Search

In this paper, we propose an algorithm for maintaining the frequent itemsets discovered in a database with minimal re-computation when new transactions are added to or old transactions are removed from the transaction database. An efficient algorithm called EFPIM (Extending FP-tree for Incremental Mining), is designed based on EFP-tree (extended FP-tree) structures. An important feature of our algorithm is that it requires no scan of the original database, and the new EFP-tree structure of the updated database can be obtained directly from the EFP-tree of the original database. We give two versions of EFPIM algorithm, called EFPIM1 (an easy vision to implement) and EFPIM2 (a fast algorithm), they both mining frequent itemsets of the updated database based on EFP-tree. Experimental results show that EFPIM outperforms the existing algorithms in terms of the execution time. [ABSTRACT FROM AUTHOR]

This paper contributes to the ongoing research on multidimensional online association rules mining by proposing a general architecture that utilizes a fuzzy data cube combined with the concepts of weight and multiple-level to mine fuzzy weighted multi-cross-level association rules. We compared the proposed approach to an existing approach that does not utilize fuzziness. Experimental results on the adult data of the United States census in year 2000 demonstrate the effectiveness and applicability of the proposed fuzzy OLAP based mining approach. Keywords: association rules, data mining, fuzzy data cube, multidimensional mining, OLAP, weighted mining. [ABSTRACT FROM AUTHOR]

The issue of maintaining privacy in frequent itemset mining has attracted considerable attentions. In most of those works, only distorted data are available which may bring a lot of issues in the data-mining process. Especially, in the dynamic update distorted database environment, it is nontrivial to mine frequent itemsets incrementally due to the high counting overhead to recompute support counts for itemsets. This paper investigates such a problem and develops an efficient algorithm SA-IFIM for incrementally mining frequent itemsets in update distorted databases. In this algorithm, some additional information is stored during the earlier mining process to support the efficient incremental computation. Especially, with the introduction of supporting aggregate and representing it with bit vector, the transaction database is transformed into machine oriented model to perform fast support computation. The performance studies show the efficiency of our algorithm. [ABSTRACT FROM AUTHOR]

Mining user profiles is a crucial task for Web usage mining, and can be accomplished by mining frequent patterns. However, in the Web usage domain, sessions tend to be very sparse, and mining the right user profiles tends to be difficult. Either too few or too many profiles tend to be mined, partly because of problems in fixing support thresholds and intolerant matching. Also, in the Web usage mining domain, there is often a need for post-processing and validation of the results of mining. In this paper, we use criterion guided optimization to mine localized and error-tolerant transaction patterns, instead of using exact counting based method, and explore the effect of different post-processing options on their quality. Experiments with real Web transaction data are presented. [ABSTRACT FROM AUTHOR]

In this paper, neural network techniques based on Kohonen's self-organising map method which can be trained in an unsupervised fashion and applicable to the processing of graph structured inputs are described. Then it is shown how such techniques can be applied to the problems of clustering of XML documents. [ABSTRACT FROM AUTHOR]

Traditional researches on data stream mining only put emphasis on building classifiers with high accuracy, which always results in classifiers with dramatic drop of accuracy when concept drifts. In this paper, we present our RTRC system that has good classification accuracy when concept drifts and enough samples are scanned in data stream. By using Markov chain and least-square method, the system is able to predict not only on which the next concept is but also on when the concept is to drift. Experimental results confirm the advantages of our system over Weighted Bagging and CVFDT, two representative systems in streaming data mining. [ABSTRACT FROM AUTHOR]

Temporal data mining is concerned with the analysis of temporal data and finding temporal patterns, regularities, trends, clusters in sets of temporal data. In this paper we extract regression features from the coefficients obtained by applying WaveSim Transform on Multi-Channel signals. WaveSim Transform is a reverse approach for generating Wavelet Transform like coefficients by using a conventional similarity measure between the function f(t) and the wavelet. WaveSim transform provides a means to analyze a temporal data at multiple resolutions. We propose a method for computing principal components when the feature is of linear regression type i.e. a line. The resultant principal component features are also lines. So through PCA we achieve dimensionality reduction and thus we show that from the first few principal component regression lines we can achieve a good classification of the objects or samples. The techniques have been tested on an EEG dataset recorded through 64 channels and the results are very encouraging. [ABSTRACT FROM AUTHOR]

