Teaching Data Cleaning in Machine Learning: A Misconception-Driven Pedagogical Framework for Explainable and Ethical Data Preprocessing
TL;DR
This paper proposes a four-phase instructional sequence — Motivational Grounding, Mechanism Diagnosis, Algorithm-Context Mapping, Algorithm-Context Mapping, and Explainability and Ethical Reflection — operationalized through a multi-dimensional Instructor Cleaning Strategy Matrix, embedding explainability and ethical reflection as a mandatory fourth phase of instruction.
Abstract
Data cleaning — the process of detecting and remedying missing, erroneous, and inconsistent records in raw datasets — is widely recognized as the most time-consuming and consequential stage of the machine learning preprocessing pipeline. Yet despite its operational centrality, how data cleaning should be taught to undergraduate computer science students has received remarkably little systematic attention in the machine learning education literature. This paper addresses that gap by developing a theoretically grounded pedagogical framework for instructors teaching data cleaning in higher education contexts. Drawing on threshold concept theory, constructivist learning principles, and the documented landscape of machine learning education challenges, we identify three pervasive misconceptions: the deletion reflex, the imputation neutrality fallacy, and the outlier criminalization fallacy. We propose a four-phase instructional sequence — Motivational Grounding, Mechanism Diagnosis, Algorithm-Context Mapping, and Explainability and Ethical Reflection — operationalized through a multi-dimensional Instructor Cleaning Strategy Matrix. The framework is empirically grounded through a case study applying the full technique repertoire to a real-world academic email dataset of over 10,000 records spanning eleven years. A cross-method comparison demonstrates that imputation methods produced NRMSE values ranging from 0.08 to 0.45, and anomaly detection methods flagged between 185 and 412 records on identical data. Implications for machine learning curriculum design, assessment, and data ethics literacy are discussed. By embedding explainability and ethical reflection as a mandatory fourth phase of instruction — grounded in GDPR, the EU AI Act, and ISO 8000 requirements — the framework equips learners to navigate the data governance demands of contemporary technology-enhanced professional environments.
