Article Preview
TopIntroduction
The amount of biomedical texts available in MEDLINE database is tremendous, and continues to grow rapidly. In fact, this raises the necessity of efficient techniques for large datasets management, especially, to access the relevant information in a short time, this itself is a challenge. Hence, the application of data mining techniques, more precisely, the clustering is a fair candidate solution for this difficult requirement (Simpson and Demner-Fushman, 2012). The clustering is an unsupervised technique of data mining. It plays an important role in document summarization and organization, automatic topic detection, and fast information retrieval and filtering. Clustering’s main aim is: dividing automatically the documents collection into subgroups such that, the documents of the same group are similar, and represent the same subject, while the documents in different subgroups are dissimilar and concern different topics and meanings.
In nowadays document clustering is a very active research field, and many approaches have been established to deal with it (Dipak and Mukesh, 2011; Oikonomakou and Vazirgiannis, 2009; Zamir and Etzioni, 1998; Vidyadhari et al., 2019). They are categorized into two major classes: the hierarchical and the partitioning based clustering. The difference between these two categories of clustering methods resides in the properties of the delivered clusters. In the partitioning based clustering, the data are directly divided into a predefined number of disjoint groups. However, in the hierarchical clustering, a dendrogram is generated in levels’ sequences, in each one, a partitioning clustering is realized with a fixed number of clusters. It varies from singleton clusters to one cluster containing all the data. Its unsupervised nature makes clustering as one of the most difficult problems of data mining. Furthermore, it is considered as an NP-hard problem (Xu and Wunsch, 2005; Jain et al., 1999; Dubes, 1993). One should notice that the time complexity of hierarchical clustering is quadratic, whereas it is almost linear in the partitioning approaches. Therefore, the partitioning approaches are more suitable for clustering large-scale datasets.
One of the most famous partitioning algorithms is K-means and its variants (Jain, 2010). It has linear time complexity, yet, it is the most efficient method for large datasets in terms of execution time. However, the K-means suffers from the problem of random initialization which gets sometimes trapped to a local solution (Dalhatu and Tie Hiang Sim, 2016; Ishak Boushaki et al., 2014b). Recently, to overcome this drawback, metaheuristics based algorithms have demonstrated their efficiency (Heraguemi et al., 2015) for NP-hard problems like the clustering (Ishak Boushaki et al., 2014a; Ishak Boushaki et al., 2018a; Nanda and Panda, 2014; Hruschka et al., 2009; Aboubi et al., 2016; Danish et al., 2019). These algorithms can reach optimal or near-optimal solutions to such problems in a reasonable time. For this purpose, clustering is seeing as a category of the optimization problem, which maximizes or minimizes an objective function called fitness function. Many research groups proved that the metaheuristic algorithms can find a global solution for the clustering problem instead of the local one and they provide successful results in linear time complexity (Nanda and Panda, 2014; Hruschka et al., 2009).