Search

Considering the complexity of trust among users, we presented a new trust factor model. In this model, trust relationships among users are represented as an adjacent matrix, and trust matrix can be decomposed into two low rank dimensionality matrixes, which are trust factor matrix of users and trusted factor matrix by users, through matrix factorization in this model. User-to-user trust values are computed as inner products of corresponding vectors., Trust users can be identified according sorts of trust values. The model can give consideration to the global structure and finer-grained characteristics of trust. The experimental results show that this model has higher accuracy of trust identification and efficiency in prediction stage. The Trust factors model not only reflect users' characteristics of trust and trusted, but also can be used to explain trust propagation in trust chain.

The way to measure similarity among users is a key factor affecting accuracy of recommendation algorithm. A similarity measuring algorithm based on degree of correlation among users and weighting of recommended items is proposed to improve less marking aspects between users and lower prediction accuracy of traditional similarity measuring algorithms. The method can solve the non-convex problem generated by non-exclusive parameters in the objective function by introducing an adjusting parameter of similarity and alternatively optimizing the relative parameters using least squares. Experimental results show that the proposed method is effective in improving the accuracy of whole system and the prediction accuracy of recommendation items.

Automatic summarization can help us accurately and efficiently obtain the information needed from the magnanimity information and has attracted more attention. In this paper, a new method for Chinese text summarization using the algorithm of Affinity Propagation Cluster (APC) is presented. It is not necessary to set the number of clusters and the initial representative exemplars in APC, so it can avoid the problems of local-optimal and instable clustering results caused by randomly selecting initial representative exemplars. And the algorithm has high computing efficiency. The results of the experiments show us that Chinese text automatic summarization based on APC has higher accuracy than that of other algorithms. APC is a suitable method for automatic text summarization.

Data dimension reduction plays an important role in the field of text representation. An effective dimension reduction method can not only reduce computation complexity, but help to improve the accuracy of text classification. This paper presents a new method of dimension reduction which is based on words semantic similarities. Being different with traditional methods which usually use the statistical information of words, natural language processing knowledge is used in our method which considers semantic information and POS information of feature terms. The experimental results show that our method is effective in dimensionality reduction of text representation and achieves a higher accuracy of text classification. The semantic similarity based method is a suitable method for text representation.

As a feature co-occurrence approach, the traditional event tracking technique encounters a dilemma when setting the tracking similarity threshold that can diminish its tracking effectiveness. In this paper we present a new event tracking method-variable query, which can automatically adjust query terms and its corresponding weights according to the event evolution stages and tracking similarity, and solve the dilemma well. Experiment shows that the method of variable query can improve the precision and decrease the false rate of tracking.

The main tasks of Opinion Analysis of Online News comments are comment objects identification and determining polarities of comments. As most online comments are short sentences, the comment polarity should be analyzed on sentences level. Inspired by n-grams rational kernels theory, a new method, Distance-Weighted Count (DWC) of polarized word, is proposed in this paper. Unlike others weighted method, DWC is an unsupervised method and it can meet the requirement of Online News comments processing. Experiments results show that our method is capable of analyzing polarity of online news comments with higher precision and can satisfy the requirements of online news comments analysis.