Exploring the Power of Outliers for Cross-Domain Literature Mining

Reducing the Search Space in Literature-Based Discovery by Exploring Outlier Documents: a Case Study in Finding Links Between Gut Microbiome and Alzheimer’s Disease

“…Results obtained by Sluban et al [10] and by Petrič et al [9] confirm the hypothesis that most bridging terms appear in outlier documents and that by considering only outlier documents the search space for b-term identification can be largely reduced. In this way, we can substantially reduce the search space for finding b-term candidates.…”

“…An example tool that we have developed in previous research is CrossBee [8], a web-based tool for bridging term (b-term) discovery and exploration, which implements an ensemble based term ranking approach to finding new connections between two predefined domains, represented by two user defined sets of biomedical articles. The research conducted by Petrič et al [9] and Sluban et al [10] complements this research by showing that bridging terms are substantially more frequent in documents that are outlier documents of their own domain, compared to their frequency in normal, non-outlier documents. Analogously to statistics, where an outlier is defined as an observation that falls outside the overall pattern of a distribution [11], an outlier document in the field of LBD is a document that lies outside the main group of documents of its own domain and is, therefore, in two domain settings more similar to the documents of the other explored domain than to the documents of the domain of its origin.…”

“…One of the techniques that can be used to detect outlier documents is by using classification algorithms [10]. The technique works as follows.…”

“…It can be shown that the majority of bridging terms can be found in outlier documents [10], as show on the gold standard migraine-magnesium domain pair, for which a confirmed list of concept bridging terms (b-terms) was made available. The experimental results showed that the sets of detected outlier documents were relatively small -including less than 5% of the entire datasets -and that they contained a great majority of bridging terms, which was significantly higher than in same-sized random subsets.…”

“…This well-known example has become the gold standard in the literature mining field and has been used as a benchmark in several studies, including our own work, developed by Juršič et al [8], Sluban et al [10] and Petrič et al [9], which is the basis for the methodology presented in this paper.…”