Zhijing Liu

Motion pattern analysis in crowded scenes by using density based clustering

Video surveillance is always a hot topic in computer vision. With the public safe issue received more and more attention, analysis for crowd motion is becoming significant, and detecting motion patterns or activities in crowded scenes from videos is one of the major problem in crowd analysis. This paper proposes a new method for learning the motion patterns in crowded scenes. We add the direction information to the motion vectors, and cluster the data by a density based clustering. We extract the feature points using KLT corner extractor and track them to obtain basic motion information by optical flow techniques. All the motion information in different frames forms the motion flow field. Improved DBSCAN method is used to divide the motion flow filed into different patterns. The result of the system is given as a graph with groups of vectors. The experiment result in real-world videos is presented to demonstrate our approach.

A Novel Approach to Naive Bayes Web Page Automatic Classification

In this paper, a novel approach of Web page classification using Naive Bayes (NB) classifier based on independent component analysis (ICA) is proposed. In order to perform the classification, a Web page is firstly represented by a vector of features with different weights, and the weight calculated method is improved. As the number of the features is big, principal component analysis (PCA) which is to select the relevant features will perform in preprocessing section as input for improved ICA algorithm (MFICA). Finally, the output of MFICA is sent to NB classifier for classification to boost the classifierpsilas performance. The experimental evaluation demonstrates that the NB classifier based on ICA model provides acceptable classification accuracy.

Parallel Data Mining Optimal Algorithm of Virtual Cluster

Based on the problem of TB level mass data lacking of parallel patterns which is distributed on Earth and accessed by Internet, we focus on the research of parallel computing architecture structure--virtual cluster based on cloud computing. Meanwhile, the parallel data mining algorithm is studied, and the effectiveness of parallel data mining algorithm based on this platform is proved.

Group Recommendation Using Topic Identification in Social Networks

Many recommendation systems recommend item (book, music, news, or restaurant) to a group of users with the help of social network services like micro-blogging, which are called as group recommendation systems. However, as the social network services always contain large amounts of information on different topics, and people have different influence on different topic, so the group recommendation systems should take the topic identification in large-scale social networks into account to get an appropriate recommendation. In this paper, we proposed a new group recommendation method, which combines topic identification and social networks for group recommendation. In detail, we firstly identify different topical sub-groups by topics in social networks. Secondly, different user factors are used to calculate the user influence (including individual and social) on the topical sub-groups, which can depict the topical sub-group characteristics in different points of view. Experimental results demonstrate that the proposed method can improve the prediction accuracy of the group recommendation.

Group Recommendation Based on the PageRank

Social network greatly improve the social recommendation application, especially the study of group recommendation. The group recommendation, analyze the social factors of the group, such as social and trust relationship between users, as the factors for the prediction model established. In this paper, PageRank algorithm is introduced in the recommendation method to calculate the member’s importance in the group respectively, and to amend the aggregate function of individual preferences. The aggregate function consider the relationship between various users in the group, and optimize the aggregate function according to users different influence on the group, which can better reflect the social characteristics of group. In short, the study on group recommended model and algorithm can take the initiative to find the users needs. Extensive experiments demonstrate the effectiveness and efficiency of the methods, which improve the prediction accuracy of the group recommended algorithms.

Graph-based KNN text classification

Vector space model is used in most text categorization methods without considering the important information such as the order and co- occurrence of words within the text. In this paper we describe a novel approach of text classification using graph-based KNN. We reduce the number of features dimensions by a combined feature selection method. Then we present an improved graph-based text representation model and describe a novel graph-based KNN algorithm to predict the category of the texts in the testing set. The result shows that our approach can outperform traditional VSM-based KNN methods in terms of both accuracy and cost time.

A Classification Method Based on Non-linear SVM Decision Tree

The induction of classification of decision tree is an important algorithm for data mining now. The support vector machine technology and the decision tree have combined into one multi-class classifier so as to solve multi-class classification problems. In this paper, SVM is extended to non-linear SVM by using kernel functions and a new method of NSVM decision tree is proposed based on traditional SVM decision tree. Classification experiments prove the method is effective.

A new partitioning based algorithm for document clustering

Document clustering is one of the key problems in text mining and information retrieval area. It groups text documents in a way that maximizes the similarity within clusters and minimizes the similarity between different clusters. Most partitioning based algorithms are sensitive to the initial centroids, the clustering result greatly depends on the initial centroids. This paper first uses unsupervised feature selection method to reduce the dimension of document feature space and then proposes a novel partitioning based algorithm which select initial cluster centriods in the process of clustering by the size and density of cluster in the datasets. The experiments on several text datasets show that the proposed approach effectively improves the quality of clustering.

Audio classification in a weighted SVM

This paper presents a novel audio classification algorithm, which combines the rule-based with model-based method in an efficient way. First, the threshold-based method is performed over each audio clip for preclassification, with three typical features utilized and majority rule applied. Next, a weighted frame-based Support Vector Machine (SVM) is presented for further classification, using a new feature Mel-ICA as classification feature and preclassification results as weights. Finally, the experimental results have shown that the presented algorithm achieved effective audio classification, with accuracy rate increased greatly, and the new Mel-ICA was more suitable for classification than traditional mel-frequency cepstral coefficients (MFCCs).

Distributed Clustering Based on K-Means and CPGA

Distributed clustering is a new research field of data mining now. In this paper, one of distributed clustering named DCBKC (distributed clustering based on K-means and coarse-grained parallel genetic algorithm) based on K-means and coarse-grained parallel genetic algorithm is advanced. The algorithm can solve local clustering problem of distributed clustering effectively, reflect all of local data characters, enhance local datapsilas perspectivity and decrease network overload at a way by adopting proper migration strategy simultaneously. Both theory analysis and experimental results confirm that DCBKC is feasible.