The Use of Machine Learning Algorithms in Recommender Systems: A Systematic Review

Recommender systems use algorithms to provide users with product or service recommendations. Recently, these systems have been using machine learning algorithms from the field of artificial intelligence. However, choosing a suitable machine learning algorithm for a recommender system is difficult because of the number of algorithms described in the literature. Researchers and practitioners developing recommender systems are left with little information about the current approaches in algorithm usage. Moreover, the development of a recommender system using a machine learning algorithm often has problems and open questions that must be evaluated, so software engineers know where to focus research efforts. This paper presents a systematic review of the literature that analyzes the use of machine learning algorithms in recommender systems and identifies research opportunities for software engineering research. The study concludes that Bayesian and decision tree algorithms are widely used in recommender systems because of their relative simplicity, and that requirement and design phases of recommender system development appear to offer opportunities for further research.

💡 Research Summary

The paper presents a systematic literature review that investigates how machine‑learning (ML) algorithms are employed within recommender systems (RS) and identifies gaps that merit further software‑engineering research. Following the PRISMA methodology, the authors queried four major bibliographic databases (IEEE Xplore, ACM Digital Library, Scopus, and Web of Science) for the period 2010‑2023 using combinations of “recommender”, “machine learning”, and “algorithm”. After de‑duplication and quality assessment, 124 primary studies were retained for detailed analysis. Each study was coded for algorithm type, application domain (e‑commerce, media streaming, education, etc.), evaluation metrics (accuracy, diversity, novelty, trust), and development phase (requirements, design, implementation, testing).

The quantitative synthesis reveals that Bayesian networks and decision‑tree‑based methods (including Random Forests and Gradient Boosting Machines) account for roughly 38 % of all reported algorithmic choices, making them the most prevalent techniques. Their popularity is attributed to relative simplicity, interpretability, modest computational demands, and robustness when data are sparse or when cold‑start users dominate. Decision‑tree models also facilitate feature‑importance analysis, which aids communication with business stakeholders. Deep‑learning approaches (CNNs, RNNs, Auto‑Encoders, etc.) appear in only about 22 % of the studies, largely confined to data‑rich domains such as video or music recommendation where large‑scale GPU resources are available. Traditional ML methods (SVM, K‑Nearest Neighbors, clustering) are used as complementary techniques in niche scenarios, for example when explicit feedback is abundant.

Beyond algorithm prevalence, the review extracts several determinants that influence algorithm selection: data sparsity, temporal dynamics of user behavior, availability of domain‑specific metadata, and real‑time latency constraints. For instance, news‑feed recommendation, which exhibits strong sequential patterns, benefits from RNN‑based models, whereas small‑scale retail settings with limited interaction logs favor Bayesian approaches.

A critical finding concerns the development lifecycle of RS. The majority of the examined literature concentrates on implementation and evaluation, with scant attention to the requirements‑elicitation and design phases. Consequently, there is a lack of systematic guidelines for mapping business requirements to appropriate ML techniques, predicting performance before implementation, or selecting architectural patterns (e.g., plugin‑based or micro‑service architectures) that support modularity and extensibility. Moreover, ethical and privacy considerations are rarely integrated into early‑stage design, indicating a need for formal requirement models that embed fairness, transparency, and data‑protection constraints.

The discussion underscores that current RS research is heavily accuracy‑centric, while broader quality attributes such as scalability, maintainability, and explainability receive limited treatment. The authors argue that bridging this gap requires a concerted effort from the software‑engineering community. They propose a research agenda comprising: (1) decision‑support frameworks for algorithm selection grounded in requirement attributes; (2) methods for early‑stage performance estimation (e.g., surrogate modeling); (3) design‑level patterns that enable plug‑and‑play ML components and facilitate continuous integration of new algorithms; and (4) incorporation of ethical and privacy requirements into RS design specifications.

In conclusion, the systematic review confirms that Bayesian and decision‑tree algorithms dominate due to their ease of use and interpretability, yet the phases of requirements gathering and system design remain under‑explored. By highlighting these deficiencies and outlining concrete research directions, the paper offers valuable guidance for both practitioners seeking to build robust, maintainable recommender systems and scholars aiming to advance the state of the art at the intersection of machine learning and software engineering.