所属 コンピュータサイエンス学部 コンピュータサイエンス学科 職種 専任講師
|著者・共著者||伏見 卓恭，佐藤 哲司，斉藤 和巳，風間 一洋|
|概要||In some types of real networks,
the tendency for some nodes in a network to connect to other nodes
that have similar features have been found.
However generally, not all nodes on the network possess the similar features
and there exist some change points of the feature distributions.
That is to say, when we extract a certain node group,
feature distribution over this node group differs in those over other nodes.
We also assume that the density of feature distribution gradually or rapidly decays
according to the distance from center of the distribution.
In this paper, we deal with feature vectors of nodes
which are obtained from activities or contents the nodes posted
and their resultant vectors with introducing distance based decay weights.
By clustering these resultant vectors, we attempt to extract
semantically and structurally cohesive node groups,
then annotate node groups with features which significantly occur in them.
Moreover, since the number of similar neighbor nodes differs in each node,
we should control the degree of decay with respect to each node.
Therefore, we explain an estimation method of exponential decay function paramter
which control distance based decay weights.
From experimental evaluations using three real networks,
we discuss the relation between estimated parameters and nature of nodes.
Furthermore, we quantitatively evaluate annotation results of extracted node groups
in terms of semantic and structural cohesiveness.