Machine learning based evaluation of scientific retractions

Researchers produce a tremendous number of publications based on their findings every single day. The frequency of retracted papers published in scientific journals has sharply increased during the last 20 years, outpacing the gradual growth in the overall number of publications. Retraction is a mechanism for correcting the literature by informing the readers about papers that include information or data that is seriously faulty or inaccurate that makes the researcher impossible to rely on the findings and conclusions. Questionable research may propagate disinformation and put people in danger, particularly in the medical literature. As of May 2020, the Retraction Watch Database shows 23,000 retracted papers and 800,000 articles that directly cite them. In this thesis, we address the problem of determining which publications are close to retraction based on their association with the retracted articles and when we need to reconsider papers that cite them. We proposed a method in which we can predict the paper that is close to retraction using different machine learning algorithms like Logistic Regression, Naïve Bayes, SVM and Random Forest. Moreover, we introduce an annotation scheme for citation functions with the help of a subject matter expert. According to this scheme, we have found that there are 221 papers out of 664 that need reconsideration. The trained model achieved approximately 82% accuracy in identifying possible retractions. In this way, we can save our resources to check some percentage of potentially harmful literature instead of reviewing the complete database. In the context of healthcare, several domains and categories may be taken into consideration for future work to form new data. More insights and trends may be discovered from the collected data. The development of deep learning models for feature engineering would also be useful.