Takuo Nakashima

Information filtering for the newspaper

A large quantity of information is spread out to our life, especially to our cyberspace. Not only are major newspaper companies providing a Web service, so are regional companies with their characteristic Web pages. We built an information filtering system using multi-agent for the article which a user wanted to read in the regional newspaper. Our system has two main agents. One is the ranking agent for the newspaper, the other is the feedback agent. The ranking agent ranks articles based on user keywords. The feedback agent feeds back to the user profile data based on the access frequency. We propose the ranking and feedback algorithms, and estimate these algorithms by experiments. Experiments show there is relation between filtered and read article by a chi-square test of 0.05 significance level. Also they show our system is efficient for the passage of time.

Neurological abnormalities in cognitively impaired but not demented elderly.

Objective– We determined the prevalence and clinical characteristics of elderly people with cognitively impaired but not demented (CIND) in the community. Material and methods– We examined 945 elderly residents (aged 65 years or more) in a cross‐sectional study of a standardized clinical examination in 27 randomly selected communities in the town of Mifune, a standard farming town in Japan. Results– The prevalence of CIND and dementia were 10.8% and 4.8%, respectively, with both rates increasing with age. The frequency of most neurological signs in CIND elderly was greater than that found in subjects with no cognitive impairment (NCI) and less than that in subjects with dementia. The mean ADL disability score in CIND subjects also lay between the mean values for those with NCI and those with dementia. There were no differences in lifestyles between the CIND group and the other groups. Conclusion– The results suggest that a majority of CIND elderly suffer from brain dysfunction due to chronic neurological disorders or aging itself.

An experimental evaluation of the total cost of NTP topology

To improve the quality of the Internet, the behavior of Internet is evaluated by active or passive measurements. For a one directional measurement method, it is assumed that the network time protocol (NTP) server can maintain the exact absolute time. But the relative time of the NTP server depends on the NTP server-client topology and the interval time of peering hosts. In this research we define the offset time of the peering hosts, propose the evaluation model of the NTP based network and define the total cost of the NTP topology. We measure the offset time between two NTP servers by experiments in various network structures and extract the parameters that cause the offset time drift. From the results of a few experiments, we discovered that (1) a large number of hosts under 100 does not affect the accuracy of NTP server, (2) the processing time of the NTP server depends on synchronous processing, (3) the difference between the sending and receiving packet delay affects the real offset value, (4) if there is no NTP server, there is a 0.028 (millisecond/second) drift, and (5) there are reasons for the offset time drift in peer host synchronization. From these results, we can get some of the characteristics of the NTP evaluation model.

Entropy Based Analysis of Anomaly Access of IP Packets

To defend DoS (denial of service) attacks, the access filtering mechanism is adopted on the end servers or the IDS (intrusion detection system). The difficulty to define the filtering rules lies where normal and anomaly packets have to be distinguished in incoming packets. The purpose of our research is to explore the early detective method for anomaly accesses based on statistic analysis. In this paper, we firstly define the entropy-based analysis, then analyze the amount of incoming packets to our collage. As the results, we were able to extract the following features for the entropy analysis. Firstly, entropy-based analysis detect distributed small amount of 80/TCP anomaly accesses compared to the frequency-based analysis. Secondly, one hour window size is most sensitive to find the 80/TCP anomaly accesses. Finally, after applying the filter of noisy accesses of ICMP anomaly packets for total amount of data sets, entropy-based analysis detects small amount of ICMP anomaly accesses.

Classification of characteristic words of electronic newspaper based on the directed relation

Newspaper articles are gradually opened to Web system. But the categories of articles are remained in conventional category types. So users of these system have to consider the appropriate searching keywords and categories when they search some articles. In this paper, we propose the new classification method for the characteristic words based on directed relation. We think this classification can change categories of articles and accessing keywords of users. First, we introduce the word vector and directed relation degree. Second, we define the classification method, reference degree and abstraction method of cluster. Finally, we experimented using one month data and can get the following results. (1)Both directional tightly connected words are suitable base words of clusters of articles. (2)Level-0 abstraction method can draw the rough border of clusters. (3)One directional tightly connected words represent the main general words or proper words.

Daily clustering for the electronic newspaper based on the analysis of trends

To classify newspaper articles automatically, the tf*idf method has been used to weight the words in an article. These methods are suitable for fixed databases, but cannot pick up the topic words of articles because the IDF methods give a low value for frequently occurring words. We propose the daily clustering method for electronic daily newspapers. Our method is based on the characteristics of articles and the change of contents. First, we define the weight function of words based on the position in the article and the change rate of content as time passes. Then we calculate the relation between articles, clustering value and the relation between clusters of different days. As a result of experiments, the accuracy of recall and precision rate improved by several percent compared with old methods.

Active measurements by the two-dimensional path-based IP performance method

Packet loss distribution on Internet nodes depends on the node performance and the bandwidth of the incoming and outgoing link. Each node knows the performance of the IP forwarding processing and bandwidth of the link, but the application can only detect congestion by packet loss. We propose a two-dimensional (path and time dimensions) IP performance evaluation method. We propose new metrics, such as node dependency, relation degree, time dependency and time lag dependency. By experiments, we can get the results showing new node characteristics as follows: (1) the node dependency of path direction is greater than anti-path direction; (2) these pre-dependency nodes are classified leakage type and suppression type; (3) node relation degree is a higher value after the first increasing point of minimum delay; (4) leakage type nodes have big burst degree; (5) the time lag relation degrees are approximately quarter of the node relation degrees; (6) the periodical delays appear in FPTT (forwarding path trip time).