Our work is within the framework of studying and implementing a sound analysis system in a telemedicine project. The task of this system is to detect situations of distress in a patient's room based on sound analysis. If such a situation is detected, an alarm will be automatically sent to the medical centre. In this paper we present our works on building domain ontology of such situations. They gather abstract concepts of sounds and these concepts, along with their properties and instances, are represented by a neural network. The ontology-based classifier uses outputs of networks to identify classes of audio scenes. The system is tested with a database extracted from films. [ABSTRACT FROM AUTHOR]

The efficiency of skyline query processing has recently received a lot of attention in database community. However, researchers often ignore that the skyline set will be beyond control in the applications which must deal with enormous data set. Consequently, it is not useful for users at all. In this paper, we propose a novel skyline reducing algorithm, i.e. SRANF. SRANF algorithm adopts the technique of noise filtering. It filters skyline noises directly on the original data set based on the acceptable difference, and returns the objects which can not be filtered from the original data set. Furthermore, our experiment demonstrated that SRANF is both efficient and effective. [ABSTRACT FROM AUTHOR]

We explore the role of sequences of extreme values for measuring tail-dependence between times series. The proposed measure concentrates on searching extreme cause-effect fluctuation pairs in the recent time interval and requires much less data than current causality and dependence approaches. The target applications of this approach are those in which there is the necessity of rapidly recognizing the interval time in which a time series may be influenced by other time series characterized by sudden and unpredictable extreme changes. This paper presents the tail-dependence measure in the field of stock markets and compares it to known causality and dependence measures. An application of the mentioned measure in the field of space physics is also presented. [ABSTRACT FROM AUTHOR]

Multi-dimensional sequential pattern mining attempts to find much more informative frequent patterns suitable for immediate use. In this paper, a novel data model called multi-dimensional concept lattice is proposed and, based on which, a new incremental multi-dimensional sequential pattern mining algorithm is developed. The proposed algorithm integrates sequential pattern mining and association pattern mining with a uniform data structure and makes the mining process more efficient. The performance of the proposed approach is evaluated on both synthetic and real-life financial date sets. [ABSTRACT FROM AUTHOR]

Sequential pattern mining is to find out all the frequent sub-sequences in a sequence database. In order to have more accurate results, constraints in addition to the support threshold need to be specified in the mining. Time-constraints cannot be managed by retrieving patterns because the support computation of patterns must validate the time attributes for every data sequence in the mining process. In this paper, we propose a memory time-indexing approach (called METISP) to discover sequential patterns with time constraints including minimum/maximum/exact gaps, sliding window, and duration. METISP scans the database into memory and constructs time-index sets for effective processing. Utilizing the index sets and the pattern-growth strategy, METISP efficiently mines the desired patterns without generating any candidate or sub-database. The comprehensive experiments show that METISP outperforms GSP and DELISP in the discovery of time-constrained sequential patterns, even with low support thresholds and very large databases. [ABSTRACT FROM AUTHOR]

Currently, many researches have been pursued for cardiovascular disease diagnosis using ECG so far. In this paper we extract multi-parametric features by HRV analysis from ECG, data preprocessing and heart disease pattern classification method. This study analyzes the clinical information as well as the time and the frequency domains of HRV, and then discovers cardiovascular disease patterns of patient groups. In each group, its patterns are a large frequency in one class, patients with coronary artery disease but are never found in the control or normal group. These patterns are called emerging patterns. We also use efficient algorithms to derive the patterns using the cohesion measure. Our studies show that the discovered patterns from 670 participants are used to classify new instances with higher accuracy than other reported methods [ABSTRACT FROM AUTHOR]

