posted on 2023-06-21, 06:01authored byFrederique Bone, Diego Chavarro, Yuxian Liu
Purpose: The authors aim at testing the performance of a set of machine learning algorithms that could improve the process of data cleaning when building datasets. Design/methodology/approach: The paper is centered on cleaning datasets gathered from publishers and online resources by the use of specific keywords. In this case, we analyzed data from the Web of Science. The accuracy of various forms of automatic classification was tested here in comparison with manual coding in order to determine their usefulness for data collection and cleaning. We assessed the performance of seven supervised classification algorithms (Support Vector Machine (SVM), Scaled Linear Discriminant Analysis, Lasso and elastic-net regularized generalized linear models, Maximum Entropy, Regression Tree, Boosting, and Random Forest) and analyzed two properties: accuracy and recall. We assessed not only each algorithm individually, but also their combinations through a voting scheme. We also tested the performance of these algorithms with different sizes of training data. When assessing the performance of different combinations, we used an indicator of coverage to account for the agreement and disagreement on classification between algorithms. Findings: We found that the performance of the algorithms used vary with the size of the sample for training. However, for the classification exercise in this paper the best performing algorithms were SVM and Boosting. The combination of these two algorithms achieved a high agreement on coverage and was highly accurate. This combination performs well with a small training dataset (10%), which may reduce the manual work needed for classification tasks. Research limitations: The dataset gathered has significantly more records related to the topic of interest compared to unrelated topics. This may affect the performance of some algorithms, especially in their identification of unrelated papers. Practical implications: Although the classification achieved by this means is not completely accurate, the amount of manual coding needed can be greatly reduced by using classification algorithms. This can be of great help when the dataset is big. With the help of accuracy, recall, and coverage measures, it is possible to have an estimation of the error involved in this classification, which could open the possibility of incorporating the use of these algorithms in software specifically designed for data cleaning and classification. Originality/value: We analyzed the performance of seven algorithms and whether combinations of these algorithms improve accuracy in data collection. Use of these algorithms could reduce time needed for manual data cleaning.