Chunxiao Xing

Similarity measure and instance selection for collaborative filtering

Collaborative filtering has been very successful in both research and applications such as information filtering and E-commerce. The k-Nearest Neighbor (KNN) method is a popular way for its realization. Its key technique is to find k nearest neighbors for a given user to predict his interests. However, this method suffers from two fundamental problems: sparsity and scalability. In this paper, we present our solutions for these two problems. We adopt two techniques: a matrix conversion method for similarity measure and an instance selection method. And then we present an improved collaborative filtering algorithm based on these two methods. In contrast with existing collaborative algorithms, our method shows its satisfactory accuracy and performance.

A new approach to identify influential spreaders in complex networks

In the research of the propagation model of complex network, it is of theoretical and practical significance to detect the most influential spreaders. Global metrics such as degree centrality, closeness centrality, betweenness centrality and K-shell centrality can be used to identify the influential spreaders. These approaches are simple but have low accuracy. We propose K-shell and Community centrality (KSC) model. This model considers not only the internal properties of nodes but also the external properties of nodes, such as the com-munity which these nodes belong to. The Susceptible-Infected-Recovered (SIR) model is used to evaluate the performance of KSC model. The experiment result shows that our method is better to identify the most influential nodes. This paper comes up with a new idea and method for the study in this field.

Sentiment classification for stock news

Web news articles play an important role in stock market. Sentiment classification of news articles can help the investors make investment decisions more efficiently. In this paper, we implemented an approach of Chinese new words detection by using N-gram model and applied the result for Chinese word segmentation and sentiment classification. Appraisal theory was introduced into sentiment analysis and Naive Bay es, K-nearest Neighbor and Support Vector Machine were used as classification algorithms. Our method was used for a Chinese stock news data set. The best accuracy reaches 82.9% in all experiments. Additionally, we developed a prototype system to demonstrate our work.

HealthQA: A Chinese QA Summary System for Smart Health

Although online health expert QA services can provide high quality information for health consumers, there is no Chinese question answering system built on knowledge from existing expert answers, leading to duplicated efforts of medical experts and reduced efficiency. To address this issue, we develop a Chinese QA system for smart health (HealthQA), which provides timely, automatic and valuable QA service. Our HealthQA collects diabetes expert question answer data from three major QA websites in China. We develop a hierarchical clustering method to group similar questions and answers, an extended similarity evaluation algorithm for retrieving relevant answers and a ranking based summarization for representing the answer. ROUGE and manual tests show that our system significantly outperforms the search engine.

Data Source Recommendation for Building Mashup Applications

The emergence of mashup is gaining tremendous popularity and its application can be seen in a large number of domains. Along with the development of mashup technology, several mashup editors have been produced by the industry which can assist users to build mashups. However, with the increasing service and information sources distributed across the entire web space, even an easy to use mashup editor for nonprogrammers is not sufficient. In this paper, we apply the item based top-N recommendation algorithm which is widely used in e-Commerce area to recommend data source to the users while they are building mashups based on collected data of existing mashups. We also conduct experiment to evaluate the parameters of the recommendation algorithm and finally achieve very satisfactory results.

THVDM: A Data Model for Video Management in Digital Library

As an important part of digital library, video library is becoming more and more significant. However, the problem in lacking of appropriate data models prevents the video library from being widely used. In this paper, we first review some existing data models, and then present a new model consisting of three layers: feature layer, semantic layer, and knowledge layer. The model supports non-procedural query language and overcomes the shortcoming of existing models, thus can meet the requirements of new applications.

Personalized Services for Digital Library

In this paper, we describe a prototype system, called PASS (Personalized Active Service System), which provides personalized services for digital libraries. User profiles are represented as probabilistic distributions of interests over different domains. The system realizes content-based filtering by computing the similarity of probabilistic distributions between documents and user profiles. In addition the system realizes collaborative filtering by clustering similar user profiles. Experimental results show its performance satisfactory.

A Packaging Approach for Massive Amounts of Small Geospatial Files with HDFS

The efficiency of dealing with massive small geospatial files deeply affects the performance of Web Geography Information System (WebGIS). The Hadoop Distributed File System (HDFS) is scalable to satisfy the requirement of massive data files storage, but not efficient in dealing with small files. In this paper, we proposed a method to pack a group of small files into one large logical file, and set up Hilbert spatial index inside the block with their spatial adjacency relation. The experimentation proved that this method reduces the size of block indices and increases the speed to search and retrieve the massive small spatial files.

A Cache Replacement Algorithm in Hierarchical Storage of Continuous Media Object

Although storage technologies are in great progress, hierarchical storage is still an efficient way to manage massive data, especially massive Continuous Media (CM) data. In some large web based CM applications such as DL (Digital Library) and VOD (Video-On-Demand), cache replacement algorithms are used to keep the hot data on the disk, and migrate out those rarely accessed. The existing algorithms always treat all files equally, and use cache miss rate to evaluate the performance. However, in actual CM applications, most video objects are stored as several separated files that will be played one by one, so the strong relationships among the access time of these files make pre-fetch practicable and efficient. If all the files are taken as the same, the performance of systems will degrade. Meanwhile, the cache miss rate cannot evaluate the performance comprehensively. In this paper, a new metric, User Waiting Rate, is defined to evaluate the performance, and a novel cache replacement algorithm, Least Waiting Probability (LWP) algorithm, is proposed. Experiment results show that it can improve the performance a lot, and is highly adaptive.

Developing Tsinghua University Architecture Digital Library for Chinese Architecture Study and University Education

The Tsinghua University Architecture Digital Library (THADL) has been developed as a prototype system with dual goals - to study the key techniques of digital libraries (DLs), and to provide rich, valuable resources for Chinese architecture research and education as well as a high level of service. This paper focuses on the following issues: 1) the novel architecture of THADL and its implementation; 2) the metadata specification--THADL Metadata 2.0; 3) the OAI PMH v2.0 implementation to make THADL interoperable with other digital resources; 4) the Web-based courseware environment for students majoring in architecture science; and 5) the personalized active service subsystem -- THPASS.