Unlike previous pattern-based biclustering methods that focus on grouping objects on the same subset of dimensions, in this paper, we propose a novel model of coherent cluster for time series gene expression data, namely td-cluster (time-delayed cluster). Under this model, objects can be coherent on different subsets of dimensions if these objects follow a certain time-delayed relationship. Such a cluster can discover the cycle time of gene expression, which is essential in revealing the gene regulatory networks. This work is missed by previous research. A novel algorithm is also presented and implemented to mine all the significant td-clusters. Experimental results from both real and synthetic microarray datasets prove its effectiveness and efficiency. [ABSTRACT FROM AUTHOR]

This paper focuses on the problem of choosing a representation of documents that can be suitable to induce more advanced semantic user profiles, in which concepts are used instead of keywords to represent user interests. We propose a method which integrates a word sense disambiguation algorithm based on the WordNet IS-A hierarchy, with two machine learning techniques to induce semantic user profiles, namely a relevance feedback method and a probabilistic one. The document representation proposed, that we called Bag-Of-Synsets improves the classic Bag-Of-Words approach, as shown by an extensive experimental session. Keywords: User Profiles, Text Categorization, Word Sense Disambiguation, WordNet. [ABSTRACT FROM AUTHOR]

Text clustering is one of the most important research areas in text mining, which handles the text automatically to discover implicit knowledge. It groups text into different clusters by contents without apriori knowledge. In this paper, different text clustering methods are studied and three text clustering validation criteria are studied and used to evaluate the experimental results. We compare and contrast the effectiveness of k-means and FIHC text clustering methods by experiments, and address the different levels of quality of the resulting text clusters. [ABSTRACT FROM AUTHOR]

This paper presents a new audiovisual integration scheme for racquet sports video structure indexing and highlight generating. Instead of using low-level features, the method is built upon the combination of visual and audio features. With respect to prior information about this kind of video content and editing rules, visual features based on dominant color and motion attention model are applied to classify shots into two classes: global view shots and non-global view shots. The classification algorithm is independent of predefined court color, and much robust to lighting conditions. Afterwards, among shots important auditory features including both ball hitting and applause are detected for identifying interesting events with strong semantic meaning, such as missed serves, aces, rallies and replays in tennis video. Finally, a reasonable model is built to rank rally events by excitement. The results showed the scheme could effectively identify typical scenes for retrieving highlights. [ABSTRACT FROM AUTHOR]

In this paper, we consider the problem of enumerating all maximal cliques in a complex network G = (V, E) with n vertices and m edges. We propose an algorithm for enumerating all maximal cliques based on researches of the complex network properties. A novel branch and bound strategy by considering the clustering coefficient of a vertex is proposed. Our algorithm runs with time O (d^2*N*S) delay and in O (n + m) space. It requires O (n*D^2) time as a preprocessing, where D, N, S, d denote the maximum degree of G, the number of maximal cliques, the size of the maximum clique, and the number of triangles of a vertex with degree D respectively. Finally, we apply our algorithm to the telecommunication customer-churn-prediction and the experimental results show that the application promotes the capabilities of the churn prediction system effectively. [ABSTRACT FROM AUTHOR]

This paper proposes an effective scoring scheme for feature selection in Text Mining, using characteristics of Small-World Phenomenon on the semantic networks of documents. Our focus is on the reservation of both syntactic and statistical information of words, rather than solely simple frequency summarization in prevailing scoring schemes, such as TFIDF. Experimental results on TREC dataset show that our scoring scheme outperforms the prevailing schemes. [ABSTRACT FROM AUTHOR]

Feature selection is an important data preprocessing step in data mining and pattern recognition. Many algorithms have been proposed in the past for simple patterns that can be characterised by a single feature vector. Unfortunately, these algorithms are hardly applicable to what are referred as complex patterns that have to be described by a finite set of feature vectors. This paper addresses the problem of feature selection for the complex patterns. First, we formulated the calculation of mutual information for complex patterns based on Gaussian mixture model. A hybrid feature selection algorithm is then proposed based on the formulated mutual information calculation (filter) and Baysian classification (wrapper). Experimental results on XM2VTS speaker recognition database have not only verified the performance of the proposed algorithm, but also demonstrated that traditional feature selection algorithms designed for simple patterns would perform poorly for complex patterns. [ABSTRACT FROM AUTHOR]

