Cheng Xueqi

Aspect-level opinion mining of online customer reviews

This paper focuses on how to improve aspect-level opinion mining for online customer reviews. We first propose a novel generative topic model, the Joint Aspect/Sentiment (JAS) model, to jointly extract aspects and aspect-dependent sentiment lexicons from online customer reviews. An aspect-dependent sentiment lexicon refers to the aspect-specific opinion words along with their aspect-aware sentiment polarities with respect to a specific aspect. We then apply the extracted aspect-dependent sentiment lexicons to a series of aspect-level opinion mining tasks, including implicit aspect identification, aspect-based extractive opinion summarization, and aspect-level sentiment classification. Experimental results demonstrate the effectiveness of the JAS model in learning aspect- dependent sentiment lexicons and the practical values of the extracted lexicons when applied to these practical tasks.

Structural-interaction link prediction in microblogs

Link prediction in Microblogs by using unsupervised methods aims to find an appropriate similarity measure between users in the network. However, the measures used by existing work lack a simple way to incorporate the structure of the network and the interactions between users. In this work, we define the retweet similarity to measure the interactions between users in Twitter, and propose a structural-interaction based matrix factorization model for following-link prediction. Experiments on the real world Twitter data show our model outperforms state-of-the-art methods.Link prediction in Microblogs by using unsupervised methods aims to find an appropriate similarity measure between users in the network. However, the measures used by existing work lack a simple way to incorporate the structure of the network and the interactions between users. In this work, we define the retweet similarity to measure the interactions between users in Twitter, and propose a structural-interaction based matrix factorization model for following-link prediction. Experiments on the real world Twitter data show our model outperforms state-of-the-art methods.

An improvement of the fast uncovering community algorithm

Community detection methods have been used in computer, sociology, physics, biology, and brain information science areas. Many methods are based on the optimization of modularity. The algorithm proposed by Blondel et al. (Blondel V D, Guillaume J L, Lambiotte R and Lefebvre E 2008 J. Stat. Mech. 10 10008) is one of the most widely used methods because of its good performance, especially in the big data era. In this paper we make some improvements to this algorithm in correctness and performance. By tests we see that different node orders bring different performances and different community structures. We find some node swings in different communities that influence the performance. So we design some strategies on the sweeping order of node to reduce the computing cost made by repetition swing. We introduce a new concept of overlapping degree (OV) that shows the strength of connection between nodes. Three improvement strategies are proposed that are based on constant OV, adaptive OV, and adaptive weighted OV, respectively. Experiments on synthetic datasets and real datasets are made, showing that our improved strategies can improve the performance and correctness.

SSInSE: an intelligent search engine based on WWW structure analysis

In this paper, characteristics and shortcomings of three generations of web search engines are analyzed, along with users’ search requirements. Based on these observations, we propose a new strategy to build large scale comprehensive search engines: to replace pages with coarser granularity-- semantically common topic page groups--as the basic unit of a search engine. To provide intelligent search service, we will make heavy use of links to reflect human thoughts.