Chuang Zhang

An optimized item-based collaborative filtering recommendation algorithm

Collaborative filtering is a very important technology in E-commerce. Unfortunately, with the increase of users and commodities, the user rating data is extremely sparse, which leads to the low efficient collaborative filtering recommendation system. To address these issues, an optimized collaborative filtering recommendation algorithm based on item is proposed. While calculating the similarity of two items, we obtain the ratio of users who rated both items to those who rated each of them. The ratio is taken into account in this method. The experimental results show that the proposed algorithm can improve the quality of collaborative filtering.

A distributed vertical crawler using crawling-period based strategy

Due to the explosive growth of the web pages, centralized crawlers are no longer sufficient to run on the web efficiently. There are many distributed crawlers in wide use; however, none of them is suitable for template-customized vertical crawling. In this paper, we present a distributed template-customized vertical crawler which is specially used for crawling Internet forums. The Client-Server architecture of the system and the function of every module are described in detail which can be extended to other fields easily. A crawling-period based distribution strategy is also proposed, with which the crawler manager can coordinate the quantity of crawling tasks and the resources of each crawler very well, and the crawler can process websites with different updating frequency flexibly. We also define a communication protocol between crawlers and crawler manager and describe how to solve the duplicated crawling problem in the distributed system. The performance of centralized vertical crawler and distributed vertical crawler are compared in the experiment. Experimental results demonstrate that the parallel operation of all the crawlers in the distributed system can greatly enhance the crawling efficiency.

Real-time quality control for crowdsourcing relevance evaluation

In recent years, crowdsourcing has become an effective method in many fields, such as relevance evaluation of search engine results. However, the problems of bad workers and quality control always stand among the major challenges within the realm. Based on our experiments carried out in TREC 2011 Crowdsourcing track, this paper demonstrates a realtime strategy in recruiting workers and monitoring the quality of their relevance and rank judgments. The effectiveness of our strategy has been verified by empirical results.

Social Media Communication Model Research Bases on Sina-weibo

The popularity of microblog brings new characters to information diffusion in social networks. Facing new challenges of understanding information propagation in microblog, the framework of information producing and receiving was proposed. A general model named competing-window is also presented based on human behavior. The detailed composition of the model and its basal mathematical description are also given. In addition, a parameter called information lost as a supplement to measure dynamics of information diffusion. Meanwhile, the further application of our model to information processing and propagating was pointed out. All those work is based on the studies on human dynamics. Finally, to verify applicability, the model was applied to empirical data crawled from Sina-weibo. The interesting patterns extracted from empirical data indicate that microblog in deed is fundamentally characterized by human dynamics.

Blog extraction with template-independent wrapper

Rich information is contributed to blogs by millions of users all around the world with the development of blogsphere. However, few work has been done on the study of blog extraction so far. Unlike the traditional template-dependent wrapper, not only blog articles but also blogroll is extracted with template-independent wrapper in this paper. In our method, blog extraction is formalized as a machine learning problem and a template-independent wrapper is learned by using labeled blog pages from a single site. Testing pages are obtained from 10 popular Chinese blog sites. And experimental results on 300 real blog pages indicate that the proposed method can correctly extract data from blogs with the accuracy of 90% or even above.

Structured diffusion: Amplify your voice in online social networks

Marketing in a social network via structured groups outperforms spontaneous social conversations. It regulates the social interactions and thus the information pathways, which boosts overall impact. We start with a case study of a strongly-organized advertising network in Weibo, a popular microblogging service, which demonstrates that commercials promoted by this group are averagely twice effective as normal tweets. Then we present a modeling framework to characterize this structured diffusion paradigm. An analytical model using heterogeneous mean-field approximation was thus worked out to capture transmission dynamics. The transmission probability incorporates the inflation and propagation of node influences that explains why and how the structures within a campaign affect the spreading speed, scale, and impact. As empirical evidences, our model numerically simulates four typical structure settings on a large Twitter network. It confirms quantitatively that, hierarchical grouping excels the loosely or randomly coupled ones and classical marketing heuristics.

Multi-level topic detection algorithm for Netnews Specials

This paper investigates the topic detection method in Netnews Specials Detection (NSD). We found that when the traditional clustering algorithms are used in NSD, the same topic is usually split into several pieces and the result is not satisfying. So a new algorithm is proposed which uses a multi-level model, better suited for NSD. Firstly, such algorithm elevates the accuracy of single-layer clustering by introducing hot search words, a selective dictionary, and an advanced weight formula. Secondly, the multiple-level model not only avoids the problem of topic over-split but also establishes a structure for Netnews Specials, which lays the foundation for quick viewing, positioning and retrieval. Experimental results show that the algorithm in the real test corpus have high accuracy, doing a better job than the traditional clustering method.

A session-oriented retrieval model based on Markov random field

In this paper, we study how to use the search session information to improve the retrieval accuracy. We propose a session-oriented retrieval model based on Markov random field. This model introduces the correlations between query terms as a retrieval factor into the retrieval process. It also presents a dynamic update algorithm based on the analysis of users search behavior. Our model implements a complete session-oriented information retrieval framework finally. We use ClueWeb09 category B dataset and TREC 2010 (2011) Session dataset to quantitatively evaluate the model. Experimental results show that our model can improve retrieval performance substantially using the search session information.

High efficient memory fetch architecture for motion compensation of video decoder

To improve the performance of the real-time video decoding under the power-constrained environment, an efficient memory fetch architecture for accessing the motion compensation memory is proposed. According to the feature of accessing the motion compensation memory, the architecture adopts the cache mechanism to buffer the reference data and chooses DMA to access external memories. The architecture can be applied for H.264/AVC, MPEG-4, AVS and other video decoding systems. Compared to the conventional memory fetch module, experimental results show that the proposed architecture reduces 12.8%∼16.7% video decoding cycles in H.264 decoding. The timing and the area of this design are both satisfied after synthesized.

An interactive framework for online image annotation

Image annotation plays a very important role in managing and retrieving abundant images over the internet nowadays. In this paper, we try to propose a Human-Computer interactive framework for online image annotation which fully considers the human judgments and all online-users feedback. First, an interactive image annotation interface is demonstrated. It is utilized to interact with online-users and collect the original manual annotations. Second, an object-based image annotation framework is proposed to provide the automatic results for users. Then, the users feedback can be recorded and integrated to help our system to achieve performance improvement. Third, a comprehensive method of annotation refinement is developed to remove the noisy keywords. Finally, experimental results demonstrate the effectiveness of our proposed system.