Semi-supervised classification uses a large amount of unlabeled data to help a little amount of labeled data for designing classifiers, which has good potential and performance when the labeled data are difficult to obtain. This paper mainly discusses semi-supervised classification based on CPN (Counter-propagation Network). CPN and its revised models have merits such as simple structure, fast training and high accuracy. Especially, its training scheme combines supervised learning and unsupervised learning, which makes it very conformable to be extended to semi-supervised classification problem. According to the characteristics of CPN, we propose a semi-supervised dynamic CPN, and compare it with other two semi-supervised CPN models using Expectation Maximization and Co-Training/Self-Training techniques respectively. The experimental results show the effectiveness of CPN based semi-supervised classification methods. [ABSTRACT FROM AUTHOR]

Pattern recognition problems that involve functional predictors has developed, specifically for spectral data. The classification of three peach varieties based on near infrared spectra was researched in the practical context. Principal component analysis (PCA) and artificial neural networks (ANN) were used for pattern recognition in this research. PCA is a very effective data mining way; it is applied to enhance species features and reduce data dimensionality. ANN with back propagation algorithm was used for the data compression tasks as well as class discrimination tasks. The first 9 principal components computed by PCA were applied as inputs to a back propagation neural network with one hidden layer. This model was used to predict the varieties of 15 unknown samples. The recognition rate of the model for the unknown sample was 100%. So this paper could offer an effective pattern recognition way. [ABSTRACT FROM AUTHOR]

We propose several simple techniques which dramatically reduce both the memory demand and computational effort in building multinomial mixture models using the EM algorithm. The reason of the dramatic improvement in performance is that the techniques make use of certain properties of the data. These properties are: the data is sparse and there are many repeating records. We claim that particular sources of data consistently satisfy these properties. Excellent examples are Clickstream and retail data which are very sparse and consist of many repititions. Using simple techniques huge speed-ups and compression rates, on real life clickstream data sets, are observed compared to the standard implementation of the EM. [ABSTRACT FROM AUTHOR]

We hypothesize that the variance in volume of high-velocity queries over time can be explained by observing that these queries are formulated in response to events in the world that users are interested in. Based on it, this paper describes a system, Zed, which automatically finds explanations for high velocity queries, by extracting descriptions of relevant and temporally-proximate events from the news stream. Zed can thus provide a meaningful explanation of what the general public is interested in at any time. We evaluated performance of several variant methods on top velocity "celebrity name" queries from Yahoo, using news stories from several sources for event extraction. Results bear out the event-causation hypothesis, in that Zed currently finds acceptable event-based explanations for about 90% of the queries examined. [ABSTRACT FROM AUTHOR]

In this paper, a new method for learning Bayesian networks structure with continuous variables is proposed. The continuous variables are discretized based on hybrid data clustering. The discrete values of a continuous variable are obtained by using father node structure and Gibbs sampling. Optimal dimension of discretized continuous variable is found by MDL principle to the Markov blanket. Dependent relationship is refined by optimization regulation to Bayesian network structure in iteration learning. [ABSTRACT FROM AUTHOR]

We present in this paper an interactive method for numeric or symbolic data visualization that allows a domain expert to extract useful knowledge and information. We propose a new approach based on points of interest (POI) but in the context of visual data mining. POIs are located on a circle, and data are displayed within this circle according to their similarities to these POI. Interactive actions are possible: selection, zoom, dynamical change of POI. We evaluate the properties of such visualization with standard data with known characteristics. We describe an industrial application which explores results from satisfaction inquiries. [ABSTRACT FROM AUTHOR]

Accurate volatility forecasting is the core task in the risk management in which various portfolios' pricing, hedging, and option strategies are exercised. Prior studies on stock market have primarily focused on estimation of stock price index by using financial time series models and data mining techniques. This paper proposes hybrid models with neural network and time series models for forecasting the volatility of stock price index in two view points: deviation and direction. It demonstrates the utility of the hybrid model for volatility forecasting. [ABSTRACT FROM AUTHOR]

