Jiwen Guan

COMBINING MULTIPLE CLASSIFIERS USING DEMPSTER'S RULE FOR TEXT CATEGORIZATION

In this paper we investigate the combination of four machine learning methods for text categorization using Dempsters rule of combination. These methods include Support Vector Machine (SVM), kNN (Nearest Neighbor), kNN model-based approach (kNNM), and Rocchio. We first present a general representation of the outputs of different classifiers, in particular, modeling it as a piece of evidence by using a novel evidence structure called focal element triplet. Furthermore, we investigate an effective method for combining pieces of evidence derived from classifiers generated by a 10-fold cross-validation. Finally, we evaluate our methods on the 20-newsgroup and Reuters-21578 benchmark data sets and perform the comparative analysis with majority voting in combining multiple classifiers along with the previous result. Our experimental results show that the best combined classifier can improve the performance of the individual classifiers and Dempsters rule of combination outperforms majority voting in combining multiple classifiers.

Combining Evidence from Classifiers in Text Categorization

In this paper, we describe a way for modelling a generalization process involved in the combination of multiple classification systems as an evidential reasoning process. We first propose a novel structure for representing multiple pieces of evidence de‘rived from multiple classifiers. This structure is called a focal element triplet. We then present a method for combining multiple pieces of evidence by using Dempster’s rule of combination. The advantage of the novel structure is that it not only facilitates the distinguishing of trivial focal elements from important ones, but it also reduces the effective computation-time from exponential as in the conventional process of combining multiple pieces of evidence to linear. In consequence, this allows Dempster’s rule of combination to be implemented in a widened range of applications.

Risk assessment of e-commerce projects using evidential reasoning

The purpose of this study is to develop a decision making system to evaluate the risks in E-Commerce (EC) projects. Competitive software businesses have the critical task of assessing the risk in the software system development life cycle. This can be conducted on the basis of conventional probabilities, but limited appropriate information is available and so a complete set of probabilities is not available. In such problems, where the analysis is highly subjective and related to vague, incomplete, uncertain or inexact information, the Dempster-Shafer (DS) theory of evidence offers a potential advantage. We use a direct way of reasoning in a single step (i.e., extended DS theory) to develop a decision making system to evaluate the risk in EC projects. This consists of five stages 1) establishing knowledge base and setting rule strengths, 2) collecting evidence and data, 3) determining evidence and rule strength to a mass distribution for each rule; i.e., the first half of a single step reasoning process, 4) combining prior mass and different rules; i.e., the second half of the single step reasoning process, 5) finally, evaluating the belief interval for the best support decision of EC project. We test the system by using potential risk factors associated with EC development and the results indicate that the system is promising way of assisting an EC project manager in identifying potential risk factors and the corresponding project risks.

Data mining for motifs in DNA sequences

In the large collections of genomic information accumulated in recent years there is potentially significant knowledge for exploitation in medicine and in the pharmaceutical industry. One interesting approach to the distillation of such knowledge is to detect strings in DNA sequences which are very repetitive within a given sequence (eg for a particular patient) or across sequences (eg from different patients who have been classified in some way eg as sharing a particular medical diagnosis). Motifs are strings that occur relatively frequently. In this paper we present basic theory and algorithms for finding such frequent and common strings. We are particularly interested in strings which are maximally frequent and, having discovered very frequent motifs we show how to mine association rules by an existing rough sets based technique. Further work and applications are in process.

A New Algorithm for Mining Sequential Patterns

AprioriAll and AprioriSome are very famous algorithms for mining sequential patterns, which are used to find motifs on a fixed min-support number. In this paper, we contribute a new algorithm that can find all motifs on any min-support numbers.

A Decision Assistant Based on Evidential Reasoning

The evidence theory provides an important reasoning mechanism in artificial intelligence, and it has been applied to complex systems to handle uncertainty reasoning. Using a set of easy-to-use interfaces, we have designed a desktop decision assistant system based on the evidence theory. Through the easy-to-use interfaces, expert knowledge and opinions or political arguments can be input to the system easily. An incremental evidence combination algorithm is embedded in a reasoning module in the system. Based on the evidential reasoning module, the system can carry out uncertainty reasoning efficiently. Therefore, the system can be applied to enhance human capability to make decisions in daily life, business, or political arguments. As the easy-to-use interfaces are designed, the decision assistant can make decisions in different situations with only a few keyboard clicks. A couple of examples of applications are also demonstrated in this paper

Discovering maximal frequent patterns in sequence groups

In this paper, we give a general treatment for some kind of sequences such as customer sequences, document sequences, and DNA sequences, etc. Large collections of transaction, document, and genomic information have been accumulated in recent years, and embedded latently in it there is potentially significant knowledge for exploitation in the retailing industry, in information retrieval, in medicine and in the pharmaceutical industry, respectively. The approach taken here to the distillation of such knowledge is to detect strings in sequences which appear frequently, either within a given sequence (eg for a particular customer, document, or patient) or across sequences (eg from different customers, documents, or patients sharing a particular transaction, information retrieval, or medical diagnosis; respectively).

Finding motifs in a set of DNA sequences: A dynamic programming approach

The search for motifs in a DNA sequence set is a generic problem area that is of great interest bioinformatics. Given a set of n DNA sequence and a support-rate threshold τ, it is useful to find maximal patterns that occur in at least τn sequences in the set. This paper presents an efficient approach to find motifs for any support-rate threshold and without any miss. The idea is to prune the candidate set of maximal patterns while finding patterns satisfying the given threshold using the dynamic programming method and adding them to the candidate set. Theoretical analysis shows that this approach is efficient and preliminary experiments show that the runtime performance of this approach is satisfactory.

A novel discretizer for knowledge discovery approaches based on rough sets

Knowledge discovery approaches based on rough sets have successful application in machine learning and data mining. As these approaches are good at dealing with discrete values, a discretizer is required when the approaches are applied to continuous attributes. In this paper, a novel adaptive discretizer based on a statistical distribution index is proposed to preprocess continuous valued attributes in an instance information system, so that the knowledge discovery approaches based on rough sets can reach a high decision accuracy. The experimental results on benchmark data sets show that the proposed discretizer is able to improve the decision accuracy

A MIXED RADIX APPROACH TO THE POOLING OF EVIDENCE

This paper proposes an innovative approach to the implementation of an evidence-pooling method used in Artificial Intelligence. We present a direct hardware implementation of the intersection table which is used as a device for calculating the Dempster-Shafer orthogonal sum. This operation is a well known means of combining evidence from different sensors for use in reasoning. The proposed hardware makes use of a mixed-radix (i.e.,