Search

Applications to massive data raise two challenges to supervised learning. First, sufficient training examples to ensure the generalization ability become unavailable, since labelling by experts is expensive; second, it is impossible to load massive data into memory, and the response time is unacceptable by a serial mode. In this paper, we gracefully combine semi-supervised learning with parallel computing to meet these two challenges together. In detail, (1) the co-training style of semi-supervised learning named tri-training is exploited and revised in order to perform learning from the labelled and the unlabelled data. In particular, the co-training process is revised by introducing data editing to remove the newly mislabelled data. (2) The learning algorithm for each individual classifier and the data editing are re-formed as the MapReduce parallel pattern. Experiments on University of California, Irvine, Machine Learning Repositoy data sets and the application to CT images detection show improvement i...

Computer-aided diagnosis (CAD) has become a major research topic in medical imaging, and one of the most important CAD applications is the detection of lung nodules. The paper is to develop a CAD system for automatically detecting lung nodules in computed tomography (CT) images. The system includes three parts: pulmonary parenchyma segmentation, ROI extraction, and nodule prediction of ROI based on ADE-Co-Forest. At the beginning, we proposed the new pulmonary parenchyma segmentation method; In the stage of ROI extraction, circle shape descriptor is exploited to reduce the false positives; Although the samples can be easily collected from routine medical examinations, it is usually impossible for medical experts to make a diagnosis for each of the collected samples. So we use the semi-supervised learning method ADE-Co-Forest to predict the nodules. Thus, in the predicting stage, we can use a few of labeled samples and a lot of unlabeled samples to learn a well-performed classifier. The experimental results demonstrate that the CAD system gets high sensitivity and low false-positive.

Mislabeling unlabeled data during the learning process is an inevitable problem for the co-training style semi-supervised learning. In this paper, the participatory learning cognition paradigm is instantiated through employing the data editing as acceptance unit and designing an arousal strategy of data editing as critic unit. Then, this participatory learning is equipped into each individual classifier of Tri-training, a co-training style semi-supervised approach, and forms a new algorithm named PL-Tri-training (participatory learning based Tri-training). In the co-training process of PL-tri training,the acceptance unit utilizes data editing to identify and remove the mislabeled data, as well as the critic unit exploits arousal strategy to inhibit the invalid activation of data editing. The experiments on UCI datasets show that PL-Tri-training can more effectively and stably exploit the unlabeled data to improve the classification performance than Tri-training and DE-Tri training, which equips the Tri-training with only the data editing acceptance unit of participatory learning.

Inference of evolutionary trees using the maximum likelihood principle is NP-hard. Therefore, all practical methods rely on heuristics. The topological transformations often used in heuristics are nearest neighbor interchange (NNI), sub-tree prune and regraft (SPR) and tree bisection and reconnection (TBR). However, these topological transformations often fall easily into local optima, since there are not many trees accessible in one step from any given tree. Another more exhaustive topological transformation is p-Edge Contraction and Refinement (p-ECR). However, due to its high computation complexity, p-ECR has rarely been used in practice. This paper proposes a method p-ECRNJ with a O(p3) time complexity to make the p-ECR move efficient by using neighbor joining (NJ) to refine the unresolved nodes produced in p-ECR. Moreover, the demonstrated topological accuracy for small datasets of NJ can guarantee the accuracy of the p-ECRNJ move. Experiments with simulated and real datasets show that p-ECRNJ can find better trees than the best-known maximum likelihood methods so far and can efficiently improve local topological transforms in reasonable time.

A method is illustrated in this paper for edge detection of moving objects in video sequences, the method synthetically combines LS-SVM (Least Squares-Support Vector Machines) algorithm with motion recovery theory for edge detection. The experimental results show that this method can provide accurate edge detection result effectively and overcomes fuzzy parts of image caused by motion. Due to low computation complexity, it has fine adaptability and robustness.

Multi-agent negotiation has been one of the key problems in the multi-agent research area. An extended contract-net negotiation model based on task coalition and genetic algorithm is presented after analyzing the advantage and disadvantage of the classical contract-net negotiation model. Formalized definition method and coalition generation algorithm are given. A specialized genetic algorithm, which is optimized by optimized initial colony selection, optimized parent crossover/mutation and the using of Metropolis rule, is used to solve the task allocation in the coalition. The algorithm improves the efficiency of task allocation and reduces the communication cost. By testing and analyzing an example of a missile defense system, it is proved that the model can reduce the negotiation cost effectively contrast with the classical contract-net model on the basis of ensuring the negotiation quality.