Clustering is a major technique in data mining. However the numerical feedback of clustering algorithms is difficult for user to have an intuitive overview of the dataset that they deal with. Visualization has been proven to be very helpful for high-dimensional data analysis. Therefore it is desirable to introduce visualization techniques with user's domain knowledge into clustering process. Whereas most existing visualization techniques used in clustering are exploration oriented. Inevitably, they are mainly stochastic and subjective in nature. In this paper, we introduce an approach called HOV3 (Hypothesis Oriented Verification and Validation by Visualization), which projects high-dimensional data on the 2D space and reflects data distribution based on user hypotheses. In addition, HOV3 enables user to adjust hypotheses iteratively in order to obtain an optimized view. As a result, HOV3 provides user an efficient and effective visualization method to explore cluster information. [ABSTRACT FROM AUTHOR]

Clustering ensembles have emerged as a powerful method for improving both the robustness and the stability of unsupervised classification solutions. However, finding a consensus clustering from multiple partitions is a difficult problem that can be approached from graph-based, combinatorial or statistical perspectives. A consensus scheme via the k-modes algorithm is proposed in this paper. A combined partition is found as a solution to the corresponding categorical data clustering problem using the k-modes algorithm. This study compares the performance of the k-modes consensus algorithm with other fusion approaches for clustering ensembles. Experimental results demonstrate the effectiveness of the proposed method. [ABSTRACT FROM AUTHOR]

This paper introduces a new method for solving clustering ensembles, that is, combining multiple clusterings over a common dataset into a final better one. The ensemble is reduced to a graph that simultaneously models as vertices the original clusters in the ensemble and the joint clusters derived from them. Only edges linking vertices from different types are considered. The resulting graph can be partitioned efficiently to produce the final clustering. Finally, the proposed method is evaluated against two graph formulations commonly used. [ABSTRACT FROM AUTHOR]

Scientific gait (walking) analysis provides valuable information about an individual's locomotion function, in turn, to assist clinical diagnosis and prevention, such as assessing treatment for patients with impaired postural control and detecting risk of falls in elderly population. While several artificial intelligence (AI) paradigms are addressed for gait analysis, they usually utilize supervised techniques where subject groups are defineda priori. In this paper, we explore to investigate gait pattern mining with clustering-based approaches, in which k-means and hierarchical clustering algorithms are employed to derive gait pattern. After feature selection and data preparation, we conduct clustering on the constructed gait data model to build up pattern-based clusters. The centroids of clusters are then treated as the subject profiles to model the various kinds of gait pattern, e.g. normal or pathological. Experiments are undertaken to visualize the derived subject clusters, evaluate the quality of clustering paradigm in terms of silhouette and mean square error and compare the results with the discovery derived from hierarchy tree analysis. In addition, analysis conducted on test data demonstrates the usability of the proposed paradigm in clinical applications. [ABSTRACT FROM AUTHOR]

This paper analyses some important characteristics of self-organization map network. Based on this analysis, we propose a method that can overcome the insufficiencies of single self-organization feature map (SOFM) network. The implementation detail of our proposed self-organizing feature map network algorithm is also discussed. Our proposed algorithm has a number of advantages. It can overcome the insufficiencies identified in other similar clustering algorithms. It is able to find clusters in different shapes and is insensitive to input data sequence. It can process noisy and multi-dimensional data well in multi-resolutions. Furthermore the proposed clustering method can find the dense or sparse areas with different data distributions. It will be convenient to discover the distribution mode and interesting relationship among data. We have conducted numerous experiments in order to justify this novel ideal of spatial data clustering. It has been shown that the proposed method can be applied to spatial clustering well. [ABSTRACT FROM AUTHOR]

