Novel Search Strategies for Conducting Systematic Literature Reviews: Comparing a Novel Citation-Based Combined Bfsearch Approach with Text-Based User-Defined Queries and Relevance Prediction Through Active Learning
15 Pages Posted: 26 Apr 2022 Publication Status: Under Review
A novel combined Backward and Forward search (BFsearch) strategy was developed in this study. Its performance was compared to common user-defined queries and novel relevance prediction through active learning based on textual information using balanced random forest classification. The three search strategies were applied to two typical case studies related to the food science domain: a systematic scoping review and a critical review focussing on trends. Data was collected for both case studies using several online databases. Tools were developed to automatically combine duplicates through unique identifiers and to speed up the categorisation process through automatic text highlighting. The performance of the search strategies was evaluated using precision-recall curves. BFsearch was superior to queries for case study 1, acting complementary in both case studies. More research is needed to further develop the BFsearch strategy. To reach a high level of sensitivity (high recall), active learning proved most efficient with a high precision. Preference was given to relevance prediction based on the abstracts and titles of the papers, as compared to title only. Currently, active learning strategies focus mostly on textual data, so it is advised to study the integration of the BFsearch approach and other citation-based approaches within active learning.
Keywords: Balanced random forest classification, Categorization interface, Citation score, Machine Learning, Precision-recall curve, Text highlighting.
Suggested Citation: Suggested Citation