Similarity determination for clustering textual documents

  • Authors:

  • Vladimir Barakhnin;Vera Nekhaeva;Anatolii Fedotov

  • Affiliations:

  • Institute of Computational Technologies, SB RAS, Novosibirsk State University, Novosibirsk, Russia;Institute of Computational Technologies, SB RAS, Novosibirsk State University, Novosibirsk, Russia;Institute of Computational Technologies, SB RAS, Novosibirsk State University, Novosibirsk, Russia

  • Venue:
  • KONT'07/KPP'07 Proceedings of the First international conference on Knowledge processing and data analysis
  • Year:
  • 2007

Quantified Score

Hi-index 0.00

Visualization

Abstract

The problem of computerized selection of textual documents on scientific subjects is treated with new improved methods; that could be of interest for an individual researcher or a research team. Attributes of a bibliographical description (authors, keywords, abstract) are proposed to be used as scales for the measure determination. The values of weight coefficients in the formula for calculating the similarity measure are determined by the assumed a posteriorireliability of the respective scale data. Three classical document clusterization methods have been analysed in order to find the ones potentially feasible for the solution of the formulated problem: clusterization by finding cliques in the full matrix of documents similarity, clusterization by Rocchio method and the method based on the so-called greedy algorithm as well as the new method suggested by N. Zagoruiko based on employing the function of rival similarity (the so-called FRiS-function). Testing showed that the FRiS algorithm proved to be the most efficient one for this problem although the greedy algorithm also yields acceptable results.