Naïve-Bayes classifiers (NB) support incremental learning. However, the lack of effective incremental discretization methods has been hindering NB's incremental learning in face of quantitative data. This problem is further compounded by the fact that quantitative data are everywhere, from temperature readings to share prices. In this paper, we present a novel incremental discretization method for NB, incremental flexible frequency discretization (IFFD). IFFD discretizes values of a quantitative attribute into a sequence of intervals of flexible sizes. It allows online insertion and splitting operation on intervals. Theoretical analysis and experimental test are conducted to compare IFFD with alternative methods. Empirical evidence suggests that IFFD is efficient and effective. NB coupled with IFFD achieves a rapport between high learning efficiency and high classification accuracy in the context of incremental learning. [ABSTRACT FROM AUTHOR]

Gene Expression Programming (GEP) aims at discovering essential rules hidden in observed data and expressing them mathematically. GEP has been proved to be a powerful tool for constructing efficient classifiers. Traditional GEP-classifiers ignore the distribution of samples, and hence decrease the efficiency and accuracy. The contributions of this paper include: (1) proposing two strategies of generating classification threshold dynamically, (2) designing a new approach called Distance Guided Evolution Algorithm (DGEA) to improve the efficiency of GEP, and (3) demonstrating the effectiveness of generating classification threshold dynamically and DGEA by extensive experiments. The results show that the new methods decrease the number of evolutional generations by 83% to 90%, and increase the accuracy by 20% compared with the traditional approach. [ABSTRACT FROM AUTHOR]

This paper presents a new approach for inducing decision trees by combining information entropy criteria with VPRS based methods. From the angle of rough set theory, when inducing decision trees, entropy based methods emphasize the effect of class distribution. Whereas the rough set based approaches emphasize the effect of certainty. The presented approach takes the advantages of both criteria for inducing decision trees. Comparisons between the presented approach and the fundamental information entropy based method on some data sets from the UCI Machine Learning Repository are also reported. [ABSTRACT FROM AUTHOR]

This paper presents a study on the combination of different classifiers for toxicity prediction. Two combination operators for the Multiple-Classifier System definition are also proposed. The classification methods used to generate classifiers for combination are chosen in terms of their representability and diversity and include the Instance-based Learning algorithm (IBL), Decision Tree learning algorithm (DT), Repeated Incremental Pruning to Produce Error Reduction (RIPPER), Multi-Layer Perceptrons (MLPs) and Support Vector Machine (SVM). An effective approach of combining class-wise expertise of diverse classifiers is proposed and evaluated on seven toxicity data sets. The experimental results show that the performance of the combined classifier based on our approach over seven data sets can achieve 69.24% classification accuracy on average, which is better than that of the best classifier (generated by MLP) and four combination schemes studied. [ABSTRACT FROM AUTHOR]

The problem of finding anomaly has received much attention recently. However, most of the anomaly detection algorithms depend on an explicit definition of anomaly, which may be impossible to elicit from a domain expert. Using discords as anomaly detectors is useful since less parameter setting is required. Keogh et al proposed an efficient method for solving this problem. However, their algorithm requires users to choose the word size for the compression of subsequences. In this paper, we propose an algorithm which can dynamically determine the word size for compression. Our method is based on some properties of the Haar wavelet transformation. Our experiments show that this method is highly effective. [ABSTRACT FROM AUTHOR]

Automated design of circuits is a vital task, which becomes more and more challenging due to the conflict of ever-growing scales and complexities of circuits and slow acquisition of relevant knowledge. Evolutionary design of circuit (EDC) combined with data mining is a promising way to solve the problem. To improve EDC in the aspects of efficiency, scalability and capability of optimization, a novel technique is developed. It features an adaptive multi-objective genetic algorithm and interactions between EDC and data mining. The proposed method is validated by the experiments on arithmetic circuits, showing many exciting results especially some novel knowledge discovered from the EDC data. [ABSTRACT FROM AUTHOR]

Efficient online detection of similar patterns under arbitrary time scaling of a given time sequence is a challenging problem in the real-time application field of time series data mining. Some methods based on sliding window have been proposed. Although their ideas are simple and easy to realize, their computational loads are very expensive. Therefore, model based methods are proposed. Recently, the segmental semi-Markov model is introduced into the field of online series pattern detection. However, it can only detect the matching sequences with approximately equal length to that of the query pattern. In this paper, an improved segmental semi-Markov model, which can solve this challenging problem, is proposed. And it is successfully demonstrated on real data sets. [ABSTRACT FROM AUTHOR]

