Marcela Mosquera

Prof. Marcela Mosquera

orcid

Research Focus & Metrics

Academic Career Length

3Years (2024-Present)

(Calculated once Scilit publication data is synced: current year − year of first publication + 1)

Representative Publications

View All Publications

24 December 2025 in IEEE Access
Verified
Journal Article
Open access
Marcela MosqueraMarcela Mosquera
Marcela Mosquera
RBRodolfo Bojorque
Rodolfo Bojorque
PFPamela Flores
Pamela Flores
Hyperparameter optimization plays a pivotal role in the reliability and generalization of machine-learning models for software quality prediction. This paper presents a comparative evaluation of three search strategies: Grid Search, Random Search, and a Hybrid approach applied to the detection of four types of code smells (BLOB, Functional Decomposition, Spaghetti Code, and Swiss Army Knife). Experiments were performed on three open-source Java systems (Azureus, ArgoUML, and Xerces) using Support Vector Machines (SVMs). The results show that all three strategies produce consistent patterns in accuracy and F1-score across the datasets, confirming the robustness of the evaluation framework. Simultaneously, Random Search tends to achieve higher recall for the Functional Decomposition smell, reflecting its ability to explore sparse and irregular hyperparameter spaces. The Hybrid approach combines the exploratory strength of Random Search with the local refinement of Grid Search, offering greater stability with a moderate computational cost. A correlation analysis of software metrics further reveals that while the Blob smell is consistently associated with size and coupling indicators, the other smells—Functional Decomposition, Spaghetti Code, and Swiss Army Knife are captured more indirectly through general structural proxies such as size, cohesion, and coupling. Overall, these findings suggest that hyperparameter tuning not only enhances predictive performance but also provides valuable insights into the structural factors underlying code smells.

8 April 2026 in Applied Sciences
Verified
Journal Article
Open access
PDF file
Marcela MosqueraMarcela Mosquera
Marcela Mosquera
RBRodolfo Bojorque
Rodolfo Bojorque
Code smells are indicators of potential design problems in software systems and are commonly used to guide refactoring activities. Recent advances in representation learning have enabled the use of embedding-based models for analyzing source code, offering an alternative to traditional approaches based on manually engineered metrics. However, the effectiveness of different embedding representations for multiclass code smell detection remains insufficiently explored. This study presents an empirical comparison of embedding models for the automatic detection of three widely studied code smells: Long Method, God Class, and Feature Envy. Using the Crowdsmelling dataset as an empirical basis, source code fragments were extracted from the original projects and transformed into vector representations using two embedding approaches: a general-purpose embedding model and the code-specialized CodeBERT model. The resulting representations were evaluated using several machine learning classifiers under a stratified group-based validation protocol. The results show that CodeBERT consistently outperforms the general-purpose embeddings across multiple evaluation metrics, including balanced accuracy, macro F1-score, and Matthews correlation coefficient. Dimensionality reduction analyses using PCA and t-SNE further indicate that CodeBERT organizes code smell instances in a more structured latent representation space, which facilitates the separation of smell categories. In particular, CodeBERT achieved a macro F1-score of 0.8619, outperforming general-purpose embeddings (0.7622) and substantially surpassing a classical TF-IDF baseline (0.4555). These findings highlight the value of this study as a controlled multiclass evaluation of embedding representations and demonstrate the practical value of domain-specific representations for improving automated code smell detection and class separability in real-world software engineering scenarios.


2 September 2026
Verified
Book Chapter
Closed access
PDF file
PFPamela Flores
Pamela Flores
Marcela MosqueraMarcela Mosquera
Marcela Mosquera
JHJorge Luis Segovia Herrera
Jorge Luis Segovia Herrera
CACarlos Anchundia
Carlos Anchundia

20 March 2026 in Preprints.org
Verified
Preprint
Open access
PDF file
Marcela MosqueraMarcela Mosquera
Marcela Mosquera
RBRodolfo Bojorque
Rodolfo Bojorque
Code smells are indicators of potential design problems in software systems and are commonly used to guide refactoring activities. Recent advances in representation learning have enabled the use of embedding-based models for analyzing source code, offering an alternative to traditional approaches based on manually engineered metrics. However, the effectiveness of different embedding representations for multiclass code smell detection remains insufficiently explored. This study presents an empirical comparison of embedding models for the automatic detection of three widely studied code smells: Long Method, God Class, and Feature Envy. Using the Crowdsmelling dataset as an empirical basis, source code fragments were extracted from the original projects and transformed into vector representations using two embedding approaches: a general-purpose embedding model and the code-specialized CodeBERT model. The resulting representations were evaluated using several machine learning classifiers under a stratified group-based validation protocol. The results show that CodeBERT consistently outperforms the general-purpose embeddings across multiple evaluation metrics, including balanced accuracy, macro F1-score, and Matthews correlation coefficient. Dimensionality reduction analyses using PCA and t-SNE further indicate that CodeBERT organizes code smell instances in a more structured latent representation space, which facilitates the separation of smell categories. These findings provide empirical evidence that domain-specific pretraining plays an important role in representation learning for software engineering tasks and that embedding choice significantly influences the separability of code smell categories in multiclass detection settings.