Tan Jianlong

Theme word subspace method for text document categorization

In this paper, a text document categorization method called Theme Word Subspace (TWS) learning is presented, which utilizes theme words jointly express class-semantic information for document classification. In a class corpus, the theme words with high probability distribution in topic structure are extracted firstly, and then these words as important theme element span class subspaces to jointly represent semantic and distribution of the class. For document categorization processing, a text document is belonged to the nearest subspace whose theme words have the best representation for test document. In our TWS, L1, L2 norm are separately used for measuring the distances of a test document to subspaces. Experiments on a large Chinese text corpus, the proposed TWS learning methods exhibit comparable performances for text document category.

Research of New Words Identification in Social Network for Monitoring Public Opinion

With the rapid development of Internet, a large number of new words have emerged and widely been used in social network. Traditional segmentation algorithm can’t identify these new words efficiently, which will greatly affect the accuracy in extracting out these hot words and keywords. Moreover, it will affect the performance of the network public opinion monitoring system. In this paper, we use tweets collected from Twitter as the experimental data-set. By calculating frequency statistics of k-gram strings, we can find out new words as candidates, and then identify new words by their practical application frequency using Twitter’s search function. The experiment shows: this segmentation algorithm can effectively identify the new keywords and is more suitable for public opinion monitoring system.

Research of Security Relationship Based on Social Networks

This paper introduces the security problems of social networks and the research topics of core nodes, relationship, structures of social network, and so on. This paper discusses the importance of “security relationship” in social networks, analyzes the relevance between security risks and events, and adopts the qualitative and quantitative mechanism based on grade partition, numerical measure, polymorphic data fusion, and the logical relevance. Based on Bayesian network, this paper proposes a kind of assessment model for the situation of security relationship in social networks so as to provide theoretical bases for the perception and prediction of security situation in social networks.

Finding δ-Closure Property of BitTorrent System and Applications Based on This Property

In the Peer-to-peer (P2P) field, a basic problem is to build a model which can exactly represent the behaviors of peers. According to the protocols of BitTorrent (BT) system, participating peers are selected randomly when downloading some content. However, by using the connection model gathered during a one-month period, we found that the BT peers are subject to the δ-closure property, which is that the content in one δ-closure tends to spread inside this δ-closure. In this paper, we defined δ-closureaccording to the complex network theory. Then we tried to design algorithm to attain δ-closure. By gathered dataset, we verified the existing of δ-closure in BT system. Finally, we proposed possible applications based on this property.

Traffic-Aware Frequent Elements Matching Algorithms for Deep Packet Inspection

Network devices are increasingly using packet content for processing incoming or outgoing packets. Many pattern matching algorithms have been proposed to improve packet matching throughput. Most of them are, however, independent of traffic pattern and may end up with longer match time against actual traffic. We present novel algorithms that utilize traffic characteristics coupled with frequent elements pattern matching to obtain high throughput. The algorithms modified and expanded the current matching procedure and data structure of classical pattern matching algorithms. The presented pattern matching algorithms, using the traffic-aware frequent elements and the dynamic pattern matching algorithms to adaptive to the traffic, have performance advantage with the true dynamic network traffic.