Technical Debt (TD) refers to the long-term costs incurred due to suboptimal decisions made for short-term gains. Algorithm Debt (AD), is a TD type that arises from the suboptimal implementation of algorithms, that can result in model degradation and poor scalability. Machine and Deep Learning (ML/DL) systems are particularly susceptible to AD due to the complexity of their methods and dependencies on large-scale data. Despite the significance of AD in ML/DL systems, its causes, effects, and mitigation strategies remain underexplored, creating a knowledge gap. To address this, we investigated the causes, effects, and mitigation strategies of AD in ML/DL systems. Using a mixed-methods approach, we conducted a systematic review of 44 primary studies, interviewed 21 ML/DL practitioners, complemented by analysis of 65 questionnaire. We also conducted experiments with ML/DL models and their embeddings to empirically evaluate their performance in the automated detection of AD. Our findings suggest that the causes of AD can be categorised into three key areas: data quality issues, ML knowledge gaps, and model training. Our experiments further revealed that Logistic Regression model trained with AD-specific features achieved the highest F1-score of 54%, outperforming other techniques. This study contributes to software engineering by proposing a framework that outlines the causes, effects, and mitigation strategies of AD in ML/DL systems. The framework serves as a guideline for practitioners and provides insights for developing effective AD management strategies in ML/DL systems.

Iko-Ojo is a PhD student under the supervision of Dr. Chirath Hettiarachchi at the Australian National University, in the CECC School of Computing. He is co supervised by Dr. Fatemeh Fard from University of British Columbia, Prof. Alex Potanin, and Prof. Hanna Suominen. His main research interests are in Software Engineering, Technical Debt, and ML/DL. He is currently working on Algorithm Debt in ML/DL Software (e.g., Deep Learning frameworks).