Graduate School of Information Science and Technology, Hokkaido University

Network Analysis

In the last decade, relational data, which can be naturally represented as a graph with nodes and edges, get great attentions in the data mining community. A number of real relational data, such as Social Network, Word Wide Web, are so large and complex that inspections and analysis by human eyes are practically impossible. As an example, let us glance the discussions in technical forum of an athletic club in Hokkaido University. You might feel that there seems no useful patterns, but by disentangling this graph carefully, we could discover, for example, groups of individuals sharing similar interest, influential persons, and many other interesting information.

Community Extraction

Many networks of interest such as computer, biological and social networks possess some clustered, dense sub-graphs, which is referred as communities. They often correspond to some practical units having special functions or roles in the networks so that a number of interesting characteristics are observed at community level. Therefore, the detection of communities are of great importance for our understanding of network systems. Many networks of interest are naturally changing over time, and communities on such networks cannot be represented by conventional community models proposed for static graphs. In this laboratory, we are trying to characterize time changing communities on social networks.

example.png

Anomaly Detection

Recent revolutions in information technologies allow us to collect huge amount of data in real time. Accordingly, a number of researchers have tried to discover some abnormal patters in real time such as fraud in transactions, defective product, failures in machines. The analysis on anomaly patterns in relational data have began recently and still many spaces for further sophistication. In our laboratory, we are trying to characterize the abnormal patterns on the basis of community models we developed [sada_2013a].

rect3428.png

Trend Tracking

In social network analysis, it is beneficial to know, on the communities inside the network, what they are interested in or when they became active. To obtain these beneficial information, we need to track communities and their interests over time. In this laboratory, we are tackling this task by using Nonnegative Tensor Factorization . “Tensor” is a multi-dimensional array. We can represent a time-series of networks by a 3-order tensor ( 3-dimensional array). Nonnegative Tensor Factorization is an algorithm that factorizes a tensor to matrices that provide a kind of interpretation on the structures and trends hidden in the time-series of the networks as shown in the following figure.

Trends.gif

Influence Maximization

Recently, as social networking services (SNS), such as Twitter and LINE, have spread widely, it has become easier for us to know how individuals are connected or how information spreads via relationships. Moreover, information can be propagated more fast and broadly than the conventional off-line word-of-mouth communication. Therefore, a research on information propagation over a social network has become more important. We often find that user reviews are more trustable than the ads of the product companies, which indicates big advertising potential of the word-of-mouth communication.

information_propagation.pngFinding opinion leaders on a social network is an important task for influence maximization over the network. For example, by selecting such opinion leaders as testers of a limited number of new product samples, the company can expect a big advertising effect through the diffusion of the reputation via word-of-mouth communication on SNS.

PAGETOP
Copyright © 情報認識学研究室 All Rights Reserved.
Powered by WordPress & BizVektor Theme by Vektor,Inc. technology.