Akira, Otsuki and Masayoshi, Kawamura: The Study of the Role Analysis Method
of Key Papers in the Academic Networks.
László Kovács: Reduction of Distance Computations in Selection of Pivot Ele-
ments for Balanced GHT Structure.
Both papers are very interesting, innovative and clearly written. These papers have
something in common. This is the use of specific elements and the search for them.
Such elements are the starting point for learning structures and replace simple but
complex searches. Both papers also make use of graphs for the representation. Relat-
ed problems have been studied in many publications. For this reason high quality
contributions are very welcome. The papers present different techniques what make
them particularly interesting.
The first paper is concerned with the search for papers in academic-journals. Eve-
ry researcher knows about this problem an
d how little help one gets from search engines. Basically, the (large) set of papers is divided into clusters for reducing the search. Clustering is a traditional semi-supervised technique in machine learning. This
is a complex task and centrally much context information is used. Then the next step
is to find key papers in a cluster. This relies on an earlier approach of “ Citation
Ranking“. These papers are the “special elements”. The approach relates the problem
to networks and therefore graph theoretic methods are central. First existing methods
for coupling nodes is investigated in great detail. Finally it is investigated how to
make use of them.
The second paper deals with search in generalized hyperplane trees using partition-
ing. The quality of a partitioning depends very much on the indexing. The paper pro-
poses an interval based matrix of distances. In such a situation specific elements,
namely the pivot elements are again of big interest and their selection is a classical
problem for metric spaces. In contrast to the other paper pivot elements have no inner
structure and the search is of purely combinatorial character. A major goal is balanc-
ing the tree structure. In this paper a combination of heuristics and local search opti-
mization is investigated. The selection of useful pivot elements is a key to useful in-
dexing. In this approach the pivot elements are discussed from the distance view only.
Distances seem to be clear and solid magnitudes. In practical life, however, this is
often violated and the author investigates uncertainty of distances as a first step.
In addition, distances may partially unkown what leads to interval-based distances. Be-
cause the main point in the paper is efficiency it is interesting to see how the interval-
based approach has a major influence for improving this. The reason is the
compactification nature what is studied in detail. The measure for improvement is
provided by a fitness function. Many of the techniques presented here are of interest
for quite other topics in machine learning.
Both papers seem to have a great value for readers in machine learning in general.
Both they also contain many concrete examples and evaluations based on them. This
shows their practical relevance and useful ness. In the overall, I can recommend to
read these papers.
