Machine learning algorithms for formative assessment: reducing classroom workload

Abstract

This study analyzes the potential of machine learning algorithms to support formative assessment in the Basic Elementary level of schools in Ambato (Tungurahua, Ecuador), aiming to reduce teachers’ operational workload without undermining pedagogical quality. A mixed-method, non-experimental applied design was implemented, combining teacher questionnaires and semi-structured interviews with the use of open educational datasets and anonymized student performance records to train supervised classification models. Among the tested algorithms, the random forest model achieved the best predictive performance for anticipating formative achievement levels, with accuracy and discrimination metrics comparable to recent early warning and educational data mining studies. Findings indicate that automating data aggregation and report generation may reduce between 1.5 and 2.5 weekly hours devoted to grading and reporting, while improving the timeliness and precision of feedback provided to students. Teachers reported a generally positive perception of AI as a support tool to prioritize at-risk students and to organize assessment information, but they also raised concerns about model transparency, labeling risks and data protection, in line with national guidelines that call for ethical and pedagogically grounded uses of AI in education. The study concludes that integrating machine learning into formative assessment is feasible and relevant in Ambato, provided that it is implemented gradually, contextually and under clear frameworks of governance and teacher professional development.