A multi-agent system (MAS) focuses on the interaction among the agents, and the communication mechanism is the basis for the interaction. This paper puts forward a MAS communication model based on a star-ring structure. This work proposes a MAS communication model which combines the characters of the platform architecture based on analysis of communication among the agents. The model assumes that the communication server's structure is more complex and the agent's communication module is relative simple. This MAS communication model logically directly transfers messages between agents, and messages are physically stored and transmitted. Analysis of this model indicates that this mechanism is agile and dependable, and can support many kinds of communication requirements in MAS systems. The model will help users conveniently and rapidly construct a stable and efficient MAS system.

Seeds based semi-supervised clustering algorithms often utilize a seeds set consisting of a small amount of labeled data to initialize cluster centroids, hence improve the performance of clustering over whole data set. Researches indicate that both the scale and quality of seeds set greatly restrict the performance of semi-supervised clustering. A novel semi-supervised clustering algorithm named DE-Tri-training semi-supervised K means is proposed. In new algorithm, prior to initializing cluster centroids, the training process of a semi-supervised classification approach named Tri-training is used to label the unlabeled data and add them into initial seeds to enlarge the scale. Meanwhile, to improve the quality of enlarged seeds set, a Nearest Neighbor Rule based data editing technique named Depuration is introduced into the Tri-training process to eliminate and correct the noise and mislabeled data among the enlarged seeds. Experiments show that novel algorithm can effectively improve the initialization of cluster centroids and enhance clustering performance.

The ever growing popularity of the Internet as a source of information, coupled with the accompanying growth in the number of documents available through the World Wide Web, is leading to an increasing demand for more efficient and accurate information retrieval tools. One of the fundamental problems in information retrieval is word mismatch. Expanding a user’s query with related words can improve the search performance, but the finding and using of related words is still an open problem. On the basis of previous approaches to query expansion, this paper proposes a new approach to query expansion that combines two popular traditional methods—thesauri and automatic relevance feedback. According to theoretical analysis and experiments, the new approach can effectively improve the web retrieval performance and out-performs the optimized conventional expansion approaches.

Being not down-compatible with classical two-value logic inference mechanism, fuzzy inference mechanism (FIM) would get absurd conclusions even in the simplest cases. To strictly evaluate down-compatibility degree and to build new FIM with high down-compatibility degree, a series of formulized criterions are presented based on compatibility of truth space and match modes, and logic consistence of conclusions derived by homologous composite premises from different FIMs. These criteria could offer both formal evaluation scales and revise guidelines. According to these criteria, some FIMs are lower in terms of down-compatibility degree and then revised up to the highest degree under its guidelines.

The similarities between the BP network and the rule representation in the uncertain reasoning based on certainty factor are introduced in this paper. According to the disadvantage of the traditional rule representation, a method is presented to train and correct the CF value of the rule and the weight of the proof, which makes the system possess the ability of self-learning. After the method was applied in the cow disease diagnosis expert system, and its performance has been improved.

Searching online text collections can be both rewarding and frustrating. On the same time valuable information can be found, typically many irrelevant documents are also retrieved and many relevant ones are missed. Word mismatches between the user's query and document contents are a main cause of retrieval failures. Expanding a user's query with related words can improve search performance, but finding and using related words is an open problem. On the basis of previous approaches to query expansion, this paper proposes a new approach to query expansion, which combines two popular traditional methods -thesauri and automatic relevance feedback. In terms of theoretical analysis and experiments, the new approach is effective to query expansion for Web retrieval and out-performs the optimized, conventional expansion approaches.

Exploration is used in Q/spl I.bar/learning because the agent would be caught in locally optimal policies due to blind exploitation. However excessive exploration would degrade the performance of Q/spl I.bar/learning and it is difficult to meet the trade-off between exploration and exploitation. The active backup is introduced into Q/spl I.bar/learning and the corresponding algorithm AB/spl I.bar/Q/spl I.bar/learning based on Dijkstra backup in dynamic programming is proposed. Then, the memory mechanism based MEAB/spl I.bar/Q/spl I.bar/Iearning algorithm is given for the agent to learn in completely unknown environment. The experimental results show that these two algorithms not only converge more quickly, but also solve the problem of local optimization.