Machine learning reduces questionnaire items while maintaining reliability and validity

Traditional long questionnaires often suffer from respondent fatigue, leading to decreased completion rates and compromised data quality. This study introduces a novel algorithmic item-reduction methodology, named NNAQ, designed to streamline high-dimensional survey instruments while strictly preserving both downstream classification capabilities and psychometric construct validity. Operating as a structured case study, the proposed multi-tiered framework strategically orchestrates linear filters, non-linear tree-embedded layers, and information-theoretic criteria. This hybrid architecture allows different mathematical paradigms to cross-validate and error-correct each other, successfully overcoming the limitations of traditional single-stage dimension reduction methods. This study evaluated the framework empirically using an expanded field dataset collected from 1,022 citizens. The experimental results demonstrate that the original 96-item questionnaire can be reduced to a parsimonious 34-item instrument, representing a 64.58% reduction in questionnaire length. Downstream evaluation via machine learning classifiers indicates that the predictive accuracy loss is kept under 5 percent, while comprehensive confirmatory factor analysis confirms that the shortened scale yields superior parsimonious fit indices and stable internal reliability. Furthermore, rigorous bootstrapping resampling and benchmarking experiments against alternative reduction methods validate the exceptional structural stability and empirical superiority of our pipeline. Overall, this research provides a robust, data-efficient, and scalability-tested framework for human-centered digital assessment, effectively mitigating respondent fatigue while securing high-fidelity data quality.