Classification and clustering for neuroinformatics: Assessing the efficacy on reverse-mapped NeuroNLP data using standard ML techniques

Abstract

NeuroinformaticsNatural LanguageProcessing (NeuroNLP) relies on clustering and classification for information categorization of biologically relevant extraction targets and for interconnections to knowledge-related patterns in event and text mined datasets. The accuracy of machine learning algorithms depended on quality of text-mined data while efficacy relied on the context of the choice of techniques. Although developments of automated keyword extraction methods have made differences in the quality of data selection, the efficacy of the Natural Language Processing (NLP) methods using verified keywords remain a challenge. In this paper, we studied the role of text classification and document clustering algorithms on datasets, where features were obtained by mapping to manually verified MESH terms published by National Library of Medicine (NLM). In this study, NLP data classification involved comparing Stechniques and unsupervised learning was performed with 6 clustering algorithms. Most classification techniques except meta-based algorithms namely stacking and vote, allowed 90% or higher training accuracy. Test accuracy was high (=>95%) probably due to limited test dataset. Logistic Model Trees had 30-fold higher runtime compared to other classification algorithms including Naive Bayes, AdaBoost, Hoeffding Tree. Grouped error rate in clustering was 0-4%. Runtime-wise, clustering was faster than classification algorithms on MESH-mapped NLP data suggesting clustering methods as adequate towards Medline-related datasets and text-mining big data analytic systems.

Mining Inter-Relationships in Online Scientific Articles and its Visualization: Natural Language Processing for Systems Biology Modeling

“…Since computational cost was increased due to clustering and mapping, the runtime efficiency was increased by saving query results in a local database and retrieving non-listed documents based on the search results. Mapping cluster labels to MeSH terms helped to avoid irrelevant cluster labels and improved machine learning accuracy [36]. By using different machine learning algorithms, it was found that reverse mapped data showed higher accuracy compared to data extracted directly from retrieved documents.…”

“…Machine learning has been used in addressing BioNLP aspects including text classification [32], tagging structured models [33], parsing and extraction [34], and unsupervised learning with structure induction and document clustering [35]. In a previous study [36], performance of several machine learning algorithms on BioNLP datasets was evaluated. Eight classifiers-based learning models on 2000-point dataset with MeSH terms as features showed ~78.2% training accuracy, while the root mean square error was <0.26.…”

“…The choice of clustering algorithms was based on estimations of accuracy of the learning models [36]. For LINGO and STC, normal clustering (single-run) and iterative automated clustering (referred as auto-clustering) modalities were implemented.…”