Yi-Jun Li

Sentiment classification for Chinese reviews: a comparison between SVM and semantic approaches

Web content mining is intended to help people to discover valuable information from large amount of unstructured data on the Web. Sentiment classification aims to mining the Web content of product reviews by classifying the reviews into positive or negative opinions. Such kind of classification approaches could help both consumers and sellers in making their decisions. But it is also a complicated task with great challenge. This paper conducted a comparison between the SVM approach and semantic approach for sentiment classification of Chinese reviews and also proposed some improvement for sentiment classification approaches. Experimental result indicated that, compared with previous researches for English reviews, the performance of both approaches for Chinese reviews sentiment classification are acceptable, while the support vector machine approach has better performance than the semantic orientation approach.

Time series forecasting with RBF neural network

Radial basis function neural network (RBF NN) has been widely used for nonlinear system identification because of its simple topological structure and its ability to reveal how learning proceeds in an explicit manner. In this paper, descriptions and original applications of RBF NN, to the time series forecasting problem is presented. Genetic algorithm technique is proposed to improve the RBF center placement quality. The research contributes to the applications of RBF NN by experiments with real-world data sets. Experimental results reveal that the prediction performance of RBF NN is significantly better than a traditional BP NN model.

Research on event prediction in time-series data

Event prediction in time series is an important problem with many real world applications. Existing statistical and machine learning methods are not suitable for the problem. This paper describes a neural network system that predicts events by identifying features extracted from time-series data. A new feature extraction method is proposed and a corresponding clustering method is given. The method is applied to real time series and the resulting generalization performance of the trained feed-forward neural network predictors is analyzed. It shows that the method is effective in event prediction.

Agent-based design and organization of intermodal freight transportation systems

In logistics of todays economy we have to deal with distributed systems. To support the demanding management task the multi-agent approach offers promising perspectives. Intermodal freight transportation (IFT) is a complex task involving a lot of information flowing across organizational boundaries. However, the interorganizational information flows are often slow, and sometimes even are blocked, due to lack of standardized data storage format, no rules for arranging automated transfer from one system to another, and so on. This paper presents a multi-agent system (MAS) approach to the design and organization of an integrated IFT information architecture. We propose to organize the IFT system as an open network of autonomous agents, each encapsulating one or more logistics roles, and each coordinating its activities with other agents. The architecture is first described using an abstract specification based on concepts such as agents, roles and role protocols, and then specified in details using an example. A major conclusion is that the MAS based architecture can support IFT parties to integrate their information systems in an effort to truly reach a seamless intermodal transport.

Application and study of spatial cluster and customer partitioning

Along with the development of database technology and information collection methods, spatial data mining has become more and more important, and presents new challenges that are for the large size of spatial data and complexity of spatial data types. This paper introduces the research of current spatial cluster algorithms, and we propose a new and efficient algorithm based on the theory of partitioning methods, grid-based methods and density methods. This algorithm can find arbitrarily-shape clusters without any previous knowledge, and scale well for large data sets due to its computational complexity not to connect with the number of objects. This paper employs spatial cluster in the customer partitioning of business management so as to solve spatial analysis and location.

Study on the location of distribution center: a genetic algorithm combining mechanism of simulated annealing

The location of distribution center (DC) is one of the most important decision issues for logistics managers. In this paper, 0-1 mixed integer linear programming model of location of distribution center of two-stage logistics chain network is given. The design tasks of this problem involve the choice of distribution center to be opened and the optimal assignment of product flow between suppliers and customers via distribution centers to meet the demand with minimum cost As the solution method, a genetic algorithm combining mechanism of simulated annealing that is proved effective in dealing with the NP-hard problem is proposed.

Dynamic pricing: ecommerce - oriented price setting algorithm

Pricing in electronic commerce is based on bargaining. Pricing models that can fast change prices during transaction on the consequence of the buyers needs is beneficial to electronic commerce. Demand sensitive model is one of the pricing models that can be used for fast changes of prices in electronic commerce. Price setting algorithm for demand sensitive model helps sellers to get decision variables, price per unit that maximizes profit for the quantity ordered by buyers. In this paper we analyze the price setting algorithms of demand sensitive model. We use a simple example to explain how the changes of price elasticity of demand changes price per unit, gross margin and quantity demanded. Also we show how the changes of quantity demanded changes the unit price and marginal cost. The investigation shows that an increase in demand ordered decreases price per unit of a good, at the same time increasing profit margin to seller and decreasing production cost. The seller extracts some of the buyers surplus value as profits with residual surplus remaining with the buyer over and above the actual price paid. Buyers do not pay the same amount of total price for the good ordered within the same group of order, because of the difference in the net browsing cost.

An improved feature extraction approach based on Rough Sets for the medical diagnosis

This paper presents a novel approach based on rough sets to extract the complicated features from the medical diagnosis corpus. Some symptoms or basic features in the medical diagnosis are usually correlated. In general, the combinations of several basic symptoms may represent the disease more precision. However, the overmuch feature can reduce the generalization ability, or even many unfit features as the noise can decrease the modelpsilas performance. This paper proposes to apply the rough set theory to mine the complicated features, even from noise or inconsistent corpus. Secondly, these complex features are added into the maximum entropy model or support vector machine etc. as a new kind of features, consequently, the feature weights can be assigned according to the performance of the whole model. The experiments in the liver-disorders repository show that our method can improve the maximum entropy model by the precision 3.51%, improve the support vector machine model by the precision 3.05%, improve the naive Bayes model by the precision 3.59%, and improve the Bayes and GoodTuring model by the precision 3.59%.

Study on calculating method of induced traffic volume

As there is some limitation of current calculating method in induced traffic volume on expressway, according to the change law of induced traffic volume, this paper introduces the mathematical model of quantitative calculating and its optimized solution in order to enhance the computation precision of induced traffic volume. In addition, after calculating the actual induced traffic volume on Xi-Bao and Hu-Ning expressway, on the basis of induction of concrete evidences, the paper has summarized the change features and the eigenvalue of induced traffic volume in different areas.

Application of artificial immune algorithm in e-commerce oriented negotiation support system

Negotiation support system has become increasingly important since the advent of electronic commerce. But the existing NSS cannot conveniently, timely and effectively help negotiators make decisions in the real-time bargaining. Since immune algorithms have the advantages of converging global search space fast, intrinsic parallelism and resolving the optimization problems with simulating the immune system, in this paper we attempt to apply artificial immune algorithm to providing solution supporting in the electronic commerce oriented NSS in order to offer optimal or near-optimal solutions for negotiators. We verify the feasibility and special superiority of this method through theory analysis and experiments.