Time delay is a general phenomenon in industrial process. Accurate evaluation on delay is important in data preprocessing when mine manufactory process data. As typical streaming time series, sensors' data of industrial process have attracted much attention recently. A new concept , trend similarity search, is proposed based on raw monotony between two industrial process variables. The new concept is for those two time series which are similar only in trend but dissimilar in shape, whereas similarity search may not do well in such condition. An algorithm DelayMine is also proposed to mine delay between two interrelated time series by trend similarity search. Moreover, the DelayMine is extended to online algorithm for processing streaming time series. The properties and performance of DelayMine is demonstrated through experiments both on systems with steady and time-varying delay. [ABSTRACT FROM AUTHOR]

Fuzzy co-clustering is a method that performs simultaneous fuzzy clustering of objects and features. In this paper, we introduce a new fuzzy co-clustering algorithm for high-dimensional datasets called Cosine-Distance-based & Dual-partitioning Fuzzy Co-clustering (CODIALING FCC). Unlike many existing fuzzy co-clustering algorithms, CODIALING FCC is a dual-partitioning algorithm. It clusters the features in the same manner as it clusters the objects, that is, by partitioning them according to their natural groupings. It is also a cosine-distance-based algorithm because it utilizes the cosine distance to capture the belongingness of objects and features in the co-clusters. Our main purpose of introducing this new algorithm is to improve the performance of some prominent existing fuzzy co-clustering algorithms in dealing with datasets with high overlaps. In our opinion, this is very crucial since most real-world datasets involve significant amount of overlaps in their inherent clustering structures. We discuss how this improvement can be made through the dual-partitioning formulation adopted. Experimental results on a toy problem and five large benchmark document datasets demonstrate the effectiveness of CODIALING FCC in handling overlaps better. [ABSTRACT FROM AUTHOR]

Ordinary decision tree classifiers are used to classify data with single-valued attributes and single-class labels. This paper develops a new decision tree classifier SSC for multi-valued and multi-labeled data, on the basis of the algorithm MMDT, improves on the core formula for measuring the similarity of label-sets, which is the essential index in determining the goodness of splitting attributes, and proposes a new approach of measuring similarity considering both same and consistent features of label-sets, and together with a dynamic approach of adjusting the calculation proportion of the two features according to current data set. SSC makes the similarity of label-sets measured more comprehensive and accurate. The empirical results prove that SSC indeed improves the accuracy of MMDT, and has better classification efficiency. [ABSTRACT FROM AUTHOR]

In order to mine spatial-temporal clusters from geo-databases, two clustering methods with close relationships are proposed, which are both based on neighborhood searching strategy, and rely on the sorted k-dist graph to automatically specify their respective algorithm arguments. We declare the most distinguishing advantage of our clustering methods is they avoid calculating the spatial-temporal distance between patterns which is a tough job. Our methods are validated with the successful extraction of seismic sequence from seismic databases, which is a typical example of spatial-temporal clusters. [ABSTRACT FROM AUTHOR]

Positive correlation mining can find such type of patterns, "the conditional probability that a customer purchasing A is likely to also purchase B is not only great enough, but also significantly greater than the probability that a customer purchases only B." However, there often exist many independence relationships between items in a correlated pattern due to the definition of a correlated pattern. Therefore, we mine mutually and positively correlated patterns, whose any two sub-patterns are both associated and positively correlated. A new correlation interestingness measure is proposed for rationally evaluating the correlation degree. In order to improve the mining efficiency, we combine association with correlation and use not only the correlation measure but also the association measure in the mining process. Our experimental results show that mutually and positively correlated pattern mining is a good approach to discovering patterns which can reflect both association and positive correlation relationships between items at the same time. Meanwhile, our experimental results show that the mining combined association with correlation is quite a valid method to decrease the execution time. [ABSTRACT FROM AUTHOR]

