video consultations with peers abroad. An initial evaluation on a single random split suggested strong accuracy (R2=0.93 and 0.81). However, an audit revealed institution-level data leakage. Under corrected, university-grouped cross-validation, both models showed weaker generalization to unseen institutions (mean R2 of 0.18 and 0.31 across five folds), indicating dataset size limitations (131 records across 72 institutions). We report this transparently, identifying dataset expansion as key future work. Combining data-driven matching with peer guidance aims to reduce reliance on third-party consultancies, though this benefit remains to be directly evaluated.