Using a single confidence threshold will result in a dilemmatic situation when simultaneously studying positive and negative association rule (PNAR), i.e., the forms A$\Rightarrow$B, A$\Rightarrow$$\neg$B, $\neg$A$\Rightarrow$B and $\neg$A$\Rightarrow$$\neg$B. A method based on four confidence thresholds for the four forms of PNARs is proposed. The relationships among the four confidences, which show the necessity of using multiple confidence thresholds, are also discussed. In addition, the chi-squared test can avoid generating misleading rules that maybe occur when simultaneously studying the PNARs. The method of how to apply chi-squared test in mining association rules is discussed. An algorithm PNARMC based on the chi-squared test and the four confidence thresholds is proposed. The experimental results demonstrate that the algorithm can not only generate PNARs rightly, but can also control the total number of rules flexibly. [ABSTRACT FROM AUTHOR]

We propose an efficient approach to mine frequent Itemsets on data streams. It is a memory efficient and accurate one-pass algorithm that can deal with batch updates. The proposed algorithm performers well by dividing all frequent itemsets into frequent equivalence classes and pruning all redundant itemsets except for those that represent GLB (Greatest Lower Bound) and LUB (Least Upper Bound) of the frequent equivalence classes. The number of GLB and LUB is much less than the number of frequent itemsets. The experimental evaluation on synthetic and real datasets shows that the algorithm is very accurate and requires significantly lower memory than other well-known one-pass algorithms. [ABSTRACT FROM AUTHOR]

Given a data stream of numerical data elements generated from multiple sources, we consider the problem of maintaining the sum of the elements for each data source over a sliding window of the data stream. The difficulties of the problem come from two parts. One is the number of data sources and the other is the number of elements in the sliding window. For massive data sources, we need a significant number of counters to maintain the sum for each data source, while for a large number of data elements in the sliding window, we need a huge space to keep all of them. We propose two methods, which shares the counters efficiently and merge the data elements systematically so that we are able to estimate the sums using a concise data structure. Two parameters, ε and δ, are needed to construct the data structure. ε controls the bounds of the estimate and δ represents the confidence level that the estimate is within the bounds. The estimates of both methods are proven to be bounded within a factor of ε at 1-δ probability. [ABSTRACT FROM AUTHOR]

In order to discover the knowledge with various granularities from amounts of spatial data, a view-angle of spatial data mining is proposed. First, the view-angle of spatial data mining is defined. In its context, the essentials of spatial data mining are further developed. And the view-angle based algorithms are also presented. Second, the view-angles of Baota landslide-monitoring data mining, and their pan-hierarchical relationships, are given. Finally, view-angle III is taken as a case study to discover quantitative, qualitative and visualized knowledge from Baota landslide-monitoring databases. The results indicate that the view-angle based data mining is practical, and the discovered knowledge with various granularities may satisfy spatial decision-making at different hierarchies. [ABSTRACT FROM AUTHOR]

In order to improve the performance of a data mining model, many researchers have employed a hybrid model approach in solving a problem. There are two types of approach to build a hybrid model, i.e., the whole data approach and the segmented data approach. In this research, we present a new structure of the latter type of hybrid model, which we shall call SePI. In the SePI, input data is segmented using the performance information of the models tried in the training phase. We applied the SePI to a real customer churn problem of a Korean company that provides streaming digital music services through Internet. The result shows that the SePI outperformed any model that employed only one data mining technique such as artificial neural network, decision tree and logistic regression. [ABSTRACT FROM AUTHOR]

This work describes the use of a recently proposed technique - gene expression programming - for knowledge discovery in the data mining task of data classification. We propose a new method for rule encoding and genetic operators that preserve rule integrity, and implemented a system, named GEPCLASS. Due to its encoding scheme, the system allows the automatic discovery of flexible rules, better fitted to data. The performance of GEPCLASS was compared with two genetic programming systems and with C4.5, over four data sets in a five-fold cross-validation procedure. The predictive accuracy for the methods compared were similar, but the computational effort needed by GEPCLASS was significantly smaller than the other. GEPCLASS was able to find simple and accurate rules as it can handle continuous and categorical attributes. [ABSTRACT FROM AUTHOR]